fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "ctree.h"
  20 #include "extent-tree.h"
  21 #include "tree-log.h"
  22 #include "disk-io.h"
  23 #include "print-tree.h"
  24 #include "volumes.h"
  25 #include "raid56.h"
  26 #include "locking.h"
  27 #include "free-space-cache.h"
  28 #include "free-space-tree.h"
  29 #include "sysfs.h"
  30 #include "qgroup.h"
  31 #include "ref-verify.h"
  32 #include "space-info.h"
  33 #include "block-rsv.h"
  34 #include "delalloc-space.h"
  35 #include "discard.h"
  36 #include "rcu-string.h"
  37 #include "zoned.h"
  38 #include "dev-replace.h"
  39 #include "fs.h"
  40 #include "accessors.h"
  41 #include "root-tree.h"
  42 #include "file-item.h"
  43 #include "orphan.h"
  44 #include "tree-checker.h"
  45
  46 #undef SCRAMBLE_DELAYED_REFS
  47
  48
  49 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  50                                struct btrfs_delayed_ref_node *node, u64 parent,
  51                                u64 root_objectid, u64 owner_objectid,
  52                                u64 owner_offset, int refs_to_drop,
  53                                struct btrfs_delayed_extent_op *extra_op);
  54 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  55                                     struct extent_buffer *leaf,
  56                                     struct btrfs_extent_item *ei);
  57 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  58                                       u64 parent, u64 root_objectid,
  59                                       u64 flags, u64 owner, u64 offset,
  60                                       struct btrfs_key *ins, int ref_mod);
  61 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  62                                      struct btrfs_delayed_ref_node *node,
  63                                      struct btrfs_delayed_extent_op *extent_op);
  64 static int find_next_key(struct btrfs_path *path, int level,
  65                          struct btrfs_key *key);
  66
  67 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  68 {
  69         return (cache->flags & bits) == bits;
  70 }
  71
  72 int btrfs_add_excluded_extent(struct btrfs_fs_info *fs_info,
  73                               u64 start, u64 num_bytes)
  74 {
  75         u64 end = start + num_bytes - 1;
  76         set_extent_bits(&fs_info->excluded_extents, start, end,
  77                         EXTENT_UPTODATE);
  78         return 0;
  79 }
  80
  81 void btrfs_free_excluded_extents(struct btrfs_block_group *cache)
  82 {
  83         struct btrfs_fs_info *fs_info = cache->fs_info;
  84         u64 start, end;
  85
  86         start = cache->start;
  87         end = start + cache->length - 1;
  88
  89         clear_extent_bits(&fs_info->excluded_extents, start, end,
  90                           EXTENT_UPTODATE);
  91 }
  92
  93 /* simple helper to search for an existing data extent at a given offset */
  94 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  95 {
  96         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  97         int ret;
  98         struct btrfs_key key;
  99         struct btrfs_path *path;
 100
 101         path = btrfs_alloc_path();
 102         if (!path)
 103                 return -ENOMEM;
 104
 105         key.objectid = start;
 106         key.offset = len;
 107         key.type = BTRFS_EXTENT_ITEM_KEY;
 108         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 109         btrfs_free_path(path);
 110         return ret;
 111 }
 112
 113 /*
 114  * helper function to lookup reference count and flags of a tree block.
 115  *
 116  * the head node for delayed ref is used to store the sum of all the
 117  * reference count modifications queued up in the rbtree. the head
 118  * node may also store the extent flags to set. This way you can check
 119  * to see what the reference count and extent flags would be if all of
 120  * the delayed refs are not processed.
 121  */
 122 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 123                              struct btrfs_fs_info *fs_info, u64 bytenr,
 124                              u64 offset, int metadata, u64 *refs, u64 *flags)
 125 {
 126         struct btrfs_root *extent_root;
 127         struct btrfs_delayed_ref_head *head;
 128         struct btrfs_delayed_ref_root *delayed_refs;
 129         struct btrfs_path *path;
 130         struct btrfs_extent_item *ei;
 131         struct extent_buffer *leaf;
 132         struct btrfs_key key;
 133         u32 item_size;
 134         u64 num_refs;
 135         u64 extent_flags;
 136         int ret;
 137
 138         /*
 139          * If we don't have skinny metadata, don't bother doing anything
 140          * different
 141          */
 142         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 143                 offset = fs_info->nodesize;
 144                 metadata = 0;
 145         }
 146
 147         path = btrfs_alloc_path();
 148         if (!path)
 149                 return -ENOMEM;
 150
 151         if (!trans) {
 152                 path->skip_locking = 1;
 153                 path->search_commit_root = 1;
 154         }
 155
 156 search_again:
 157         key.objectid = bytenr;
 158         key.offset = offset;
 159         if (metadata)
 160                 key.type = BTRFS_METADATA_ITEM_KEY;
 161         else
 162                 key.type = BTRFS_EXTENT_ITEM_KEY;
 163
 164         extent_root = btrfs_extent_root(fs_info, bytenr);
 165         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 166         if (ret < 0)
 167                 goto out_free;
 168
 169         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 170                 if (path->slots[0]) {
 171                         path->slots[0]--;
 172                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 173                                               path->slots[0]);
 174                         if (key.objectid == bytenr &&
 175                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 176                             key.offset == fs_info->nodesize)
 177                                 ret = 0;
 178                 }
 179         }
 180
 181         if (ret == 0) {
 182                 leaf = path->nodes[0];
 183                 item_size = btrfs_item_size(leaf, path->slots[0]);
 184                 if (item_size >= sizeof(*ei)) {
 185                         ei = btrfs_item_ptr(leaf, path->slots[0],
 186                                             struct btrfs_extent_item);
 187                         num_refs = btrfs_extent_refs(leaf, ei);
 188                         extent_flags = btrfs_extent_flags(leaf, ei);
 189                 } else {
 190                         ret = -EINVAL;
 191                         btrfs_print_v0_err(fs_info);
 192                         if (trans)
 193                                 btrfs_abort_transaction(trans, ret);
 194                         else
 195                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 196
 197                         goto out_free;
 198                 }
 199
 200                 BUG_ON(num_refs == 0);
 201         } else {
 202                 num_refs = 0;
 203                 extent_flags = 0;
 204                 ret = 0;
 205         }
 206
 207         if (!trans)
 208                 goto out;
 209
 210         delayed_refs = &trans->transaction->delayed_refs;
 211         spin_lock(&delayed_refs->lock);
 212         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 213         if (head) {
 214                 if (!mutex_trylock(&head->mutex)) {
 215                         refcount_inc(&head->refs);
 216                         spin_unlock(&delayed_refs->lock);
 217
 218                         btrfs_release_path(path);
 219
 220                         /*
 221                          * Mutex was contended, block until it's released and try
 222                          * again
 223                          */
 224                         mutex_lock(&head->mutex);
 225                         mutex_unlock(&head->mutex);
 226                         btrfs_put_delayed_ref_head(head);
 227                         goto search_again;
 228                 }
 229                 spin_lock(&head->lock);
 230                 if (head->extent_op && head->extent_op->update_flags)
 231                         extent_flags |= head->extent_op->flags_to_set;
 232                 else
 233                         BUG_ON(num_refs == 0);
 234
 235                 num_refs += head->ref_mod;
 236                 spin_unlock(&head->lock);
 237                 mutex_unlock(&head->mutex);
 238         }
 239         spin_unlock(&delayed_refs->lock);
 240 out:
 241         WARN_ON(num_refs == 0);
 242         if (refs)
 243                 *refs = num_refs;
 244         if (flags)
 245                 *flags = extent_flags;
 246 out_free:
 247         btrfs_free_path(path);
 248         return ret;
 249 }
 250
 251 /*
 252  * Back reference rules.  Back refs have three main goals:
 253  *
 254  * 1) differentiate between all holders of references to an extent so that
 255  *    when a reference is dropped we can make sure it was a valid reference
 256  *    before freeing the extent.
 257  *
 258  * 2) Provide enough information to quickly find the holders of an extent
 259  *    if we notice a given block is corrupted or bad.
 260  *
 261  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 262  *    maintenance.  This is actually the same as #2, but with a slightly
 263  *    different use case.
 264  *
 265  * There are two kinds of back refs. The implicit back refs is optimized
 266  * for pointers in non-shared tree blocks. For a given pointer in a block,
 267  * back refs of this kind provide information about the block's owner tree
 268  * and the pointer's key. These information allow us to find the block by
 269  * b-tree searching. The full back refs is for pointers in tree blocks not
 270  * referenced by their owner trees. The location of tree block is recorded
 271  * in the back refs. Actually the full back refs is generic, and can be
 272  * used in all cases the implicit back refs is used. The major shortcoming
 273  * of the full back refs is its overhead. Every time a tree block gets
 274  * COWed, we have to update back refs entry for all pointers in it.
 275  *
 276  * For a newly allocated tree block, we use implicit back refs for
 277  * pointers in it. This means most tree related operations only involve
 278  * implicit back refs. For a tree block created in old transaction, the
 279  * only way to drop a reference to it is COW it. So we can detect the
 280  * event that tree block loses its owner tree's reference and do the
 281  * back refs conversion.
 282  *
 283  * When a tree block is COWed through a tree, there are four cases:
 284  *
 285  * The reference count of the block is one and the tree is the block's
 286  * owner tree. Nothing to do in this case.
 287  *
 288  * The reference count of the block is one and the tree is not the
 289  * block's owner tree. In this case, full back refs is used for pointers
 290  * in the block. Remove these full back refs, add implicit back refs for
 291  * every pointers in the new block.
 292  *
 293  * The reference count of the block is greater than one and the tree is
 294  * the block's owner tree. In this case, implicit back refs is used for
 295  * pointers in the block. Add full back refs for every pointers in the
 296  * block, increase lower level extents' reference counts. The original
 297  * implicit back refs are entailed to the new block.
 298  *
 299  * The reference count of the block is greater than one and the tree is
 300  * not the block's owner tree. Add implicit back refs for every pointer in
 301  * the new block, increase lower level extents' reference count.
 302  *
 303  * Back Reference Key composing:
 304  *
 305  * The key objectid corresponds to the first byte in the extent,
 306  * The key type is used to differentiate between types of back refs.
 307  * There are different meanings of the key offset for different types
 308  * of back refs.
 309  *
 310  * File extents can be referenced by:
 311  *
 312  * - multiple snapshots, subvolumes, or different generations in one subvol
 313  * - different files inside a single subvolume
 314  * - different offsets inside a file (bookend extents in file.c)
 315  *
 316  * The extent ref structure for the implicit back refs has fields for:
 317  *
 318  * - Objectid of the subvolume root
 319  * - objectid of the file holding the reference
 320  * - original offset in the file
 321  * - how many bookend extents
 322  *
 323  * The key offset for the implicit back refs is hash of the first
 324  * three fields.
 325  *
 326  * The extent ref structure for the full back refs has field for:
 327  *
 328  * - number of pointers in the tree leaf
 329  *
 330  * The key offset for the implicit back refs is the first byte of
 331  * the tree leaf
 332  *
 333  * When a file extent is allocated, The implicit back refs is used.
 334  * the fields are filled in:
 335  *
 336  *     (root_key.objectid, inode objectid, offset in file, 1)
 337  *
 338  * When a file extent is removed file truncation, we find the
 339  * corresponding implicit back refs and check the following fields:
 340  *
 341  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 342  *
 343  * Btree extents can be referenced by:
 344  *
 345  * - Different subvolumes
 346  *
 347  * Both the implicit back refs and the full back refs for tree blocks
 348  * only consist of key. The key offset for the implicit back refs is
 349  * objectid of block's owner tree. The key offset for the full back refs
 350  * is the first byte of parent block.
 351  *
 352  * When implicit back refs is used, information about the lowest key and
 353  * level of the tree block are required. These information are stored in
 354  * tree block info structure.
 355  */
 356
 357 /*
 358  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 359  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 360  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 361  */
 362 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 363                                      struct btrfs_extent_inline_ref *iref,
 364                                      enum btrfs_inline_ref_type is_data)
 365 {
 366         int type = btrfs_extent_inline_ref_type(eb, iref);
 367         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 368
 369         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 370             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 371             type == BTRFS_SHARED_DATA_REF_KEY ||
 372             type == BTRFS_EXTENT_DATA_REF_KEY) {
 373                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 374                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 375                                 return type;
 376                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 377                                 ASSERT(eb->fs_info);
 378                                 /*
 379                                  * Every shared one has parent tree block,
 380                                  * which must be aligned to sector size.
 381                                  */
 382                                 if (offset &&
 383                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 384                                         return type;
 385                         }
 386                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 387                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 388                                 return type;
 389                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 390                                 ASSERT(eb->fs_info);
 391                                 /*
 392                                  * Every shared one has parent tree block,
 393                                  * which must be aligned to sector size.
 394                                  */
 395                                 if (offset &&
 396                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 397                                         return type;
 398                         }
 399                 } else {
 400                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 401                         return type;
 402                 }
 403         }
 404
 405         btrfs_print_leaf((struct extent_buffer *)eb);
 406         btrfs_err(eb->fs_info,
 407                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 408                   eb->start, (unsigned long)iref, type);
 409         WARN_ON(1);
 410
 411         return BTRFS_REF_TYPE_INVALID;
 412 }
 413
 414 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 415 {
 416         u32 high_crc = ~(u32)0;
 417         u32 low_crc = ~(u32)0;
 418         __le64 lenum;
 419
 420         lenum = cpu_to_le64(root_objectid);
 421         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
 422         lenum = cpu_to_le64(owner);
 423         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 424         lenum = cpu_to_le64(offset);
 425         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 426
 427         return ((u64)high_crc << 31) ^ (u64)low_crc;
 428 }
 429
 430 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 431                                      struct btrfs_extent_data_ref *ref)
 432 {
 433         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 434                                     btrfs_extent_data_ref_objectid(leaf, ref),
 435                                     btrfs_extent_data_ref_offset(leaf, ref));
 436 }
 437
 438 static int match_extent_data_ref(struct extent_buffer *leaf,
 439                                  struct btrfs_extent_data_ref *ref,
 440                                  u64 root_objectid, u64 owner, u64 offset)
 441 {
 442         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 443             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 444             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 445                 return 0;
 446         return 1;
 447 }
 448
 449 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 450                                            struct btrfs_path *path,
 451                                            u64 bytenr, u64 parent,
 452                                            u64 root_objectid,
 453                                            u64 owner, u64 offset)
 454 {
 455         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 456         struct btrfs_key key;
 457         struct btrfs_extent_data_ref *ref;
 458         struct extent_buffer *leaf;
 459         u32 nritems;
 460         int ret;
 461         int recow;
 462         int err = -ENOENT;
 463
 464         key.objectid = bytenr;
 465         if (parent) {
 466                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 467                 key.offset = parent;
 468         } else {
 469                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 470                 key.offset = hash_extent_data_ref(root_objectid,
 471                                                   owner, offset);
 472         }
 473 again:
 474         recow = 0;
 475         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 476         if (ret < 0) {
 477                 err = ret;
 478                 goto fail;
 479         }
 480
 481         if (parent) {
 482                 if (!ret)
 483                         return 0;
 484                 goto fail;
 485         }
 486
 487         leaf = path->nodes[0];
 488         nritems = btrfs_header_nritems(leaf);
 489         while (1) {
 490                 if (path->slots[0] >= nritems) {
 491                         ret = btrfs_next_leaf(root, path);
 492                         if (ret < 0)
 493                                 err = ret;
 494                         if (ret)
 495                                 goto fail;
 496
 497                         leaf = path->nodes[0];
 498                         nritems = btrfs_header_nritems(leaf);
 499                         recow = 1;
 500                 }
 501
 502                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 503                 if (key.objectid != bytenr ||
 504                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 505                         goto fail;
 506
 507                 ref = btrfs_item_ptr(leaf, path->slots[0],
 508                                      struct btrfs_extent_data_ref);
 509
 510                 if (match_extent_data_ref(leaf, ref, root_objectid,
 511                                           owner, offset)) {
 512                         if (recow) {
 513                                 btrfs_release_path(path);
 514                                 goto again;
 515                         }
 516                         err = 0;
 517                         break;
 518                 }
 519                 path->slots[0]++;
 520         }
 521 fail:
 522         return err;
 523 }
 524
 525 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 526                                            struct btrfs_path *path,
 527                                            u64 bytenr, u64 parent,
 528                                            u64 root_objectid, u64 owner,
 529                                            u64 offset, int refs_to_add)
 530 {
 531         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 532         struct btrfs_key key;
 533         struct extent_buffer *leaf;
 534         u32 size;
 535         u32 num_refs;
 536         int ret;
 537
 538         key.objectid = bytenr;
 539         if (parent) {
 540                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 541                 key.offset = parent;
 542                 size = sizeof(struct btrfs_shared_data_ref);
 543         } else {
 544                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 545                 key.offset = hash_extent_data_ref(root_objectid,
 546                                                   owner, offset);
 547                 size = sizeof(struct btrfs_extent_data_ref);
 548         }
 549
 550         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 551         if (ret && ret != -EEXIST)
 552                 goto fail;
 553
 554         leaf = path->nodes[0];
 555         if (parent) {
 556                 struct btrfs_shared_data_ref *ref;
 557                 ref = btrfs_item_ptr(leaf, path->slots[0],
 558                                      struct btrfs_shared_data_ref);
 559                 if (ret == 0) {
 560                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 561                 } else {
 562                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 563                         num_refs += refs_to_add;
 564                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 565                 }
 566         } else {
 567                 struct btrfs_extent_data_ref *ref;
 568                 while (ret == -EEXIST) {
 569                         ref = btrfs_item_ptr(leaf, path->slots[0],
 570                                              struct btrfs_extent_data_ref);
 571                         if (match_extent_data_ref(leaf, ref, root_objectid,
 572                                                   owner, offset))
 573                                 break;
 574                         btrfs_release_path(path);
 575                         key.offset++;
 576                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 577                                                       size);
 578                         if (ret && ret != -EEXIST)
 579                                 goto fail;
 580
 581                         leaf = path->nodes[0];
 582                 }
 583                 ref = btrfs_item_ptr(leaf, path->slots[0],
 584                                      struct btrfs_extent_data_ref);
 585                 if (ret == 0) {
 586                         btrfs_set_extent_data_ref_root(leaf, ref,
 587                                                        root_objectid);
 588                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 589                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 590                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 591                 } else {
 592                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 593                         num_refs += refs_to_add;
 594                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 595                 }
 596         }
 597         btrfs_mark_buffer_dirty(leaf);
 598         ret = 0;
 599 fail:
 600         btrfs_release_path(path);
 601         return ret;
 602 }
 603
 604 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 605                                            struct btrfs_root *root,
 606                                            struct btrfs_path *path,
 607                                            int refs_to_drop)
 608 {
 609         struct btrfs_key key;
 610         struct btrfs_extent_data_ref *ref1 = NULL;
 611         struct btrfs_shared_data_ref *ref2 = NULL;
 612         struct extent_buffer *leaf;
 613         u32 num_refs = 0;
 614         int ret = 0;
 615
 616         leaf = path->nodes[0];
 617         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 618
 619         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 620                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 621                                       struct btrfs_extent_data_ref);
 622                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 623         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 624                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 625                                       struct btrfs_shared_data_ref);
 626                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 627         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 628                 btrfs_print_v0_err(trans->fs_info);
 629                 btrfs_abort_transaction(trans, -EINVAL);
 630                 return -EINVAL;
 631         } else {
 632                 BUG();
 633         }
 634
 635         BUG_ON(num_refs < refs_to_drop);
 636         num_refs -= refs_to_drop;
 637
 638         if (num_refs == 0) {
 639                 ret = btrfs_del_item(trans, root, path);
 640         } else {
 641                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 642                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 643                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 644                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 645                 btrfs_mark_buffer_dirty(leaf);
 646         }
 647         return ret;
 648 }
 649
 650 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 651                                           struct btrfs_extent_inline_ref *iref)
 652 {
 653         struct btrfs_key key;
 654         struct extent_buffer *leaf;
 655         struct btrfs_extent_data_ref *ref1;
 656         struct btrfs_shared_data_ref *ref2;
 657         u32 num_refs = 0;
 658         int type;
 659
 660         leaf = path->nodes[0];
 661         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 662
 663         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 664         if (iref) {
 665                 /*
 666                  * If type is invalid, we should have bailed out earlier than
 667                  * this call.
 668                  */
 669                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 670                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 671                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 672                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 673                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 674                 } else {
 675                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 676                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 677                 }
 678         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 679                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 680                                       struct btrfs_extent_data_ref);
 681                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 682         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 683                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 684                                       struct btrfs_shared_data_ref);
 685                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 686         } else {
 687                 WARN_ON(1);
 688         }
 689         return num_refs;
 690 }
 691
 692 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 693                                           struct btrfs_path *path,
 694                                           u64 bytenr, u64 parent,
 695                                           u64 root_objectid)
 696 {
 697         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 698         struct btrfs_key key;
 699         int ret;
 700
 701         key.objectid = bytenr;
 702         if (parent) {
 703                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 704                 key.offset = parent;
 705         } else {
 706                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 707                 key.offset = root_objectid;
 708         }
 709
 710         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 711         if (ret > 0)
 712                 ret = -ENOENT;
 713         return ret;
 714 }
 715
 716 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 717                                           struct btrfs_path *path,
 718                                           u64 bytenr, u64 parent,
 719                                           u64 root_objectid)
 720 {
 721         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 722         struct btrfs_key key;
 723         int ret;
 724
 725         key.objectid = bytenr;
 726         if (parent) {
 727                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 728                 key.offset = parent;
 729         } else {
 730                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 731                 key.offset = root_objectid;
 732         }
 733
 734         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 735         btrfs_release_path(path);
 736         return ret;
 737 }
 738
 739 static inline int extent_ref_type(u64 parent, u64 owner)
 740 {
 741         int type;
 742         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 743                 if (parent > 0)
 744                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 745                 else
 746                         type = BTRFS_TREE_BLOCK_REF_KEY;
 747         } else {
 748                 if (parent > 0)
 749                         type = BTRFS_SHARED_DATA_REF_KEY;
 750                 else
 751                         type = BTRFS_EXTENT_DATA_REF_KEY;
 752         }
 753         return type;
 754 }
 755
 756 static int find_next_key(struct btrfs_path *path, int level,
 757                          struct btrfs_key *key)
 758
 759 {
 760         for (; level < BTRFS_MAX_LEVEL; level++) {
 761                 if (!path->nodes[level])
 762                         break;
 763                 if (path->slots[level] + 1 >=
 764                     btrfs_header_nritems(path->nodes[level]))
 765                         continue;
 766                 if (level == 0)
 767                         btrfs_item_key_to_cpu(path->nodes[level], key,
 768                                               path->slots[level] + 1);
 769                 else
 770                         btrfs_node_key_to_cpu(path->nodes[level], key,
 771                                               path->slots[level] + 1);
 772                 return 0;
 773         }
 774         return 1;
 775 }
 776
 777 /*
 778  * look for inline back ref. if back ref is found, *ref_ret is set
 779  * to the address of inline back ref, and 0 is returned.
 780  *
 781  * if back ref isn't found, *ref_ret is set to the address where it
 782  * should be inserted, and -ENOENT is returned.
 783  *
 784  * if insert is true and there are too many inline back refs, the path
 785  * points to the extent item, and -EAGAIN is returned.
 786  *
 787  * NOTE: inline back refs are ordered in the same way that back ref
 788  *       items in the tree are ordered.
 789  */
 790 static noinline_for_stack
 791 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 792                                  struct btrfs_path *path,
 793                                  struct btrfs_extent_inline_ref **ref_ret,
 794                                  u64 bytenr, u64 num_bytes,
 795                                  u64 parent, u64 root_objectid,
 796                                  u64 owner, u64 offset, int insert)
 797 {
 798         struct btrfs_fs_info *fs_info = trans->fs_info;
 799         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 800         struct btrfs_key key;
 801         struct extent_buffer *leaf;
 802         struct btrfs_extent_item *ei;
 803         struct btrfs_extent_inline_ref *iref;
 804         u64 flags;
 805         u64 item_size;
 806         unsigned long ptr;
 807         unsigned long end;
 808         int extra_size;
 809         int type;
 810         int want;
 811         int ret;
 812         int err = 0;
 813         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 814         int needed;
 815
 816         key.objectid = bytenr;
 817         key.type = BTRFS_EXTENT_ITEM_KEY;
 818         key.offset = num_bytes;
 819
 820         want = extent_ref_type(parent, owner);
 821         if (insert) {
 822                 extra_size = btrfs_extent_inline_ref_size(want);
 823                 path->search_for_extension = 1;
 824                 path->keep_locks = 1;
 825         } else
 826                 extra_size = -1;
 827
 828         /*
 829          * Owner is our level, so we can just add one to get the level for the
 830          * block we are interested in.
 831          */
 832         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 833                 key.type = BTRFS_METADATA_ITEM_KEY;
 834                 key.offset = owner;
 835         }
 836
 837 again:
 838         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 839         if (ret < 0) {
 840                 err = ret;
 841                 goto out;
 842         }
 843
 844         /*
 845          * We may be a newly converted file system which still has the old fat
 846          * extent entries for metadata, so try and see if we have one of those.
 847          */
 848         if (ret > 0 && skinny_metadata) {
 849                 skinny_metadata = false;
 850                 if (path->slots[0]) {
 851                         path->slots[0]--;
 852                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 853                                               path->slots[0]);
 854                         if (key.objectid == bytenr &&
 855                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 856                             key.offset == num_bytes)
 857                                 ret = 0;
 858                 }
 859                 if (ret) {
 860                         key.objectid = bytenr;
 861                         key.type = BTRFS_EXTENT_ITEM_KEY;
 862                         key.offset = num_bytes;
 863                         btrfs_release_path(path);
 864                         goto again;
 865                 }
 866         }
 867
 868         if (ret && !insert) {
 869                 err = -ENOENT;
 870                 goto out;
 871         } else if (WARN_ON(ret)) {
 872                 err = -EIO;
 873                 goto out;
 874         }
 875
 876         leaf = path->nodes[0];
 877         item_size = btrfs_item_size(leaf, path->slots[0]);
 878         if (unlikely(item_size < sizeof(*ei))) {
 879                 err = -EINVAL;
 880                 btrfs_print_v0_err(fs_info);
 881                 btrfs_abort_transaction(trans, err);
 882                 goto out;
 883         }
 884
 885         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 886         flags = btrfs_extent_flags(leaf, ei);
 887
 888         ptr = (unsigned long)(ei + 1);
 889         end = (unsigned long)ei + item_size;
 890
 891         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 892                 ptr += sizeof(struct btrfs_tree_block_info);
 893                 BUG_ON(ptr > end);
 894         }
 895
 896         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 897                 needed = BTRFS_REF_TYPE_DATA;
 898         else
 899                 needed = BTRFS_REF_TYPE_BLOCK;
 900
 901         err = -ENOENT;
 902         while (1) {
 903                 if (ptr >= end) {
 904                         if (ptr > end) {
 905                                 err = -EUCLEAN;
 906                                 btrfs_print_leaf(path->nodes[0]);
 907                                 btrfs_crit(fs_info,
 908 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 909                                         path->slots[0], root_objectid, owner, offset, parent);
 910                         }
 911                         break;
 912                 }
 913                 iref = (struct btrfs_extent_inline_ref *)ptr;
 914                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 915                 if (type == BTRFS_REF_TYPE_INVALID) {
 916                         err = -EUCLEAN;
 917                         goto out;
 918                 }
 919
 920                 if (want < type)
 921                         break;
 922                 if (want > type) {
 923                         ptr += btrfs_extent_inline_ref_size(type);
 924                         continue;
 925                 }
 926
 927                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 928                         struct btrfs_extent_data_ref *dref;
 929                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 930                         if (match_extent_data_ref(leaf, dref, root_objectid,
 931                                                   owner, offset)) {
 932                                 err = 0;
 933                                 break;
 934                         }
 935                         if (hash_extent_data_ref_item(leaf, dref) <
 936                             hash_extent_data_ref(root_objectid, owner, offset))
 937                                 break;
 938                 } else {
 939                         u64 ref_offset;
 940                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 941                         if (parent > 0) {
 942                                 if (parent == ref_offset) {
 943                                         err = 0;
 944                                         break;
 945                                 }
 946                                 if (ref_offset < parent)
 947                                         break;
 948                         } else {
 949                                 if (root_objectid == ref_offset) {
 950                                         err = 0;
 951                                         break;
 952                                 }
 953                                 if (ref_offset < root_objectid)
 954                                         break;
 955                         }
 956                 }
 957                 ptr += btrfs_extent_inline_ref_size(type);
 958         }
 959         if (err == -ENOENT && insert) {
 960                 if (item_size + extra_size >=
 961                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 962                         err = -EAGAIN;
 963                         goto out;
 964                 }
 965                 /*
 966                  * To add new inline back ref, we have to make sure
 967                  * there is no corresponding back ref item.
 968                  * For simplicity, we just do not add new inline back
 969                  * ref if there is any kind of item for this block
 970                  */
 971                 if (find_next_key(path, 0, &key) == 0 &&
 972                     key.objectid == bytenr &&
 973                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 974                         err = -EAGAIN;
 975                         goto out;
 976                 }
 977         }
 978         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 979 out:
 980         if (insert) {
 981                 path->keep_locks = 0;
 982                 path->search_for_extension = 0;
 983                 btrfs_unlock_up_safe(path, 1);
 984         }
 985         return err;
 986 }
 987
 988 /*
 989  * helper to add new inline back ref
 990  */
 991 static noinline_for_stack
 992 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
 993                                  struct btrfs_path *path,
 994                                  struct btrfs_extent_inline_ref *iref,
 995                                  u64 parent, u64 root_objectid,
 996                                  u64 owner, u64 offset, int refs_to_add,
 997                                  struct btrfs_delayed_extent_op *extent_op)
 998 {
 999         struct extent_buffer *leaf;
1000         struct btrfs_extent_item *ei;
1001         unsigned long ptr;
1002         unsigned long end;
1003         unsigned long item_offset;
1004         u64 refs;
1005         int size;
1006         int type;
1007
1008         leaf = path->nodes[0];
1009         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1010         item_offset = (unsigned long)iref - (unsigned long)ei;
1011
1012         type = extent_ref_type(parent, owner);
1013         size = btrfs_extent_inline_ref_size(type);
1014
1015         btrfs_extend_item(path, size);
1016
1017         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1018         refs = btrfs_extent_refs(leaf, ei);
1019         refs += refs_to_add;
1020         btrfs_set_extent_refs(leaf, ei, refs);
1021         if (extent_op)
1022                 __run_delayed_extent_op(extent_op, leaf, ei);
1023
1024         ptr = (unsigned long)ei + item_offset;
1025         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1026         if (ptr < end - size)
1027                 memmove_extent_buffer(leaf, ptr + size, ptr,
1028                                       end - size - ptr);
1029
1030         iref = (struct btrfs_extent_inline_ref *)ptr;
1031         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1032         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1033                 struct btrfs_extent_data_ref *dref;
1034                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1035                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1036                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1037                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1038                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1039         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1040                 struct btrfs_shared_data_ref *sref;
1041                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1042                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1043                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1044         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1045                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1046         } else {
1047                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1048         }
1049         btrfs_mark_buffer_dirty(leaf);
1050 }
1051
1052 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1053                                  struct btrfs_path *path,
1054                                  struct btrfs_extent_inline_ref **ref_ret,
1055                                  u64 bytenr, u64 num_bytes, u64 parent,
1056                                  u64 root_objectid, u64 owner, u64 offset)
1057 {
1058         int ret;
1059
1060         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1061                                            num_bytes, parent, root_objectid,
1062                                            owner, offset, 0);
1063         if (ret != -ENOENT)
1064                 return ret;
1065
1066         btrfs_release_path(path);
1067         *ref_ret = NULL;
1068
1069         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1070                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1071                                             root_objectid);
1072         } else {
1073                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1074                                              root_objectid, owner, offset);
1075         }
1076         return ret;
1077 }
1078
1079 /*
1080  * helper to update/remove inline back ref
1081  */
1082 static noinline_for_stack
1083 void update_inline_extent_backref(struct btrfs_path *path,
1084                                   struct btrfs_extent_inline_ref *iref,
1085                                   int refs_to_mod,
1086                                   struct btrfs_delayed_extent_op *extent_op)
1087 {
1088         struct extent_buffer *leaf = path->nodes[0];
1089         struct btrfs_extent_item *ei;
1090         struct btrfs_extent_data_ref *dref = NULL;
1091         struct btrfs_shared_data_ref *sref = NULL;
1092         unsigned long ptr;
1093         unsigned long end;
1094         u32 item_size;
1095         int size;
1096         int type;
1097         u64 refs;
1098
1099         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1100         refs = btrfs_extent_refs(leaf, ei);
1101         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1102         refs += refs_to_mod;
1103         btrfs_set_extent_refs(leaf, ei, refs);
1104         if (extent_op)
1105                 __run_delayed_extent_op(extent_op, leaf, ei);
1106
1107         /*
1108          * If type is invalid, we should have bailed out after
1109          * lookup_inline_extent_backref().
1110          */
1111         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1112         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1113
1114         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1115                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1116                 refs = btrfs_extent_data_ref_count(leaf, dref);
1117         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1118                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1119                 refs = btrfs_shared_data_ref_count(leaf, sref);
1120         } else {
1121                 refs = 1;
1122                 BUG_ON(refs_to_mod != -1);
1123         }
1124
1125         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1126         refs += refs_to_mod;
1127
1128         if (refs > 0) {
1129                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1130                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1131                 else
1132                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1133         } else {
1134                 size =  btrfs_extent_inline_ref_size(type);
1135                 item_size = btrfs_item_size(leaf, path->slots[0]);
1136                 ptr = (unsigned long)iref;
1137                 end = (unsigned long)ei + item_size;
1138                 if (ptr + size < end)
1139                         memmove_extent_buffer(leaf, ptr, ptr + size,
1140                                               end - ptr - size);
1141                 item_size -= size;
1142                 btrfs_truncate_item(path, item_size, 1);
1143         }
1144         btrfs_mark_buffer_dirty(leaf);
1145 }
1146
1147 static noinline_for_stack
1148 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1149                                  struct btrfs_path *path,
1150                                  u64 bytenr, u64 num_bytes, u64 parent,
1151                                  u64 root_objectid, u64 owner,
1152                                  u64 offset, int refs_to_add,
1153                                  struct btrfs_delayed_extent_op *extent_op)
1154 {
1155         struct btrfs_extent_inline_ref *iref;
1156         int ret;
1157
1158         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1159                                            num_bytes, parent, root_objectid,
1160                                            owner, offset, 1);
1161         if (ret == 0) {
1162                 /*
1163                  * We're adding refs to a tree block we already own, this
1164                  * should not happen at all.
1165                  */
1166                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1167                         btrfs_crit(trans->fs_info,
1168 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu",
1169                                    bytenr, num_bytes, root_objectid);
1170                         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
1171                                 WARN_ON(1);
1172                                 btrfs_crit(trans->fs_info,
1173                         "path->slots[0]=%d path->nodes[0]:", path->slots[0]);
1174                                 btrfs_print_leaf(path->nodes[0]);
1175                         }
1176                         return -EUCLEAN;
1177                 }
1178                 update_inline_extent_backref(path, iref, refs_to_add, extent_op);
1179         } else if (ret == -ENOENT) {
1180                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1181                                             root_objectid, owner, offset,
1182                                             refs_to_add, extent_op);
1183                 ret = 0;
1184         }
1185         return ret;
1186 }
1187
1188 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1189                                  struct btrfs_root *root,
1190                                  struct btrfs_path *path,
1191                                  struct btrfs_extent_inline_ref *iref,
1192                                  int refs_to_drop, int is_data)
1193 {
1194         int ret = 0;
1195
1196         BUG_ON(!is_data && refs_to_drop != 1);
1197         if (iref)
1198                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL);
1199         else if (is_data)
1200                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1201         else
1202                 ret = btrfs_del_item(trans, root, path);
1203         return ret;
1204 }
1205
1206 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1207                                u64 *discarded_bytes)
1208 {
1209         int j, ret = 0;
1210         u64 bytes_left, end;
1211         u64 aligned_start = ALIGN(start, 1 << 9);
1212
1213         if (WARN_ON(start != aligned_start)) {
1214                 len -= aligned_start - start;
1215                 len = round_down(len, 1 << 9);
1216                 start = aligned_start;
1217         }
1218
1219         *discarded_bytes = 0;
1220
1221         if (!len)
1222                 return 0;
1223
1224         end = start + len;
1225         bytes_left = len;
1226
1227         /* Skip any superblocks on this device. */
1228         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1229                 u64 sb_start = btrfs_sb_offset(j);
1230                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1231                 u64 size = sb_start - start;
1232
1233                 if (!in_range(sb_start, start, bytes_left) &&
1234                     !in_range(sb_end, start, bytes_left) &&
1235                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1236                         continue;
1237
1238                 /*
1239                  * Superblock spans beginning of range.  Adjust start and
1240                  * try again.
1241                  */
1242                 if (sb_start <= start) {
1243                         start += sb_end - start;
1244                         if (start > end) {
1245                                 bytes_left = 0;
1246                                 break;
1247                         }
1248                         bytes_left = end - start;
1249                         continue;
1250                 }
1251
1252                 if (size) {
1253                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1254                                                    GFP_NOFS);
1255                         if (!ret)
1256                                 *discarded_bytes += size;
1257                         else if (ret != -EOPNOTSUPP)
1258                                 return ret;
1259                 }
1260
1261                 start = sb_end;
1262                 if (start > end) {
1263                         bytes_left = 0;
1264                         break;
1265                 }
1266                 bytes_left = end - start;
1267         }
1268
1269         if (bytes_left) {
1270                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1271                                            GFP_NOFS);
1272                 if (!ret)
1273                         *discarded_bytes += bytes_left;
1274         }
1275         return ret;
1276 }
1277
1278 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1279 {
1280         struct btrfs_device *dev = stripe->dev;
1281         struct btrfs_fs_info *fs_info = dev->fs_info;
1282         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1283         u64 phys = stripe->physical;
1284         u64 len = stripe->length;
1285         u64 discarded = 0;
1286         int ret = 0;
1287
1288         /* Zone reset on a zoned filesystem */
1289         if (btrfs_can_zone_reset(dev, phys, len)) {
1290                 u64 src_disc;
1291
1292                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1293                 if (ret)
1294                         goto out;
1295
1296                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1297                     dev != dev_replace->srcdev)
1298                         goto out;
1299
1300                 src_disc = discarded;
1301
1302                 /* Send to replace target as well */
1303                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1304                                               &discarded);
1305                 discarded += src_disc;
1306         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1307                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1308         } else {
1309                 ret = 0;
1310                 *bytes = 0;
1311         }
1312
1313 out:
1314         *bytes = discarded;
1315         return ret;
1316 }
1317
1318 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1319                          u64 num_bytes, u64 *actual_bytes)
1320 {
1321         int ret = 0;
1322         u64 discarded_bytes = 0;
1323         u64 end = bytenr + num_bytes;
1324         u64 cur = bytenr;
1325
1326         /*
1327          * Avoid races with device replace and make sure the devices in the
1328          * stripes don't go away while we are discarding.
1329          */
1330         btrfs_bio_counter_inc_blocked(fs_info);
1331         while (cur < end) {
1332                 struct btrfs_discard_stripe *stripes;
1333                 unsigned int num_stripes;
1334                 int i;
1335
1336                 num_bytes = end - cur;
1337                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1338                 if (IS_ERR(stripes)) {
1339                         ret = PTR_ERR(stripes);
1340                         if (ret == -EOPNOTSUPP)
1341                                 ret = 0;
1342                         break;
1343                 }
1344
1345                 for (i = 0; i < num_stripes; i++) {
1346                         struct btrfs_discard_stripe *stripe = stripes + i;
1347                         u64 bytes;
1348
1349                         if (!stripe->dev->bdev) {
1350                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1351                                 continue;
1352                         }
1353
1354                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1355                                         &stripe->dev->dev_state))
1356                                 continue;
1357
1358                         ret = do_discard_extent(stripe, &bytes);
1359                         if (ret) {
1360                                 /*
1361                                  * Keep going if discard is not supported by the
1362                                  * device.
1363                                  */
1364                                 if (ret != -EOPNOTSUPP)
1365                                         break;
1366                                 ret = 0;
1367                         } else {
1368                                 discarded_bytes += bytes;
1369                         }
1370                 }
1371                 kfree(stripes);
1372                 if (ret)
1373                         break;
1374                 cur += num_bytes;
1375         }
1376         btrfs_bio_counter_dec(fs_info);
1377         if (actual_bytes)
1378                 *actual_bytes = discarded_bytes;
1379         return ret;
1380 }
1381
1382 /* Can return -ENOMEM */
1383 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1384                          struct btrfs_ref *generic_ref)
1385 {
1386         struct btrfs_fs_info *fs_info = trans->fs_info;
1387         int ret;
1388
1389         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1390                generic_ref->action);
1391         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1392                generic_ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID);
1393
1394         if (generic_ref->type == BTRFS_REF_METADATA)
1395                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1396         else
1397                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1398
1399         btrfs_ref_tree_mod(fs_info, generic_ref);
1400
1401         return ret;
1402 }
1403
1404 /*
1405  * __btrfs_inc_extent_ref - insert backreference for a given extent
1406  *
1407  * The counterpart is in __btrfs_free_extent(), with examples and more details
1408  * how it works.
1409  *
1410  * @trans:          Handle of transaction
1411  *
1412  * @node:           The delayed ref node used to get the bytenr/length for
1413  *                  extent whose references are incremented.
1414  *
1415  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1416  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1417  *                  bytenr of the parent block. Since new extents are always
1418  *                  created with indirect references, this will only be the case
1419  *                  when relocating a shared extent. In that case, root_objectid
1420  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1421  *                  be 0
1422  *
1423  * @root_objectid:  The id of the root where this modification has originated,
1424  *                  this can be either one of the well-known metadata trees or
1425  *                  the subvolume id which references this extent.
1426  *
1427  * @owner:          For data extents it is the inode number of the owning file.
1428  *                  For metadata extents this parameter holds the level in the
1429  *                  tree of the extent.
1430  *
1431  * @offset:         For metadata extents the offset is ignored and is currently
1432  *                  always passed as 0. For data extents it is the fileoffset
1433  *                  this extent belongs to.
1434  *
1435  * @refs_to_add     Number of references to add
1436  *
1437  * @extent_op       Pointer to a structure, holding information necessary when
1438  *                  updating a tree block's flags
1439  *
1440  */
1441 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1442                                   struct btrfs_delayed_ref_node *node,
1443                                   u64 parent, u64 root_objectid,
1444                                   u64 owner, u64 offset, int refs_to_add,
1445                                   struct btrfs_delayed_extent_op *extent_op)
1446 {
1447         struct btrfs_path *path;
1448         struct extent_buffer *leaf;
1449         struct btrfs_extent_item *item;
1450         struct btrfs_key key;
1451         u64 bytenr = node->bytenr;
1452         u64 num_bytes = node->num_bytes;
1453         u64 refs;
1454         int ret;
1455
1456         path = btrfs_alloc_path();
1457         if (!path)
1458                 return -ENOMEM;
1459
1460         /* this will setup the path even if it fails to insert the back ref */
1461         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1462                                            parent, root_objectid, owner,
1463                                            offset, refs_to_add, extent_op);
1464         if ((ret < 0 && ret != -EAGAIN) || !ret)
1465                 goto out;
1466
1467         /*
1468          * Ok we had -EAGAIN which means we didn't have space to insert and
1469          * inline extent ref, so just update the reference count and add a
1470          * normal backref.
1471          */
1472         leaf = path->nodes[0];
1473         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1474         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1475         refs = btrfs_extent_refs(leaf, item);
1476         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1477         if (extent_op)
1478                 __run_delayed_extent_op(extent_op, leaf, item);
1479
1480         btrfs_mark_buffer_dirty(leaf);
1481         btrfs_release_path(path);
1482
1483         /* now insert the actual backref */
1484         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1485                 BUG_ON(refs_to_add != 1);
1486                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1487                                             root_objectid);
1488         } else {
1489                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1490                                              root_objectid, owner, offset,
1491                                              refs_to_add);
1492         }
1493         if (ret)
1494                 btrfs_abort_transaction(trans, ret);
1495 out:
1496         btrfs_free_path(path);
1497         return ret;
1498 }
1499
1500 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1501                                 struct btrfs_delayed_ref_node *node,
1502                                 struct btrfs_delayed_extent_op *extent_op,
1503                                 int insert_reserved)
1504 {
1505         int ret = 0;
1506         struct btrfs_delayed_data_ref *ref;
1507         struct btrfs_key ins;
1508         u64 parent = 0;
1509         u64 ref_root = 0;
1510         u64 flags = 0;
1511
1512         ins.objectid = node->bytenr;
1513         ins.offset = node->num_bytes;
1514         ins.type = BTRFS_EXTENT_ITEM_KEY;
1515
1516         ref = btrfs_delayed_node_to_data_ref(node);
1517         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1518
1519         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1520                 parent = ref->parent;
1521         ref_root = ref->root;
1522
1523         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1524                 if (extent_op)
1525                         flags |= extent_op->flags_to_set;
1526                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
1527                                                  flags, ref->objectid,
1528                                                  ref->offset, &ins,
1529                                                  node->ref_mod);
1530         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1531                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1532                                              ref->objectid, ref->offset,
1533                                              node->ref_mod, extent_op);
1534         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1535                 ret = __btrfs_free_extent(trans, node, parent,
1536                                           ref_root, ref->objectid,
1537                                           ref->offset, node->ref_mod,
1538                                           extent_op);
1539         } else {
1540                 BUG();
1541         }
1542         return ret;
1543 }
1544
1545 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1546                                     struct extent_buffer *leaf,
1547                                     struct btrfs_extent_item *ei)
1548 {
1549         u64 flags = btrfs_extent_flags(leaf, ei);
1550         if (extent_op->update_flags) {
1551                 flags |= extent_op->flags_to_set;
1552                 btrfs_set_extent_flags(leaf, ei, flags);
1553         }
1554
1555         if (extent_op->update_key) {
1556                 struct btrfs_tree_block_info *bi;
1557                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1558                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1559                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1560         }
1561 }
1562
1563 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1564                                  struct btrfs_delayed_ref_head *head,
1565                                  struct btrfs_delayed_extent_op *extent_op)
1566 {
1567         struct btrfs_fs_info *fs_info = trans->fs_info;
1568         struct btrfs_root *root;
1569         struct btrfs_key key;
1570         struct btrfs_path *path;
1571         struct btrfs_extent_item *ei;
1572         struct extent_buffer *leaf;
1573         u32 item_size;
1574         int ret;
1575         int err = 0;
1576         int metadata = 1;
1577
1578         if (TRANS_ABORTED(trans))
1579                 return 0;
1580
1581         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1582                 metadata = 0;
1583
1584         path = btrfs_alloc_path();
1585         if (!path)
1586                 return -ENOMEM;
1587
1588         key.objectid = head->bytenr;
1589
1590         if (metadata) {
1591                 key.type = BTRFS_METADATA_ITEM_KEY;
1592                 key.offset = extent_op->level;
1593         } else {
1594                 key.type = BTRFS_EXTENT_ITEM_KEY;
1595                 key.offset = head->num_bytes;
1596         }
1597
1598         root = btrfs_extent_root(fs_info, key.objectid);
1599 again:
1600         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1601         if (ret < 0) {
1602                 err = ret;
1603                 goto out;
1604         }
1605         if (ret > 0) {
1606                 if (metadata) {
1607                         if (path->slots[0] > 0) {
1608                                 path->slots[0]--;
1609                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1610                                                       path->slots[0]);
1611                                 if (key.objectid == head->bytenr &&
1612                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1613                                     key.offset == head->num_bytes)
1614                                         ret = 0;
1615                         }
1616                         if (ret > 0) {
1617                                 btrfs_release_path(path);
1618                                 metadata = 0;
1619
1620                                 key.objectid = head->bytenr;
1621                                 key.offset = head->num_bytes;
1622                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1623                                 goto again;
1624                         }
1625                 } else {
1626                         err = -EIO;
1627                         goto out;
1628                 }
1629         }
1630
1631         leaf = path->nodes[0];
1632         item_size = btrfs_item_size(leaf, path->slots[0]);
1633
1634         if (unlikely(item_size < sizeof(*ei))) {
1635                 err = -EINVAL;
1636                 btrfs_print_v0_err(fs_info);
1637                 btrfs_abort_transaction(trans, err);
1638                 goto out;
1639         }
1640
1641         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1642         __run_delayed_extent_op(extent_op, leaf, ei);
1643
1644         btrfs_mark_buffer_dirty(leaf);
1645 out:
1646         btrfs_free_path(path);
1647         return err;
1648 }
1649
1650 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1651                                 struct btrfs_delayed_ref_node *node,
1652                                 struct btrfs_delayed_extent_op *extent_op,
1653                                 int insert_reserved)
1654 {
1655         int ret = 0;
1656         struct btrfs_delayed_tree_ref *ref;
1657         u64 parent = 0;
1658         u64 ref_root = 0;
1659
1660         ref = btrfs_delayed_node_to_tree_ref(node);
1661         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1662
1663         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1664                 parent = ref->parent;
1665         ref_root = ref->root;
1666
1667         if (node->ref_mod != 1) {
1668                 btrfs_err(trans->fs_info,
1669         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
1670                           node->bytenr, node->ref_mod, node->action, ref_root,
1671                           parent);
1672                 return -EIO;
1673         }
1674         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1675                 BUG_ON(!extent_op || !extent_op->update_flags);
1676                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1677         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1678                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1679                                              ref->level, 0, 1, extent_op);
1680         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1681                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1682                                           ref->level, 0, 1, extent_op);
1683         } else {
1684                 BUG();
1685         }
1686         return ret;
1687 }
1688
1689 /* helper function to actually process a single delayed ref entry */
1690 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1691                                struct btrfs_delayed_ref_node *node,
1692                                struct btrfs_delayed_extent_op *extent_op,
1693                                int insert_reserved)
1694 {
1695         int ret = 0;
1696
1697         if (TRANS_ABORTED(trans)) {
1698                 if (insert_reserved)
1699                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1700                 return 0;
1701         }
1702
1703         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1704             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1705                 ret = run_delayed_tree_ref(trans, node, extent_op,
1706                                            insert_reserved);
1707         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1708                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1709                 ret = run_delayed_data_ref(trans, node, extent_op,
1710                                            insert_reserved);
1711         else
1712                 BUG();
1713         if (ret && insert_reserved)
1714                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1715         if (ret < 0)
1716                 btrfs_err(trans->fs_info,
1717 "failed to run delayed ref for logical %llu num_bytes %llu type %u action %u ref_mod %d: %d",
1718                           node->bytenr, node->num_bytes, node->type,
1719                           node->action, node->ref_mod, ret);
1720         return ret;
1721 }
1722
1723 static inline struct btrfs_delayed_ref_node *
1724 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1725 {
1726         struct btrfs_delayed_ref_node *ref;
1727
1728         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1729                 return NULL;
1730
1731         /*
1732          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1733          * This is to prevent a ref count from going down to zero, which deletes
1734          * the extent item from the extent tree, when there still are references
1735          * to add, which would fail because they would not find the extent item.
1736          */
1737         if (!list_empty(&head->ref_add_list))
1738                 return list_first_entry(&head->ref_add_list,
1739                                 struct btrfs_delayed_ref_node, add_list);
1740
1741         ref = rb_entry(rb_first_cached(&head->ref_tree),
1742                        struct btrfs_delayed_ref_node, ref_node);
1743         ASSERT(list_empty(&ref->add_list));
1744         return ref;
1745 }
1746
1747 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1748                                       struct btrfs_delayed_ref_head *head)
1749 {
1750         spin_lock(&delayed_refs->lock);
1751         head->processing = 0;
1752         delayed_refs->num_heads_ready++;
1753         spin_unlock(&delayed_refs->lock);
1754         btrfs_delayed_ref_unlock(head);
1755 }
1756
1757 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1758                                 struct btrfs_delayed_ref_head *head)
1759 {
1760         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1761
1762         if (!extent_op)
1763                 return NULL;
1764
1765         if (head->must_insert_reserved) {
1766                 head->extent_op = NULL;
1767                 btrfs_free_delayed_extent_op(extent_op);
1768                 return NULL;
1769         }
1770         return extent_op;
1771 }
1772
1773 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1774                                      struct btrfs_delayed_ref_head *head)
1775 {
1776         struct btrfs_delayed_extent_op *extent_op;
1777         int ret;
1778
1779         extent_op = cleanup_extent_op(head);
1780         if (!extent_op)
1781                 return 0;
1782         head->extent_op = NULL;
1783         spin_unlock(&head->lock);
1784         ret = run_delayed_extent_op(trans, head, extent_op);
1785         btrfs_free_delayed_extent_op(extent_op);
1786         return ret ? ret : 1;
1787 }
1788
1789 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1790                                   struct btrfs_delayed_ref_root *delayed_refs,
1791                                   struct btrfs_delayed_ref_head *head)
1792 {
1793         int nr_items = 1;       /* Dropping this ref head update. */
1794
1795         /*
1796          * We had csum deletions accounted for in our delayed refs rsv, we need
1797          * to drop the csum leaves for this update from our delayed_refs_rsv.
1798          */
1799         if (head->total_ref_mod < 0 && head->is_data) {
1800                 spin_lock(&delayed_refs->lock);
1801                 delayed_refs->pending_csums -= head->num_bytes;
1802                 spin_unlock(&delayed_refs->lock);
1803                 nr_items += btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1804         }
1805
1806         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
1807 }
1808
1809 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1810                             struct btrfs_delayed_ref_head *head)
1811 {
1812
1813         struct btrfs_fs_info *fs_info = trans->fs_info;
1814         struct btrfs_delayed_ref_root *delayed_refs;
1815         int ret;
1816
1817         delayed_refs = &trans->transaction->delayed_refs;
1818
1819         ret = run_and_cleanup_extent_op(trans, head);
1820         if (ret < 0) {
1821                 unselect_delayed_ref_head(delayed_refs, head);
1822                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1823                 return ret;
1824         } else if (ret) {
1825                 return ret;
1826         }
1827
1828         /*
1829          * Need to drop our head ref lock and re-acquire the delayed ref lock
1830          * and then re-check to make sure nobody got added.
1831          */
1832         spin_unlock(&head->lock);
1833         spin_lock(&delayed_refs->lock);
1834         spin_lock(&head->lock);
1835         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1836                 spin_unlock(&head->lock);
1837                 spin_unlock(&delayed_refs->lock);
1838                 return 1;
1839         }
1840         btrfs_delete_ref_head(delayed_refs, head);
1841         spin_unlock(&head->lock);
1842         spin_unlock(&delayed_refs->lock);
1843
1844         if (head->must_insert_reserved) {
1845                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1846                 if (head->is_data) {
1847                         struct btrfs_root *csum_root;
1848
1849                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1850                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1851                                               head->num_bytes);
1852                 }
1853         }
1854
1855         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1856
1857         trace_run_delayed_ref_head(fs_info, head, 0);
1858         btrfs_delayed_ref_unlock(head);
1859         btrfs_put_delayed_ref_head(head);
1860         return ret;
1861 }
1862
1863 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1864                                         struct btrfs_trans_handle *trans)
1865 {
1866         struct btrfs_delayed_ref_root *delayed_refs =
1867                 &trans->transaction->delayed_refs;
1868         struct btrfs_delayed_ref_head *head = NULL;
1869         int ret;
1870
1871         spin_lock(&delayed_refs->lock);
1872         head = btrfs_select_ref_head(delayed_refs);
1873         if (!head) {
1874                 spin_unlock(&delayed_refs->lock);
1875                 return head;
1876         }
1877
1878         /*
1879          * Grab the lock that says we are going to process all the refs for
1880          * this head
1881          */
1882         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1883         spin_unlock(&delayed_refs->lock);
1884
1885         /*
1886          * We may have dropped the spin lock to get the head mutex lock, and
1887          * that might have given someone else time to free the head.  If that's
1888          * true, it has been removed from our list and we can move on.
1889          */
1890         if (ret == -EAGAIN)
1891                 head = ERR_PTR(-EAGAIN);
1892
1893         return head;
1894 }
1895
1896 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1897                                            struct btrfs_delayed_ref_head *locked_ref)
1898 {
1899         struct btrfs_fs_info *fs_info = trans->fs_info;
1900         struct btrfs_delayed_ref_root *delayed_refs;
1901         struct btrfs_delayed_extent_op *extent_op;
1902         struct btrfs_delayed_ref_node *ref;
1903         int must_insert_reserved = 0;
1904         int ret;
1905
1906         delayed_refs = &trans->transaction->delayed_refs;
1907
1908         lockdep_assert_held(&locked_ref->mutex);
1909         lockdep_assert_held(&locked_ref->lock);
1910
1911         while ((ref = select_delayed_ref(locked_ref))) {
1912                 if (ref->seq &&
1913                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1914                         spin_unlock(&locked_ref->lock);
1915                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1916                         return -EAGAIN;
1917                 }
1918
1919                 ref->in_tree = 0;
1920                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1921                 RB_CLEAR_NODE(&ref->ref_node);
1922                 if (!list_empty(&ref->add_list))
1923                         list_del(&ref->add_list);
1924                 /*
1925                  * When we play the delayed ref, also correct the ref_mod on
1926                  * head
1927                  */
1928                 switch (ref->action) {
1929                 case BTRFS_ADD_DELAYED_REF:
1930                 case BTRFS_ADD_DELAYED_EXTENT:
1931                         locked_ref->ref_mod -= ref->ref_mod;
1932                         break;
1933                 case BTRFS_DROP_DELAYED_REF:
1934                         locked_ref->ref_mod += ref->ref_mod;
1935                         break;
1936                 default:
1937                         WARN_ON(1);
1938                 }
1939                 atomic_dec(&delayed_refs->num_entries);
1940
1941                 /*
1942                  * Record the must_insert_reserved flag before we drop the
1943                  * spin lock.
1944                  */
1945                 must_insert_reserved = locked_ref->must_insert_reserved;
1946                 locked_ref->must_insert_reserved = 0;
1947
1948                 extent_op = locked_ref->extent_op;
1949                 locked_ref->extent_op = NULL;
1950                 spin_unlock(&locked_ref->lock);
1951
1952                 ret = run_one_delayed_ref(trans, ref, extent_op,
1953                                           must_insert_reserved);
1954
1955                 btrfs_free_delayed_extent_op(extent_op);
1956                 if (ret) {
1957                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1958                         btrfs_put_delayed_ref(ref);
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(fs_info, 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         int ret;
1983         unsigned long count = 0;
1984
1985         delayed_refs = &trans->transaction->delayed_refs;
1986         do {
1987                 if (!locked_ref) {
1988                         locked_ref = btrfs_obtain_ref_head(trans);
1989                         if (IS_ERR_OR_NULL(locked_ref)) {
1990                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
1991                                         continue;
1992                                 } else {
1993                                         break;
1994                                 }
1995                         }
1996                         count++;
1997                 }
1998                 /*
1999                  * We need to try and merge add/drops of the same ref since we
2000                  * can run into issues with relocate dropping the implicit ref
2001                  * and then it being added back again before the drop can
2002                  * finish.  If we merged anything we need to re-loop so we can
2003                  * get a good ref.
2004                  * Or we can get node references of the same type that weren't
2005                  * merged when created due to bumps in the tree mod seq, and
2006                  * we need to merge them to prevent adding an inline extent
2007                  * backref before dropping it (triggering a BUG_ON at
2008                  * insert_inline_extent_backref()).
2009                  */
2010                 spin_lock(&locked_ref->lock);
2011                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2012
2013                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref);
2014                 if (ret < 0 && ret != -EAGAIN) {
2015                         /*
2016                          * Error, btrfs_run_delayed_refs_for_head already
2017                          * unlocked everything so just bail out
2018                          */
2019                         return ret;
2020                 } else if (!ret) {
2021                         /*
2022                          * Success, perform the usual cleanup of a processed
2023                          * head
2024                          */
2025                         ret = cleanup_ref_head(trans, locked_ref);
2026                         if (ret > 0 ) {
2027                                 /* We dropped our lock, we need to loop. */
2028                                 ret = 0;
2029                                 continue;
2030                         } else if (ret) {
2031                                 return ret;
2032                         }
2033                 }
2034
2035                 /*
2036                  * Either success case or btrfs_run_delayed_refs_for_head
2037                  * returned -EAGAIN, meaning we need to select another head
2038                  */
2039
2040                 locked_ref = NULL;
2041                 cond_resched();
2042         } while ((nr != -1 && count < nr) || locked_ref);
2043
2044         return 0;
2045 }
2046
2047 #ifdef SCRAMBLE_DELAYED_REFS
2048 /*
2049  * Normally delayed refs get processed in ascending bytenr order. This
2050  * correlates in most cases to the order added. To expose dependencies on this
2051  * order, we start to process the tree in the middle instead of the beginning
2052  */
2053 static u64 find_middle(struct rb_root *root)
2054 {
2055         struct rb_node *n = root->rb_node;
2056         struct btrfs_delayed_ref_node *entry;
2057         int alt = 1;
2058         u64 middle;
2059         u64 first = 0, last = 0;
2060
2061         n = rb_first(root);
2062         if (n) {
2063                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2064                 first = entry->bytenr;
2065         }
2066         n = rb_last(root);
2067         if (n) {
2068                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2069                 last = entry->bytenr;
2070         }
2071         n = root->rb_node;
2072
2073         while (n) {
2074                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2075                 WARN_ON(!entry->in_tree);
2076
2077                 middle = entry->bytenr;
2078
2079                 if (alt)
2080                         n = n->rb_left;
2081                 else
2082                         n = n->rb_right;
2083
2084                 alt = 1 - alt;
2085         }
2086         return middle;
2087 }
2088 #endif
2089
2090 /*
2091  * this starts processing the delayed reference count updates and
2092  * extent insertions we have queued up so far.  count can be
2093  * 0, which means to process everything in the tree at the start
2094  * of the run (but not newly added entries), or it can be some target
2095  * number you'd like to process.
2096  *
2097  * Returns 0 on success or if called with an aborted transaction
2098  * Returns <0 on error and aborts the transaction
2099  */
2100 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2101                            unsigned long count)
2102 {
2103         struct btrfs_fs_info *fs_info = trans->fs_info;
2104         struct rb_node *node;
2105         struct btrfs_delayed_ref_root *delayed_refs;
2106         struct btrfs_delayed_ref_head *head;
2107         int ret;
2108         int run_all = count == (unsigned long)-1;
2109
2110         /* We'll clean this up in btrfs_cleanup_transaction */
2111         if (TRANS_ABORTED(trans))
2112                 return 0;
2113
2114         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2115                 return 0;
2116
2117         delayed_refs = &trans->transaction->delayed_refs;
2118         if (count == 0)
2119                 count = delayed_refs->num_heads_ready;
2120
2121 again:
2122 #ifdef SCRAMBLE_DELAYED_REFS
2123         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2124 #endif
2125         ret = __btrfs_run_delayed_refs(trans, count);
2126         if (ret < 0) {
2127                 btrfs_abort_transaction(trans, ret);
2128                 return ret;
2129         }
2130
2131         if (run_all) {
2132                 btrfs_create_pending_block_groups(trans);
2133
2134                 spin_lock(&delayed_refs->lock);
2135                 node = rb_first_cached(&delayed_refs->href_root);
2136                 if (!node) {
2137                         spin_unlock(&delayed_refs->lock);
2138                         goto out;
2139                 }
2140                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2141                                 href_node);
2142                 refcount_inc(&head->refs);
2143                 spin_unlock(&delayed_refs->lock);
2144
2145                 /* Mutex was contended, block until it's released and retry. */
2146                 mutex_lock(&head->mutex);
2147                 mutex_unlock(&head->mutex);
2148
2149                 btrfs_put_delayed_ref_head(head);
2150                 cond_resched();
2151                 goto again;
2152         }
2153 out:
2154         return 0;
2155 }
2156
2157 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2158                                 struct extent_buffer *eb, u64 flags,
2159                                 int level)
2160 {
2161         struct btrfs_delayed_extent_op *extent_op;
2162         int ret;
2163
2164         extent_op = btrfs_alloc_delayed_extent_op();
2165         if (!extent_op)
2166                 return -ENOMEM;
2167
2168         extent_op->flags_to_set = flags;
2169         extent_op->update_flags = true;
2170         extent_op->update_key = false;
2171         extent_op->level = level;
2172
2173         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2174         if (ret)
2175                 btrfs_free_delayed_extent_op(extent_op);
2176         return ret;
2177 }
2178
2179 static noinline int check_delayed_ref(struct btrfs_root *root,
2180                                       struct btrfs_path *path,
2181                                       u64 objectid, u64 offset, u64 bytenr)
2182 {
2183         struct btrfs_delayed_ref_head *head;
2184         struct btrfs_delayed_ref_node *ref;
2185         struct btrfs_delayed_data_ref *data_ref;
2186         struct btrfs_delayed_ref_root *delayed_refs;
2187         struct btrfs_transaction *cur_trans;
2188         struct rb_node *node;
2189         int ret = 0;
2190
2191         spin_lock(&root->fs_info->trans_lock);
2192         cur_trans = root->fs_info->running_transaction;
2193         if (cur_trans)
2194                 refcount_inc(&cur_trans->use_count);
2195         spin_unlock(&root->fs_info->trans_lock);
2196         if (!cur_trans)
2197                 return 0;
2198
2199         delayed_refs = &cur_trans->delayed_refs;
2200         spin_lock(&delayed_refs->lock);
2201         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2202         if (!head) {
2203                 spin_unlock(&delayed_refs->lock);
2204                 btrfs_put_transaction(cur_trans);
2205                 return 0;
2206         }
2207
2208         if (!mutex_trylock(&head->mutex)) {
2209                 if (path->nowait) {
2210                         spin_unlock(&delayed_refs->lock);
2211                         btrfs_put_transaction(cur_trans);
2212                         return -EAGAIN;
2213                 }
2214
2215                 refcount_inc(&head->refs);
2216                 spin_unlock(&delayed_refs->lock);
2217
2218                 btrfs_release_path(path);
2219
2220                 /*
2221                  * Mutex was contended, block until it's released and let
2222                  * caller try again
2223                  */
2224                 mutex_lock(&head->mutex);
2225                 mutex_unlock(&head->mutex);
2226                 btrfs_put_delayed_ref_head(head);
2227                 btrfs_put_transaction(cur_trans);
2228                 return -EAGAIN;
2229         }
2230         spin_unlock(&delayed_refs->lock);
2231
2232         spin_lock(&head->lock);
2233         /*
2234          * XXX: We should replace this with a proper search function in the
2235          * future.
2236          */
2237         for (node = rb_first_cached(&head->ref_tree); node;
2238              node = rb_next(node)) {
2239                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2240                 /* If it's a shared ref we know a cross reference exists */
2241                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2242                         ret = 1;
2243                         break;
2244                 }
2245
2246                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2247
2248                 /*
2249                  * If our ref doesn't match the one we're currently looking at
2250                  * then we have a cross reference.
2251                  */
2252                 if (data_ref->root != root->root_key.objectid ||
2253                     data_ref->objectid != objectid ||
2254                     data_ref->offset != offset) {
2255                         ret = 1;
2256                         break;
2257                 }
2258         }
2259         spin_unlock(&head->lock);
2260         mutex_unlock(&head->mutex);
2261         btrfs_put_transaction(cur_trans);
2262         return ret;
2263 }
2264
2265 static noinline int check_committed_ref(struct btrfs_root *root,
2266                                         struct btrfs_path *path,
2267                                         u64 objectid, u64 offset, u64 bytenr,
2268                                         bool strict)
2269 {
2270         struct btrfs_fs_info *fs_info = root->fs_info;
2271         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2272         struct extent_buffer *leaf;
2273         struct btrfs_extent_data_ref *ref;
2274         struct btrfs_extent_inline_ref *iref;
2275         struct btrfs_extent_item *ei;
2276         struct btrfs_key key;
2277         u32 item_size;
2278         int type;
2279         int ret;
2280
2281         key.objectid = bytenr;
2282         key.offset = (u64)-1;
2283         key.type = BTRFS_EXTENT_ITEM_KEY;
2284
2285         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2286         if (ret < 0)
2287                 goto out;
2288         BUG_ON(ret == 0); /* Corruption */
2289
2290         ret = -ENOENT;
2291         if (path->slots[0] == 0)
2292                 goto out;
2293
2294         path->slots[0]--;
2295         leaf = path->nodes[0];
2296         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2297
2298         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2299                 goto out;
2300
2301         ret = 1;
2302         item_size = btrfs_item_size(leaf, path->slots[0]);
2303         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2304
2305         /* If extent item has more than 1 inline ref then it's shared */
2306         if (item_size != sizeof(*ei) +
2307             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2308                 goto out;
2309
2310         /*
2311          * If extent created before last snapshot => it's shared unless the
2312          * snapshot has been deleted. Use the heuristic if strict is false.
2313          */
2314         if (!strict &&
2315             (btrfs_extent_generation(leaf, ei) <=
2316              btrfs_root_last_snapshot(&root->root_item)))
2317                 goto out;
2318
2319         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2320
2321         /* If this extent has SHARED_DATA_REF then it's shared */
2322         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2323         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2324                 goto out;
2325
2326         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2327         if (btrfs_extent_refs(leaf, ei) !=
2328             btrfs_extent_data_ref_count(leaf, ref) ||
2329             btrfs_extent_data_ref_root(leaf, ref) !=
2330             root->root_key.objectid ||
2331             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2332             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2333                 goto out;
2334
2335         ret = 0;
2336 out:
2337         return ret;
2338 }
2339
2340 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2341                           u64 bytenr, bool strict, struct btrfs_path *path)
2342 {
2343         int ret;
2344
2345         do {
2346                 ret = check_committed_ref(root, path, objectid,
2347                                           offset, bytenr, strict);
2348                 if (ret && ret != -ENOENT)
2349                         goto out;
2350
2351                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2352         } while (ret == -EAGAIN);
2353
2354 out:
2355         btrfs_release_path(path);
2356         if (btrfs_is_data_reloc_root(root))
2357                 WARN_ON(ret > 0);
2358         return ret;
2359 }
2360
2361 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2362                            struct btrfs_root *root,
2363                            struct extent_buffer *buf,
2364                            int full_backref, int inc)
2365 {
2366         struct btrfs_fs_info *fs_info = root->fs_info;
2367         u64 bytenr;
2368         u64 num_bytes;
2369         u64 parent;
2370         u64 ref_root;
2371         u32 nritems;
2372         struct btrfs_key key;
2373         struct btrfs_file_extent_item *fi;
2374         struct btrfs_ref generic_ref = { 0 };
2375         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2376         int i;
2377         int action;
2378         int level;
2379         int ret = 0;
2380
2381         if (btrfs_is_testing(fs_info))
2382                 return 0;
2383
2384         ref_root = btrfs_header_owner(buf);
2385         nritems = btrfs_header_nritems(buf);
2386         level = btrfs_header_level(buf);
2387
2388         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2389                 return 0;
2390
2391         if (full_backref)
2392                 parent = buf->start;
2393         else
2394                 parent = 0;
2395         if (inc)
2396                 action = BTRFS_ADD_DELAYED_REF;
2397         else
2398                 action = BTRFS_DROP_DELAYED_REF;
2399
2400         for (i = 0; i < nritems; i++) {
2401                 if (level == 0) {
2402                         btrfs_item_key_to_cpu(buf, &key, i);
2403                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2404                                 continue;
2405                         fi = btrfs_item_ptr(buf, i,
2406                                             struct btrfs_file_extent_item);
2407                         if (btrfs_file_extent_type(buf, fi) ==
2408                             BTRFS_FILE_EXTENT_INLINE)
2409                                 continue;
2410                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2411                         if (bytenr == 0)
2412                                 continue;
2413
2414                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2415                         key.offset -= btrfs_file_extent_offset(buf, fi);
2416                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2417                                                num_bytes, parent);
2418                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2419                                             key.offset, root->root_key.objectid,
2420                                             for_reloc);
2421                         if (inc)
2422                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2423                         else
2424                                 ret = btrfs_free_extent(trans, &generic_ref);
2425                         if (ret)
2426                                 goto fail;
2427                 } else {
2428                         bytenr = btrfs_node_blockptr(buf, i);
2429                         num_bytes = fs_info->nodesize;
2430                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2431                                                num_bytes, parent);
2432                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root,
2433                                             root->root_key.objectid, for_reloc);
2434                         if (inc)
2435                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2436                         else
2437                                 ret = btrfs_free_extent(trans, &generic_ref);
2438                         if (ret)
2439                                 goto fail;
2440                 }
2441         }
2442         return 0;
2443 fail:
2444         return ret;
2445 }
2446
2447 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2448                   struct extent_buffer *buf, int full_backref)
2449 {
2450         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2451 }
2452
2453 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2454                   struct extent_buffer *buf, int full_backref)
2455 {
2456         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2457 }
2458
2459 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2460 {
2461         struct btrfs_fs_info *fs_info = root->fs_info;
2462         u64 flags;
2463         u64 ret;
2464
2465         if (data)
2466                 flags = BTRFS_BLOCK_GROUP_DATA;
2467         else if (root == fs_info->chunk_root)
2468                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2469         else
2470                 flags = BTRFS_BLOCK_GROUP_METADATA;
2471
2472         ret = btrfs_get_alloc_profile(fs_info, flags);
2473         return ret;
2474 }
2475
2476 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2477 {
2478         struct rb_node *leftmost;
2479         u64 bytenr = 0;
2480
2481         read_lock(&fs_info->block_group_cache_lock);
2482         /* Get the block group with the lowest logical start address. */
2483         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2484         if (leftmost) {
2485                 struct btrfs_block_group *bg;
2486
2487                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2488                 bytenr = bg->start;
2489         }
2490         read_unlock(&fs_info->block_group_cache_lock);
2491
2492         return bytenr;
2493 }
2494
2495 static int pin_down_extent(struct btrfs_trans_handle *trans,
2496                            struct btrfs_block_group *cache,
2497                            u64 bytenr, u64 num_bytes, int reserved)
2498 {
2499         struct btrfs_fs_info *fs_info = cache->fs_info;
2500
2501         spin_lock(&cache->space_info->lock);
2502         spin_lock(&cache->lock);
2503         cache->pinned += num_bytes;
2504         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2505                                              num_bytes);
2506         if (reserved) {
2507                 cache->reserved -= num_bytes;
2508                 cache->space_info->bytes_reserved -= num_bytes;
2509         }
2510         spin_unlock(&cache->lock);
2511         spin_unlock(&cache->space_info->lock);
2512
2513         set_extent_dirty(&trans->transaction->pinned_extents, bytenr,
2514                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
2515         return 0;
2516 }
2517
2518 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2519                      u64 bytenr, u64 num_bytes, int reserved)
2520 {
2521         struct btrfs_block_group *cache;
2522
2523         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2524         BUG_ON(!cache); /* Logic error */
2525
2526         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2527
2528         btrfs_put_block_group(cache);
2529         return 0;
2530 }
2531
2532 /*
2533  * this function must be called within transaction
2534  */
2535 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2536                                     u64 bytenr, u64 num_bytes)
2537 {
2538         struct btrfs_block_group *cache;
2539         int ret;
2540
2541         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2542         if (!cache)
2543                 return -EINVAL;
2544
2545         /*
2546          * Fully cache the free space first so that our pin removes the free space
2547          * from the cache.
2548          */
2549         ret = btrfs_cache_block_group(cache, true);
2550         if (ret)
2551                 goto out;
2552
2553         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2554
2555         /* remove us from the free space cache (if we're there at all) */
2556         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2557 out:
2558         btrfs_put_block_group(cache);
2559         return ret;
2560 }
2561
2562 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2563                                    u64 start, u64 num_bytes)
2564 {
2565         int ret;
2566         struct btrfs_block_group *block_group;
2567
2568         block_group = btrfs_lookup_block_group(fs_info, start);
2569         if (!block_group)
2570                 return -EINVAL;
2571
2572         ret = btrfs_cache_block_group(block_group, true);
2573         if (ret)
2574                 goto out;
2575
2576         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2577 out:
2578         btrfs_put_block_group(block_group);
2579         return ret;
2580 }
2581
2582 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2583 {
2584         struct btrfs_fs_info *fs_info = eb->fs_info;
2585         struct btrfs_file_extent_item *item;
2586         struct btrfs_key key;
2587         int found_type;
2588         int i;
2589         int ret = 0;
2590
2591         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2592                 return 0;
2593
2594         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2595                 btrfs_item_key_to_cpu(eb, &key, i);
2596                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2597                         continue;
2598                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2599                 found_type = btrfs_file_extent_type(eb, item);
2600                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2601                         continue;
2602                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2603                         continue;
2604                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2605                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2606                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2607                 if (ret)
2608                         break;
2609         }
2610
2611         return ret;
2612 }
2613
2614 static void
2615 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2616 {
2617         atomic_inc(&bg->reservations);
2618 }
2619
2620 /*
2621  * Returns the free cluster for the given space info and sets empty_cluster to
2622  * what it should be based on the mount options.
2623  */
2624 static struct btrfs_free_cluster *
2625 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2626                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2627 {
2628         struct btrfs_free_cluster *ret = NULL;
2629
2630         *empty_cluster = 0;
2631         if (btrfs_mixed_space_info(space_info))
2632                 return ret;
2633
2634         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2635                 ret = &fs_info->meta_alloc_cluster;
2636                 if (btrfs_test_opt(fs_info, SSD))
2637                         *empty_cluster = SZ_2M;
2638                 else
2639                         *empty_cluster = SZ_64K;
2640         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2641                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2642                 *empty_cluster = SZ_2M;
2643                 ret = &fs_info->data_alloc_cluster;
2644         }
2645
2646         return ret;
2647 }
2648
2649 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2650                               u64 start, u64 end,
2651                               const bool return_free_space)
2652 {
2653         struct btrfs_block_group *cache = NULL;
2654         struct btrfs_space_info *space_info;
2655         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2656         struct btrfs_free_cluster *cluster = NULL;
2657         u64 len;
2658         u64 total_unpinned = 0;
2659         u64 empty_cluster = 0;
2660         bool readonly;
2661
2662         while (start <= end) {
2663                 readonly = false;
2664                 if (!cache ||
2665                     start >= cache->start + cache->length) {
2666                         if (cache)
2667                                 btrfs_put_block_group(cache);
2668                         total_unpinned = 0;
2669                         cache = btrfs_lookup_block_group(fs_info, start);
2670                         BUG_ON(!cache); /* Logic error */
2671
2672                         cluster = fetch_cluster_info(fs_info,
2673                                                      cache->space_info,
2674                                                      &empty_cluster);
2675                         empty_cluster <<= 1;
2676                 }
2677
2678                 len = cache->start + cache->length - start;
2679                 len = min(len, end + 1 - start);
2680
2681                 if (return_free_space)
2682                         btrfs_add_free_space(cache, start, len);
2683
2684                 start += len;
2685                 total_unpinned += len;
2686                 space_info = cache->space_info;
2687
2688                 /*
2689                  * If this space cluster has been marked as fragmented and we've
2690                  * unpinned enough in this block group to potentially allow a
2691                  * cluster to be created inside of it go ahead and clear the
2692                  * fragmented check.
2693                  */
2694                 if (cluster && cluster->fragmented &&
2695                     total_unpinned > empty_cluster) {
2696                         spin_lock(&cluster->lock);
2697                         cluster->fragmented = 0;
2698                         spin_unlock(&cluster->lock);
2699                 }
2700
2701                 spin_lock(&space_info->lock);
2702                 spin_lock(&cache->lock);
2703                 cache->pinned -= len;
2704                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2705                 space_info->max_extent_size = 0;
2706                 if (cache->ro) {
2707                         space_info->bytes_readonly += len;
2708                         readonly = true;
2709                 } else if (btrfs_is_zoned(fs_info)) {
2710                         /* Need reset before reusing in a zoned block group */
2711                         space_info->bytes_zone_unusable += len;
2712                         readonly = true;
2713                 }
2714                 spin_unlock(&cache->lock);
2715                 if (!readonly && return_free_space &&
2716                     global_rsv->space_info == space_info) {
2717                         spin_lock(&global_rsv->lock);
2718                         if (!global_rsv->full) {
2719                                 u64 to_add = min(len, global_rsv->size -
2720                                                       global_rsv->reserved);
2721
2722                                 global_rsv->reserved += to_add;
2723                                 btrfs_space_info_update_bytes_may_use(fs_info,
2724                                                 space_info, to_add);
2725                                 if (global_rsv->reserved >= global_rsv->size)
2726                                         global_rsv->full = 1;
2727                                 len -= to_add;
2728                         }
2729                         spin_unlock(&global_rsv->lock);
2730                 }
2731                 /* Add to any tickets we may have */
2732                 if (!readonly && return_free_space && len)
2733                         btrfs_try_granting_tickets(fs_info, space_info);
2734                 spin_unlock(&space_info->lock);
2735         }
2736
2737         if (cache)
2738                 btrfs_put_block_group(cache);
2739         return 0;
2740 }
2741
2742 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2743 {
2744         struct btrfs_fs_info *fs_info = trans->fs_info;
2745         struct btrfs_block_group *block_group, *tmp;
2746         struct list_head *deleted_bgs;
2747         struct extent_io_tree *unpin;
2748         u64 start;
2749         u64 end;
2750         int ret;
2751
2752         unpin = &trans->transaction->pinned_extents;
2753
2754         while (!TRANS_ABORTED(trans)) {
2755                 struct extent_state *cached_state = NULL;
2756
2757                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2758                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2759                                             EXTENT_DIRTY, &cached_state);
2760                 if (ret) {
2761                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2762                         break;
2763                 }
2764
2765                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2766                         ret = btrfs_discard_extent(fs_info, start,
2767                                                    end + 1 - start, NULL);
2768
2769                 clear_extent_dirty(unpin, start, end, &cached_state);
2770                 unpin_extent_range(fs_info, start, end, true);
2771                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2772                 free_extent_state(cached_state);
2773                 cond_resched();
2774         }
2775
2776         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2777                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2778                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2779         }
2780
2781         /*
2782          * Transaction is finished.  We don't need the lock anymore.  We
2783          * do need to clean up the block groups in case of a transaction
2784          * abort.
2785          */
2786         deleted_bgs = &trans->transaction->deleted_bgs;
2787         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2788                 u64 trimmed = 0;
2789
2790                 ret = -EROFS;
2791                 if (!TRANS_ABORTED(trans))
2792                         ret = btrfs_discard_extent(fs_info,
2793                                                    block_group->start,
2794                                                    block_group->length,
2795                                                    &trimmed);
2796
2797                 list_del_init(&block_group->bg_list);
2798                 btrfs_unfreeze_block_group(block_group);
2799                 btrfs_put_block_group(block_group);
2800
2801                 if (ret) {
2802                         const char *errstr = btrfs_decode_error(ret);
2803                         btrfs_warn(fs_info,
2804                            "discard failed while removing blockgroup: errno=%d %s",
2805                                    ret, errstr);
2806                 }
2807         }
2808
2809         return 0;
2810 }
2811
2812 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2813                                      u64 bytenr, u64 num_bytes, bool is_data)
2814 {
2815         int ret;
2816
2817         if (is_data) {
2818                 struct btrfs_root *csum_root;
2819
2820                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
2821                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
2822                 if (ret) {
2823                         btrfs_abort_transaction(trans, ret);
2824                         return ret;
2825                 }
2826         }
2827
2828         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
2829         if (ret) {
2830                 btrfs_abort_transaction(trans, ret);
2831                 return ret;
2832         }
2833
2834         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
2835         if (ret)
2836                 btrfs_abort_transaction(trans, ret);
2837
2838         return ret;
2839 }
2840
2841 /*
2842  * Drop one or more refs of @node.
2843  *
2844  * 1. Locate the extent refs.
2845  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2846  *    Locate it, then reduce the refs number or remove the ref line completely.
2847  *
2848  * 2. Update the refs count in EXTENT/METADATA_ITEM
2849  *
2850  * Inline backref case:
2851  *
2852  * in extent tree we have:
2853  *
2854  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2855  *              refs 2 gen 6 flags DATA
2856  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2857  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
2858  *
2859  * This function gets called with:
2860  *
2861  *    node->bytenr = 13631488
2862  *    node->num_bytes = 1048576
2863  *    root_objectid = FS_TREE
2864  *    owner_objectid = 257
2865  *    owner_offset = 0
2866  *    refs_to_drop = 1
2867  *
2868  * Then we should get some like:
2869  *
2870  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2871  *              refs 1 gen 6 flags DATA
2872  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2873  *
2874  * Keyed backref case:
2875  *
2876  * in extent tree we have:
2877  *
2878  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2879  *              refs 754 gen 6 flags DATA
2880  *      [...]
2881  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
2882  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
2883  *
2884  * This function get called with:
2885  *
2886  *    node->bytenr = 13631488
2887  *    node->num_bytes = 1048576
2888  *    root_objectid = FS_TREE
2889  *    owner_objectid = 866
2890  *    owner_offset = 0
2891  *    refs_to_drop = 1
2892  *
2893  * Then we should get some like:
2894  *
2895  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2896  *              refs 753 gen 6 flags DATA
2897  *
2898  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
2899  */
2900 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2901                                struct btrfs_delayed_ref_node *node, u64 parent,
2902                                u64 root_objectid, u64 owner_objectid,
2903                                u64 owner_offset, int refs_to_drop,
2904                                struct btrfs_delayed_extent_op *extent_op)
2905 {
2906         struct btrfs_fs_info *info = trans->fs_info;
2907         struct btrfs_key key;
2908         struct btrfs_path *path;
2909         struct btrfs_root *extent_root;
2910         struct extent_buffer *leaf;
2911         struct btrfs_extent_item *ei;
2912         struct btrfs_extent_inline_ref *iref;
2913         int ret;
2914         int is_data;
2915         int extent_slot = 0;
2916         int found_extent = 0;
2917         int num_to_del = 1;
2918         u32 item_size;
2919         u64 refs;
2920         u64 bytenr = node->bytenr;
2921         u64 num_bytes = node->num_bytes;
2922         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2923
2924         extent_root = btrfs_extent_root(info, bytenr);
2925         ASSERT(extent_root);
2926
2927         path = btrfs_alloc_path();
2928         if (!path)
2929                 return -ENOMEM;
2930
2931         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2932
2933         if (!is_data && refs_to_drop != 1) {
2934                 btrfs_crit(info,
2935 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
2936                            node->bytenr, refs_to_drop);
2937                 ret = -EINVAL;
2938                 btrfs_abort_transaction(trans, ret);
2939                 goto out;
2940         }
2941
2942         if (is_data)
2943                 skinny_metadata = false;
2944
2945         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2946                                     parent, root_objectid, owner_objectid,
2947                                     owner_offset);
2948         if (ret == 0) {
2949                 /*
2950                  * Either the inline backref or the SHARED_DATA_REF/
2951                  * SHARED_BLOCK_REF is found
2952                  *
2953                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
2954                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
2955                  */
2956                 extent_slot = path->slots[0];
2957                 while (extent_slot >= 0) {
2958                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2959                                               extent_slot);
2960                         if (key.objectid != bytenr)
2961                                 break;
2962                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
2963                             key.offset == num_bytes) {
2964                                 found_extent = 1;
2965                                 break;
2966                         }
2967                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
2968                             key.offset == owner_objectid) {
2969                                 found_extent = 1;
2970                                 break;
2971                         }
2972
2973                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
2974                         if (path->slots[0] - extent_slot > 5)
2975                                 break;
2976                         extent_slot--;
2977                 }
2978
2979                 if (!found_extent) {
2980                         if (iref) {
2981                                 btrfs_crit(info,
2982 "invalid iref, no EXTENT/METADATA_ITEM found but has inline extent ref");
2983                                 btrfs_abort_transaction(trans, -EUCLEAN);
2984                                 goto err_dump;
2985                         }
2986                         /* Must be SHARED_* item, remove the backref first */
2987                         ret = remove_extent_backref(trans, extent_root, path,
2988                                                     NULL, refs_to_drop, is_data);
2989                         if (ret) {
2990                                 btrfs_abort_transaction(trans, ret);
2991                                 goto out;
2992                         }
2993                         btrfs_release_path(path);
2994
2995                         /* Slow path to locate EXTENT/METADATA_ITEM */
2996                         key.objectid = bytenr;
2997                         key.type = BTRFS_EXTENT_ITEM_KEY;
2998                         key.offset = num_bytes;
2999
3000                         if (!is_data && skinny_metadata) {
3001                                 key.type = BTRFS_METADATA_ITEM_KEY;
3002                                 key.offset = owner_objectid;
3003                         }
3004
3005                         ret = btrfs_search_slot(trans, extent_root,
3006                                                 &key, path, -1, 1);
3007                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3008                                 /*
3009                                  * Couldn't find our skinny metadata item,
3010                                  * see if we have ye olde extent item.
3011                                  */
3012                                 path->slots[0]--;
3013                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3014                                                       path->slots[0]);
3015                                 if (key.objectid == bytenr &&
3016                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3017                                     key.offset == num_bytes)
3018                                         ret = 0;
3019                         }
3020
3021                         if (ret > 0 && skinny_metadata) {
3022                                 skinny_metadata = false;
3023                                 key.objectid = bytenr;
3024                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3025                                 key.offset = num_bytes;
3026                                 btrfs_release_path(path);
3027                                 ret = btrfs_search_slot(trans, extent_root,
3028                                                         &key, path, -1, 1);
3029                         }
3030
3031                         if (ret) {
3032                                 btrfs_err(info,
3033                                           "umm, got %d back from search, was looking for %llu",
3034                                           ret, bytenr);
3035                                 if (ret > 0)
3036                                         btrfs_print_leaf(path->nodes[0]);
3037                         }
3038                         if (ret < 0) {
3039                                 btrfs_abort_transaction(trans, ret);
3040                                 goto out;
3041                         }
3042                         extent_slot = path->slots[0];
3043                 }
3044         } else if (WARN_ON(ret == -ENOENT)) {
3045                 btrfs_print_leaf(path->nodes[0]);
3046                 btrfs_err(info,
3047                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
3048                         bytenr, parent, root_objectid, owner_objectid,
3049                         owner_offset);
3050                 btrfs_abort_transaction(trans, ret);
3051                 goto out;
3052         } else {
3053                 btrfs_abort_transaction(trans, ret);
3054                 goto out;
3055         }
3056
3057         leaf = path->nodes[0];
3058         item_size = btrfs_item_size(leaf, extent_slot);
3059         if (unlikely(item_size < sizeof(*ei))) {
3060                 ret = -EINVAL;
3061                 btrfs_print_v0_err(info);
3062                 btrfs_abort_transaction(trans, ret);
3063                 goto out;
3064         }
3065         ei = btrfs_item_ptr(leaf, extent_slot,
3066                             struct btrfs_extent_item);
3067         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3068             key.type == BTRFS_EXTENT_ITEM_KEY) {
3069                 struct btrfs_tree_block_info *bi;
3070                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3071                         btrfs_crit(info,
3072 "invalid extent item size for key (%llu, %u, %llu) owner %llu, has %u expect >= %zu",
3073                                    key.objectid, key.type, key.offset,
3074                                    owner_objectid, item_size,
3075                                    sizeof(*ei) + sizeof(*bi));
3076                         btrfs_abort_transaction(trans, -EUCLEAN);
3077                         goto err_dump;
3078                 }
3079                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3080                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3081         }
3082
3083         refs = btrfs_extent_refs(leaf, ei);
3084         if (refs < refs_to_drop) {
3085                 btrfs_crit(info,
3086                 "trying to drop %d refs but we only have %llu for bytenr %llu",
3087                           refs_to_drop, refs, bytenr);
3088                 btrfs_abort_transaction(trans, -EUCLEAN);
3089                 goto err_dump;
3090         }
3091         refs -= refs_to_drop;
3092
3093         if (refs > 0) {
3094                 if (extent_op)
3095                         __run_delayed_extent_op(extent_op, leaf, ei);
3096                 /*
3097                  * In the case of inline back ref, reference count will
3098                  * be updated by remove_extent_backref
3099                  */
3100                 if (iref) {
3101                         if (!found_extent) {
3102                                 btrfs_crit(info,
3103 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found");
3104                                 btrfs_abort_transaction(trans, -EUCLEAN);
3105                                 goto err_dump;
3106                         }
3107                 } else {
3108                         btrfs_set_extent_refs(leaf, ei, refs);
3109                         btrfs_mark_buffer_dirty(leaf);
3110                 }
3111                 if (found_extent) {
3112                         ret = remove_extent_backref(trans, extent_root, path,
3113                                                     iref, refs_to_drop, is_data);
3114                         if (ret) {
3115                                 btrfs_abort_transaction(trans, ret);
3116                                 goto out;
3117                         }
3118                 }
3119         } else {
3120                 /* In this branch refs == 1 */
3121                 if (found_extent) {
3122                         if (is_data && refs_to_drop !=
3123                             extent_data_ref_count(path, iref)) {
3124                                 btrfs_crit(info,
3125                 "invalid refs_to_drop, current refs %u refs_to_drop %u",
3126                                            extent_data_ref_count(path, iref),
3127                                            refs_to_drop);
3128                                 btrfs_abort_transaction(trans, -EUCLEAN);
3129                                 goto err_dump;
3130                         }
3131                         if (iref) {
3132                                 if (path->slots[0] != extent_slot) {
3133                                         btrfs_crit(info,
3134 "invalid iref, extent item key (%llu %u %llu) doesn't have wanted iref",
3135                                                    key.objectid, key.type,
3136                                                    key.offset);
3137                                         btrfs_abort_transaction(trans, -EUCLEAN);
3138                                         goto err_dump;
3139                                 }
3140                         } else {
3141                                 /*
3142                                  * No inline ref, we must be at SHARED_* item,
3143                                  * And it's single ref, it must be:
3144                                  * |    extent_slot       ||extent_slot + 1|
3145                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3146                                  */
3147                                 if (path->slots[0] != extent_slot + 1) {
3148                                         btrfs_crit(info,
3149         "invalid SHARED_* item, previous item is not EXTENT/METADATA_ITEM");
3150                                         btrfs_abort_transaction(trans, -EUCLEAN);
3151                                         goto err_dump;
3152                                 }
3153                                 path->slots[0] = extent_slot;
3154                                 num_to_del = 2;
3155                         }
3156                 }
3157
3158                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3159                                       num_to_del);
3160                 if (ret) {
3161                         btrfs_abort_transaction(trans, ret);
3162                         goto out;
3163                 }
3164                 btrfs_release_path(path);
3165
3166                 ret = do_free_extent_accounting(trans, bytenr, num_bytes, is_data);
3167         }
3168         btrfs_release_path(path);
3169
3170 out:
3171         btrfs_free_path(path);
3172         return ret;
3173 err_dump:
3174         /*
3175          * Leaf dump can take up a lot of log buffer, so we only do full leaf
3176          * dump for debug build.
3177          */
3178         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
3179                 btrfs_crit(info, "path->slots[0]=%d extent_slot=%d",
3180                            path->slots[0], extent_slot);
3181                 btrfs_print_leaf(path->nodes[0]);
3182         }
3183
3184         btrfs_free_path(path);
3185         return -EUCLEAN;
3186 }
3187
3188 /*
3189  * when we free an block, it is possible (and likely) that we free the last
3190  * delayed ref for that extent as well.  This searches the delayed ref tree for
3191  * a given extent, and if there are no other delayed refs to be processed, it
3192  * removes it from the tree.
3193  */
3194 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3195                                       u64 bytenr)
3196 {
3197         struct btrfs_delayed_ref_head *head;
3198         struct btrfs_delayed_ref_root *delayed_refs;
3199         int ret = 0;
3200
3201         delayed_refs = &trans->transaction->delayed_refs;
3202         spin_lock(&delayed_refs->lock);
3203         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3204         if (!head)
3205                 goto out_delayed_unlock;
3206
3207         spin_lock(&head->lock);
3208         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3209                 goto out;
3210
3211         if (cleanup_extent_op(head) != NULL)
3212                 goto out;
3213
3214         /*
3215          * waiting for the lock here would deadlock.  If someone else has it
3216          * locked they are already in the process of dropping it anyway
3217          */
3218         if (!mutex_trylock(&head->mutex))
3219                 goto out;
3220
3221         btrfs_delete_ref_head(delayed_refs, head);
3222         head->processing = 0;
3223
3224         spin_unlock(&head->lock);
3225         spin_unlock(&delayed_refs->lock);
3226
3227         BUG_ON(head->extent_op);
3228         if (head->must_insert_reserved)
3229                 ret = 1;
3230
3231         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3232         mutex_unlock(&head->mutex);
3233         btrfs_put_delayed_ref_head(head);
3234         return ret;
3235 out:
3236         spin_unlock(&head->lock);
3237
3238 out_delayed_unlock:
3239         spin_unlock(&delayed_refs->lock);
3240         return 0;
3241 }
3242
3243 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3244                            u64 root_id,
3245                            struct extent_buffer *buf,
3246                            u64 parent, int last_ref)
3247 {
3248         struct btrfs_fs_info *fs_info = trans->fs_info;
3249         struct btrfs_ref generic_ref = { 0 };
3250         int ret;
3251
3252         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3253                                buf->start, buf->len, parent);
3254         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3255                             root_id, 0, false);
3256
3257         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3258                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3259                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3260                 BUG_ON(ret); /* -ENOMEM */
3261         }
3262
3263         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3264                 struct btrfs_block_group *cache;
3265                 bool must_pin = false;
3266
3267                 if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3268                         ret = check_ref_cleanup(trans, buf->start);
3269                         if (!ret) {
3270                                 btrfs_redirty_list_add(trans->transaction, buf);
3271                                 goto out;
3272                         }
3273                 }
3274
3275                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3276
3277                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3278                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3279                         btrfs_put_block_group(cache);
3280                         goto out;
3281                 }
3282
3283                 /*
3284                  * If there are tree mod log users we may have recorded mod log
3285                  * operations for this node.  If we re-allocate this node we
3286                  * could replay operations on this node that happened when it
3287                  * existed in a completely different root.  For example if it
3288                  * was part of root A, then was reallocated to root B, and we
3289                  * are doing a btrfs_old_search_slot(root b), we could replay
3290                  * operations that happened when the block was part of root A,
3291                  * giving us an inconsistent view of the btree.
3292                  *
3293                  * We are safe from races here because at this point no other
3294                  * node or root points to this extent buffer, so if after this
3295                  * check a new tree mod log user joins we will not have an
3296                  * existing log of operations on this node that we have to
3297                  * contend with.
3298                  */
3299                 if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
3300                         must_pin = true;
3301
3302                 if (must_pin || btrfs_is_zoned(fs_info)) {
3303                         btrfs_redirty_list_add(trans->transaction, buf);
3304                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3305                         btrfs_put_block_group(cache);
3306                         goto out;
3307                 }
3308
3309                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3310
3311                 btrfs_add_free_space(cache, buf->start, buf->len);
3312                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3313                 btrfs_put_block_group(cache);
3314                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3315         }
3316 out:
3317         if (last_ref) {
3318                 /*
3319                  * Deleting the buffer, clear the corrupt flag since it doesn't
3320                  * matter anymore.
3321                  */
3322                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3323         }
3324 }
3325
3326 /* Can return -ENOMEM */
3327 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3328 {
3329         struct btrfs_fs_info *fs_info = trans->fs_info;
3330         int ret;
3331
3332         if (btrfs_is_testing(fs_info))
3333                 return 0;
3334
3335         /*
3336          * tree log blocks never actually go into the extent allocation
3337          * tree, just update pinning info and exit early.
3338          */
3339         if ((ref->type == BTRFS_REF_METADATA &&
3340              ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3341             (ref->type == BTRFS_REF_DATA &&
3342              ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)) {
3343                 /* unlocks the pinned mutex */
3344                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3345                 ret = 0;
3346         } else if (ref->type == BTRFS_REF_METADATA) {
3347                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3348         } else {
3349                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3350         }
3351
3352         if (!((ref->type == BTRFS_REF_METADATA &&
3353                ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3354               (ref->type == BTRFS_REF_DATA &&
3355                ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)))
3356                 btrfs_ref_tree_mod(fs_info, ref);
3357
3358         return ret;
3359 }
3360
3361 enum btrfs_loop_type {
3362         LOOP_CACHING_NOWAIT,
3363         LOOP_CACHING_WAIT,
3364         LOOP_UNSET_SIZE_CLASS,
3365         LOOP_ALLOC_CHUNK,
3366         LOOP_WRONG_SIZE_CLASS,
3367         LOOP_NO_EMPTY_SIZE,
3368 };
3369
3370 static inline void
3371 btrfs_lock_block_group(struct btrfs_block_group *cache,
3372                        int delalloc)
3373 {
3374         if (delalloc)
3375                 down_read(&cache->data_rwsem);
3376 }
3377
3378 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3379                        int delalloc)
3380 {
3381         btrfs_get_block_group(cache);
3382         if (delalloc)
3383                 down_read(&cache->data_rwsem);
3384 }
3385
3386 static struct btrfs_block_group *btrfs_lock_cluster(
3387                    struct btrfs_block_group *block_group,
3388                    struct btrfs_free_cluster *cluster,
3389                    int delalloc)
3390         __acquires(&cluster->refill_lock)
3391 {
3392         struct btrfs_block_group *used_bg = NULL;
3393
3394         spin_lock(&cluster->refill_lock);
3395         while (1) {
3396                 used_bg = cluster->block_group;
3397                 if (!used_bg)
3398                         return NULL;
3399
3400                 if (used_bg == block_group)
3401                         return used_bg;
3402
3403                 btrfs_get_block_group(used_bg);
3404
3405                 if (!delalloc)
3406                         return used_bg;
3407
3408                 if (down_read_trylock(&used_bg->data_rwsem))
3409                         return used_bg;
3410
3411                 spin_unlock(&cluster->refill_lock);
3412
3413                 /* We should only have one-level nested. */
3414                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3415
3416                 spin_lock(&cluster->refill_lock);
3417                 if (used_bg == cluster->block_group)
3418                         return used_bg;
3419
3420                 up_read(&used_bg->data_rwsem);
3421                 btrfs_put_block_group(used_bg);
3422         }
3423 }
3424
3425 static inline void
3426 btrfs_release_block_group(struct btrfs_block_group *cache,
3427                          int delalloc)
3428 {
3429         if (delalloc)
3430                 up_read(&cache->data_rwsem);
3431         btrfs_put_block_group(cache);
3432 }
3433
3434 /*
3435  * Helper function for find_free_extent().
3436  *
3437  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3438  * Return -EAGAIN to inform caller that we need to re-search this block group
3439  * Return >0 to inform caller that we find nothing
3440  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3441  */
3442 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3443                                       struct find_free_extent_ctl *ffe_ctl,
3444                                       struct btrfs_block_group **cluster_bg_ret)
3445 {
3446         struct btrfs_block_group *cluster_bg;
3447         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3448         u64 aligned_cluster;
3449         u64 offset;
3450         int ret;
3451
3452         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3453         if (!cluster_bg)
3454                 goto refill_cluster;
3455         if (cluster_bg != bg && (cluster_bg->ro ||
3456             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3457                 goto release_cluster;
3458
3459         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3460                         ffe_ctl->num_bytes, cluster_bg->start,
3461                         &ffe_ctl->max_extent_size);
3462         if (offset) {
3463                 /* We have a block, we're done */
3464                 spin_unlock(&last_ptr->refill_lock);
3465                 trace_btrfs_reserve_extent_cluster(cluster_bg, ffe_ctl);
3466                 *cluster_bg_ret = cluster_bg;
3467                 ffe_ctl->found_offset = offset;
3468                 return 0;
3469         }
3470         WARN_ON(last_ptr->block_group != cluster_bg);
3471
3472 release_cluster:
3473         /*
3474          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3475          * lets just skip it and let the allocator find whatever block it can
3476          * find. If we reach this point, we will have tried the cluster
3477          * allocator plenty of times and not have found anything, so we are
3478          * likely way too fragmented for the clustering stuff to find anything.
3479          *
3480          * However, if the cluster is taken from the current block group,
3481          * release the cluster first, so that we stand a better chance of
3482          * succeeding in the unclustered allocation.
3483          */
3484         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3485                 spin_unlock(&last_ptr->refill_lock);
3486                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3487                 return -ENOENT;
3488         }
3489
3490         /* This cluster didn't work out, free it and start over */
3491         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3492
3493         if (cluster_bg != bg)
3494                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3495
3496 refill_cluster:
3497         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3498                 spin_unlock(&last_ptr->refill_lock);
3499                 return -ENOENT;
3500         }
3501
3502         aligned_cluster = max_t(u64,
3503                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3504                         bg->full_stripe_len);
3505         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3506                         ffe_ctl->num_bytes, aligned_cluster);
3507         if (ret == 0) {
3508                 /* Now pull our allocation out of this cluster */
3509                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3510                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3511                                 &ffe_ctl->max_extent_size);
3512                 if (offset) {
3513                         /* We found one, proceed */
3514                         spin_unlock(&last_ptr->refill_lock);
3515                         ffe_ctl->found_offset = offset;
3516                         trace_btrfs_reserve_extent_cluster(bg, ffe_ctl);
3517                         return 0;
3518                 }
3519         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
3520                    !ffe_ctl->retry_clustered) {
3521                 spin_unlock(&last_ptr->refill_lock);
3522
3523                 ffe_ctl->retry_clustered = true;
3524                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3525                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
3526                 return -EAGAIN;
3527         }
3528         /*
3529          * At this point we either didn't find a cluster or we weren't able to
3530          * allocate a block from our cluster.  Free the cluster we've been
3531          * trying to use, and go to the next block group.
3532          */
3533         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3534         spin_unlock(&last_ptr->refill_lock);
3535         return 1;
3536 }
3537
3538 /*
3539  * Return >0 to inform caller that we find nothing
3540  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3541  * Return -EAGAIN to inform caller that we need to re-search this block group
3542  */
3543 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3544                                         struct find_free_extent_ctl *ffe_ctl)
3545 {
3546         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3547         u64 offset;
3548
3549         /*
3550          * We are doing an unclustered allocation, set the fragmented flag so
3551          * we don't bother trying to setup a cluster again until we get more
3552          * space.
3553          */
3554         if (unlikely(last_ptr)) {
3555                 spin_lock(&last_ptr->lock);
3556                 last_ptr->fragmented = 1;
3557                 spin_unlock(&last_ptr->lock);
3558         }
3559         if (ffe_ctl->cached) {
3560                 struct btrfs_free_space_ctl *free_space_ctl;
3561
3562                 free_space_ctl = bg->free_space_ctl;
3563                 spin_lock(&free_space_ctl->tree_lock);
3564                 if (free_space_ctl->free_space <
3565                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3566                     ffe_ctl->empty_size) {
3567                         ffe_ctl->total_free_space = max_t(u64,
3568                                         ffe_ctl->total_free_space,
3569                                         free_space_ctl->free_space);
3570                         spin_unlock(&free_space_ctl->tree_lock);
3571                         return 1;
3572                 }
3573                 spin_unlock(&free_space_ctl->tree_lock);
3574         }
3575
3576         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3577                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3578                         &ffe_ctl->max_extent_size);
3579
3580         /*
3581          * If we didn't find a chunk, and we haven't failed on this block group
3582          * before, and this block group is in the middle of caching and we are
3583          * ok with waiting, then go ahead and wait for progress to be made, and
3584          * set @retry_unclustered to true.
3585          *
3586          * If @retry_unclustered is true then we've already waited on this
3587          * block group once and should move on to the next block group.
3588          */
3589         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
3590             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
3591                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3592                                                       ffe_ctl->empty_size);
3593                 ffe_ctl->retry_unclustered = true;
3594                 return -EAGAIN;
3595         } else if (!offset) {
3596                 return 1;
3597         }
3598         ffe_ctl->found_offset = offset;
3599         return 0;
3600 }
3601
3602 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3603                                    struct find_free_extent_ctl *ffe_ctl,
3604                                    struct btrfs_block_group **bg_ret)
3605 {
3606         int ret;
3607
3608         /* We want to try and use the cluster allocator, so lets look there */
3609         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3610                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3611                 if (ret >= 0 || ret == -EAGAIN)
3612                         return ret;
3613                 /* ret == -ENOENT case falls through */
3614         }
3615
3616         return find_free_extent_unclustered(block_group, ffe_ctl);
3617 }
3618
3619 /*
3620  * Tree-log block group locking
3621  * ============================
3622  *
3623  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3624  * indicates the starting address of a block group, which is reserved only
3625  * for tree-log metadata.
3626  *
3627  * Lock nesting
3628  * ============
3629  *
3630  * space_info::lock
3631  *   block_group::lock
3632  *     fs_info::treelog_bg_lock
3633  */
3634
3635 /*
3636  * Simple allocator for sequential-only block group. It only allows sequential
3637  * allocation. No need to play with trees. This function also reserves the
3638  * bytes as in btrfs_add_reserved_bytes.
3639  */
3640 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3641                                struct find_free_extent_ctl *ffe_ctl,
3642                                struct btrfs_block_group **bg_ret)
3643 {
3644         struct btrfs_fs_info *fs_info = block_group->fs_info;
3645         struct btrfs_space_info *space_info = block_group->space_info;
3646         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3647         u64 start = block_group->start;
3648         u64 num_bytes = ffe_ctl->num_bytes;
3649         u64 avail;
3650         u64 bytenr = block_group->start;
3651         u64 log_bytenr;
3652         u64 data_reloc_bytenr;
3653         int ret = 0;
3654         bool skip = false;
3655
3656         ASSERT(btrfs_is_zoned(block_group->fs_info));
3657
3658         /*
3659          * Do not allow non-tree-log blocks in the dedicated tree-log block
3660          * group, and vice versa.
3661          */
3662         spin_lock(&fs_info->treelog_bg_lock);
3663         log_bytenr = fs_info->treelog_bg;
3664         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3665                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3666                 skip = true;
3667         spin_unlock(&fs_info->treelog_bg_lock);
3668         if (skip)
3669                 return 1;
3670
3671         /*
3672          * Do not allow non-relocation blocks in the dedicated relocation block
3673          * group, and vice versa.
3674          */
3675         spin_lock(&fs_info->relocation_bg_lock);
3676         data_reloc_bytenr = fs_info->data_reloc_bg;
3677         if (data_reloc_bytenr &&
3678             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3679              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3680                 skip = true;
3681         spin_unlock(&fs_info->relocation_bg_lock);
3682         if (skip)
3683                 return 1;
3684
3685         /* Check RO and no space case before trying to activate it */
3686         spin_lock(&block_group->lock);
3687         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3688                 ret = 1;
3689                 /*
3690                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3691                  * Return the error after taking the locks.
3692                  */
3693         }
3694         spin_unlock(&block_group->lock);
3695
3696         if (!ret && !btrfs_zone_activate(block_group)) {
3697                 ret = 1;
3698                 /*
3699                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3700                  * Return the error after taking the locks.
3701                  */
3702         }
3703
3704         spin_lock(&space_info->lock);
3705         spin_lock(&block_group->lock);
3706         spin_lock(&fs_info->treelog_bg_lock);
3707         spin_lock(&fs_info->relocation_bg_lock);
3708
3709         if (ret)
3710                 goto out;
3711
3712         ASSERT(!ffe_ctl->for_treelog ||
3713                block_group->start == fs_info->treelog_bg ||
3714                fs_info->treelog_bg == 0);
3715         ASSERT(!ffe_ctl->for_data_reloc ||
3716                block_group->start == fs_info->data_reloc_bg ||
3717                fs_info->data_reloc_bg == 0);
3718
3719         if (block_group->ro ||
3720             test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
3721                 ret = 1;
3722                 goto out;
3723         }
3724
3725         /*
3726          * Do not allow currently using block group to be tree-log dedicated
3727          * block group.
3728          */
3729         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3730             (block_group->used || block_group->reserved)) {
3731                 ret = 1;
3732                 goto out;
3733         }
3734
3735         /*
3736          * Do not allow currently used block group to be the data relocation
3737          * dedicated block group.
3738          */
3739         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3740             (block_group->used || block_group->reserved)) {
3741                 ret = 1;
3742                 goto out;
3743         }
3744
3745         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3746         avail = block_group->zone_capacity - block_group->alloc_offset;
3747         if (avail < num_bytes) {
3748                 if (ffe_ctl->max_extent_size < avail) {
3749                         /*
3750                          * With sequential allocator, free space is always
3751                          * contiguous
3752                          */
3753                         ffe_ctl->max_extent_size = avail;
3754                         ffe_ctl->total_free_space = avail;
3755                 }
3756                 ret = 1;
3757                 goto out;
3758         }
3759
3760         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3761                 fs_info->treelog_bg = block_group->start;
3762
3763         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg)
3764                 fs_info->data_reloc_bg = block_group->start;
3765
3766         ffe_ctl->found_offset = start + block_group->alloc_offset;
3767         block_group->alloc_offset += num_bytes;
3768         spin_lock(&ctl->tree_lock);
3769         ctl->free_space -= num_bytes;
3770         spin_unlock(&ctl->tree_lock);
3771
3772         /*
3773          * We do not check if found_offset is aligned to stripesize. The
3774          * address is anyway rewritten when using zone append writing.
3775          */
3776
3777         ffe_ctl->search_start = ffe_ctl->found_offset;
3778
3779 out:
3780         if (ret && ffe_ctl->for_treelog)
3781                 fs_info->treelog_bg = 0;
3782         if (ret && ffe_ctl->for_data_reloc &&
3783             fs_info->data_reloc_bg == block_group->start) {
3784                 /*
3785                  * Do not allow further allocations from this block group.
3786                  * Compared to increasing the ->ro, setting the
3787                  * ->zoned_data_reloc_ongoing flag still allows nocow
3788                  *  writers to come in. See btrfs_inc_nocow_writers().
3789                  *
3790                  * We need to disable an allocation to avoid an allocation of
3791                  * regular (non-relocation data) extent. With mix of relocation
3792                  * extents and regular extents, we can dispatch WRITE commands
3793                  * (for relocation extents) and ZONE APPEND commands (for
3794                  * regular extents) at the same time to the same zone, which
3795                  * easily break the write pointer.
3796                  */
3797                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
3798                 fs_info->data_reloc_bg = 0;
3799         }
3800         spin_unlock(&fs_info->relocation_bg_lock);
3801         spin_unlock(&fs_info->treelog_bg_lock);
3802         spin_unlock(&block_group->lock);
3803         spin_unlock(&space_info->lock);
3804         return ret;
3805 }
3806
3807 static int do_allocation(struct btrfs_block_group *block_group,
3808                          struct find_free_extent_ctl *ffe_ctl,
3809                          struct btrfs_block_group **bg_ret)
3810 {
3811         switch (ffe_ctl->policy) {
3812         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3813                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3814         case BTRFS_EXTENT_ALLOC_ZONED:
3815                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3816         default:
3817                 BUG();
3818         }
3819 }
3820
3821 static void release_block_group(struct btrfs_block_group *block_group,
3822                                 struct find_free_extent_ctl *ffe_ctl,
3823                                 int delalloc)
3824 {
3825         switch (ffe_ctl->policy) {
3826         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3827                 ffe_ctl->retry_clustered = false;
3828                 ffe_ctl->retry_unclustered = false;
3829                 break;
3830         case BTRFS_EXTENT_ALLOC_ZONED:
3831                 /* Nothing to do */
3832                 break;
3833         default:
3834                 BUG();
3835         }
3836
3837         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3838                ffe_ctl->index);
3839         btrfs_release_block_group(block_group, delalloc);
3840 }
3841
3842 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3843                                    struct btrfs_key *ins)
3844 {
3845         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3846
3847         if (!ffe_ctl->use_cluster && last_ptr) {
3848                 spin_lock(&last_ptr->lock);
3849                 last_ptr->window_start = ins->objectid;
3850                 spin_unlock(&last_ptr->lock);
3851         }
3852 }
3853
3854 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3855                          struct btrfs_key *ins)
3856 {
3857         switch (ffe_ctl->policy) {
3858         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3859                 found_extent_clustered(ffe_ctl, ins);
3860                 break;
3861         case BTRFS_EXTENT_ALLOC_ZONED:
3862                 /* Nothing to do */
3863                 break;
3864         default:
3865                 BUG();
3866         }
3867 }
3868
3869 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
3870                                     struct find_free_extent_ctl *ffe_ctl)
3871 {
3872         /* If we can activate new zone, just allocate a chunk and use it */
3873         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
3874                 return 0;
3875
3876         /*
3877          * We already reached the max active zones. Try to finish one block
3878          * group to make a room for a new block group. This is only possible
3879          * for a data block group because btrfs_zone_finish() may need to wait
3880          * for a running transaction which can cause a deadlock for metadata
3881          * allocation.
3882          */
3883         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
3884                 int ret = btrfs_zone_finish_one_bg(fs_info);
3885
3886                 if (ret == 1)
3887                         return 0;
3888                 else if (ret < 0)
3889                         return ret;
3890         }
3891
3892         /*
3893          * If we have enough free space left in an already active block group
3894          * and we can't activate any other zone now, do not allow allocating a
3895          * new chunk and let find_free_extent() retry with a smaller size.
3896          */
3897         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
3898                 return -ENOSPC;
3899
3900         /*
3901          * Even min_alloc_size is not left in any block groups. Since we cannot
3902          * activate a new block group, allocating it may not help. Let's tell a
3903          * caller to try again and hope it progress something by writing some
3904          * parts of the region. That is only possible for data block groups,
3905          * where a part of the region can be written.
3906          */
3907         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
3908                 return -EAGAIN;
3909
3910         /*
3911          * We cannot activate a new block group and no enough space left in any
3912          * block groups. So, allocating a new block group may not help. But,
3913          * there is nothing to do anyway, so let's go with it.
3914          */
3915         return 0;
3916 }
3917
3918 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
3919                               struct find_free_extent_ctl *ffe_ctl)
3920 {
3921         switch (ffe_ctl->policy) {
3922         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3923                 return 0;
3924         case BTRFS_EXTENT_ALLOC_ZONED:
3925                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
3926         default:
3927                 BUG();
3928         }
3929 }
3930
3931 /*
3932  * Return >0 means caller needs to re-search for free extent
3933  * Return 0 means we have the needed free extent.
3934  * Return <0 means we failed to locate any free extent.
3935  */
3936 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
3937                                         struct btrfs_key *ins,
3938                                         struct find_free_extent_ctl *ffe_ctl,
3939                                         bool full_search)
3940 {
3941         struct btrfs_root *root = fs_info->chunk_root;
3942         int ret;
3943
3944         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
3945             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
3946                 ffe_ctl->orig_have_caching_bg = true;
3947
3948         if (ins->objectid) {
3949                 found_extent(ffe_ctl, ins);
3950                 return 0;
3951         }
3952
3953         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
3954                 return 1;
3955
3956         ffe_ctl->index++;
3957         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
3958                 return 1;
3959
3960         /*
3961          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
3962          *                      caching kthreads as we move along
3963          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
3964          * LOOP_UNSET_SIZE_CLASS, allow unset size class
3965          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
3966          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
3967          *                     again
3968          */
3969         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
3970                 ffe_ctl->index = 0;
3971                 /*
3972                  * We want to skip the LOOP_CACHING_WAIT step if we don't have
3973                  * any uncached bgs and we've already done a full search
3974                  * through.
3975                  */
3976                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT &&
3977                     (!ffe_ctl->orig_have_caching_bg && full_search))
3978                         ffe_ctl->loop++;
3979                 ffe_ctl->loop++;
3980
3981                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
3982                         struct btrfs_trans_handle *trans;
3983                         int exist = 0;
3984
3985                         /* Check if allocation policy allows to create a new chunk */
3986                         ret = can_allocate_chunk(fs_info, ffe_ctl);
3987                         if (ret)
3988                                 return ret;
3989
3990                         trans = current->journal_info;
3991                         if (trans)
3992                                 exist = 1;
3993                         else
3994                                 trans = btrfs_join_transaction(root);
3995
3996                         if (IS_ERR(trans)) {
3997                                 ret = PTR_ERR(trans);
3998                                 return ret;
3999                         }
4000
4001                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4002                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4003
4004                         /* Do not bail out on ENOSPC since we can do more. */
4005                         if (ret == -ENOSPC) {
4006                                 ret = 0;
4007                                 ffe_ctl->loop++;
4008                         }
4009                         else if (ret < 0)
4010                                 btrfs_abort_transaction(trans, ret);
4011                         else
4012                                 ret = 0;
4013                         if (!exist)
4014                                 btrfs_end_transaction(trans);
4015                         if (ret)
4016                                 return ret;
4017                 }
4018
4019                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4020                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4021                                 return -ENOSPC;
4022
4023                         /*
4024                          * Don't loop again if we already have no empty_size and
4025                          * no empty_cluster.
4026                          */
4027                         if (ffe_ctl->empty_size == 0 &&
4028                             ffe_ctl->empty_cluster == 0)
4029                                 return -ENOSPC;
4030                         ffe_ctl->empty_size = 0;
4031                         ffe_ctl->empty_cluster = 0;
4032                 }
4033                 return 1;
4034         }
4035         return -ENOSPC;
4036 }
4037
4038 static bool find_free_extent_check_size_class(struct find_free_extent_ctl *ffe_ctl,
4039                                               struct btrfs_block_group *bg)
4040 {
4041         if (ffe_ctl->policy == BTRFS_EXTENT_ALLOC_ZONED)
4042                 return true;
4043         if (!btrfs_block_group_should_use_size_class(bg))
4044                 return true;
4045         if (ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS)
4046                 return true;
4047         if (ffe_ctl->loop >= LOOP_UNSET_SIZE_CLASS &&
4048             bg->size_class == BTRFS_BG_SZ_NONE)
4049                 return true;
4050         return ffe_ctl->size_class == bg->size_class;
4051 }
4052
4053 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4054                                         struct find_free_extent_ctl *ffe_ctl,
4055                                         struct btrfs_space_info *space_info,
4056                                         struct btrfs_key *ins)
4057 {
4058         /*
4059          * If our free space is heavily fragmented we may not be able to make
4060          * big contiguous allocations, so instead of doing the expensive search
4061          * for free space, simply return ENOSPC with our max_extent_size so we
4062          * can go ahead and search for a more manageable chunk.
4063          *
4064          * If our max_extent_size is large enough for our allocation simply
4065          * disable clustering since we will likely not be able to find enough
4066          * space to create a cluster and induce latency trying.
4067          */
4068         if (space_info->max_extent_size) {
4069                 spin_lock(&space_info->lock);
4070                 if (space_info->max_extent_size &&
4071                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4072                         ins->offset = space_info->max_extent_size;
4073                         spin_unlock(&space_info->lock);
4074                         return -ENOSPC;
4075                 } else if (space_info->max_extent_size) {
4076                         ffe_ctl->use_cluster = false;
4077                 }
4078                 spin_unlock(&space_info->lock);
4079         }
4080
4081         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4082                                                &ffe_ctl->empty_cluster);
4083         if (ffe_ctl->last_ptr) {
4084                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4085
4086                 spin_lock(&last_ptr->lock);
4087                 if (last_ptr->block_group)
4088                         ffe_ctl->hint_byte = last_ptr->window_start;
4089                 if (last_ptr->fragmented) {
4090                         /*
4091                          * We still set window_start so we can keep track of the
4092                          * last place we found an allocation to try and save
4093                          * some time.
4094                          */
4095                         ffe_ctl->hint_byte = last_ptr->window_start;
4096                         ffe_ctl->use_cluster = false;
4097                 }
4098                 spin_unlock(&last_ptr->lock);
4099         }
4100
4101         return 0;
4102 }
4103
4104 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4105                               struct find_free_extent_ctl *ffe_ctl,
4106                               struct btrfs_space_info *space_info,
4107                               struct btrfs_key *ins)
4108 {
4109         switch (ffe_ctl->policy) {
4110         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4111                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4112                                                     space_info, ins);
4113         case BTRFS_EXTENT_ALLOC_ZONED:
4114                 if (ffe_ctl->for_treelog) {
4115                         spin_lock(&fs_info->treelog_bg_lock);
4116                         if (fs_info->treelog_bg)
4117                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4118                         spin_unlock(&fs_info->treelog_bg_lock);
4119                 }
4120                 if (ffe_ctl->for_data_reloc) {
4121                         spin_lock(&fs_info->relocation_bg_lock);
4122                         if (fs_info->data_reloc_bg)
4123                                 ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4124                         spin_unlock(&fs_info->relocation_bg_lock);
4125                 }
4126                 return 0;
4127         default:
4128                 BUG();
4129         }
4130 }
4131
4132 /*
4133  * walks the btree of allocated extents and find a hole of a given size.
4134  * The key ins is changed to record the hole:
4135  * ins->objectid == start position
4136  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4137  * ins->offset == the size of the hole.
4138  * Any available blocks before search_start are skipped.
4139  *
4140  * If there is no suitable free space, we will record the max size of
4141  * the free space extent currently.
4142  *
4143  * The overall logic and call chain:
4144  *
4145  * find_free_extent()
4146  * |- Iterate through all block groups
4147  * |  |- Get a valid block group
4148  * |  |- Try to do clustered allocation in that block group
4149  * |  |- Try to do unclustered allocation in that block group
4150  * |  |- Check if the result is valid
4151  * |  |  |- If valid, then exit
4152  * |  |- Jump to next block group
4153  * |
4154  * |- Push harder to find free extents
4155  *    |- If not found, re-iterate all block groups
4156  */
4157 static noinline int find_free_extent(struct btrfs_root *root,
4158                                      struct btrfs_key *ins,
4159                                      struct find_free_extent_ctl *ffe_ctl)
4160 {
4161         struct btrfs_fs_info *fs_info = root->fs_info;
4162         int ret = 0;
4163         int cache_block_group_error = 0;
4164         struct btrfs_block_group *block_group = NULL;
4165         struct btrfs_space_info *space_info;
4166         bool full_search = false;
4167
4168         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4169
4170         ffe_ctl->search_start = 0;
4171         /* For clustered allocation */
4172         ffe_ctl->empty_cluster = 0;
4173         ffe_ctl->last_ptr = NULL;
4174         ffe_ctl->use_cluster = true;
4175         ffe_ctl->have_caching_bg = false;
4176         ffe_ctl->orig_have_caching_bg = false;
4177         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4178         ffe_ctl->loop = 0;
4179         /* For clustered allocation */
4180         ffe_ctl->retry_clustered = false;
4181         ffe_ctl->retry_unclustered = false;
4182         ffe_ctl->cached = 0;
4183         ffe_ctl->max_extent_size = 0;
4184         ffe_ctl->total_free_space = 0;
4185         ffe_ctl->found_offset = 0;
4186         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4187         ffe_ctl->size_class = btrfs_calc_block_group_size_class(ffe_ctl->num_bytes);
4188
4189         if (btrfs_is_zoned(fs_info))
4190                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4191
4192         ins->type = BTRFS_EXTENT_ITEM_KEY;
4193         ins->objectid = 0;
4194         ins->offset = 0;
4195
4196         trace_find_free_extent(root, ffe_ctl);
4197
4198         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4199         if (!space_info) {
4200                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4201                 return -ENOSPC;
4202         }
4203
4204         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4205         if (ret < 0)
4206                 return ret;
4207
4208         ffe_ctl->search_start = max(ffe_ctl->search_start,
4209                                     first_logical_byte(fs_info));
4210         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4211         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4212                 block_group = btrfs_lookup_block_group(fs_info,
4213                                                        ffe_ctl->search_start);
4214                 /*
4215                  * we don't want to use the block group if it doesn't match our
4216                  * allocation bits, or if its not cached.
4217                  *
4218                  * However if we are re-searching with an ideal block group
4219                  * picked out then we don't care that the block group is cached.
4220                  */
4221                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4222                     block_group->cached != BTRFS_CACHE_NO) {
4223                         down_read(&space_info->groups_sem);
4224                         if (list_empty(&block_group->list) ||
4225                             block_group->ro) {
4226                                 /*
4227                                  * someone is removing this block group,
4228                                  * we can't jump into the have_block_group
4229                                  * target because our list pointers are not
4230                                  * valid
4231                                  */
4232                                 btrfs_put_block_group(block_group);
4233                                 up_read(&space_info->groups_sem);
4234                         } else {
4235                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4236                                                         block_group->flags);
4237                                 btrfs_lock_block_group(block_group,
4238                                                        ffe_ctl->delalloc);
4239                                 ffe_ctl->hinted = true;
4240                                 goto have_block_group;
4241                         }
4242                 } else if (block_group) {
4243                         btrfs_put_block_group(block_group);
4244                 }
4245         }
4246 search:
4247         trace_find_free_extent_search_loop(root, ffe_ctl);
4248         ffe_ctl->have_caching_bg = false;
4249         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4250             ffe_ctl->index == 0)
4251                 full_search = true;
4252         down_read(&space_info->groups_sem);
4253         list_for_each_entry(block_group,
4254                             &space_info->block_groups[ffe_ctl->index], list) {
4255                 struct btrfs_block_group *bg_ret;
4256
4257                 ffe_ctl->hinted = false;
4258                 /* If the block group is read-only, we can skip it entirely. */
4259                 if (unlikely(block_group->ro)) {
4260                         if (ffe_ctl->for_treelog)
4261                                 btrfs_clear_treelog_bg(block_group);
4262                         if (ffe_ctl->for_data_reloc)
4263                                 btrfs_clear_data_reloc_bg(block_group);
4264                         continue;
4265                 }
4266
4267                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4268                 ffe_ctl->search_start = block_group->start;
4269
4270                 /*
4271                  * this can happen if we end up cycling through all the
4272                  * raid types, but we want to make sure we only allocate
4273                  * for the proper type.
4274                  */
4275                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4276                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4277                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4278                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4279                                 BTRFS_BLOCK_GROUP_RAID10;
4280
4281                         /*
4282                          * if they asked for extra copies and this block group
4283                          * doesn't provide them, bail.  This does allow us to
4284                          * fill raid0 from raid1.
4285                          */
4286                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4287                                 goto loop;
4288
4289                         /*
4290                          * This block group has different flags than we want.
4291                          * It's possible that we have MIXED_GROUP flag but no
4292                          * block group is mixed.  Just skip such block group.
4293                          */
4294                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4295                         continue;
4296                 }
4297
4298 have_block_group:
4299                 trace_find_free_extent_have_block_group(root, ffe_ctl, block_group);
4300                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4301                 if (unlikely(!ffe_ctl->cached)) {
4302                         ffe_ctl->have_caching_bg = true;
4303                         ret = btrfs_cache_block_group(block_group, false);
4304
4305                         /*
4306                          * If we get ENOMEM here or something else we want to
4307                          * try other block groups, because it may not be fatal.
4308                          * However if we can't find anything else we need to
4309                          * save our return here so that we return the actual
4310                          * error that caused problems, not ENOSPC.
4311                          */
4312                         if (ret < 0) {
4313                                 if (!cache_block_group_error)
4314                                         cache_block_group_error = ret;
4315                                 ret = 0;
4316                                 goto loop;
4317                         }
4318                         ret = 0;
4319                 }
4320
4321                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
4322                         goto loop;
4323
4324                 if (!find_free_extent_check_size_class(ffe_ctl, block_group))
4325                         goto loop;
4326
4327                 bg_ret = NULL;
4328                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4329                 if (ret == 0) {
4330                         if (bg_ret && bg_ret != block_group) {
4331                                 btrfs_release_block_group(block_group,
4332                                                           ffe_ctl->delalloc);
4333                                 block_group = bg_ret;
4334                         }
4335                 } else if (ret == -EAGAIN) {
4336                         goto have_block_group;
4337                 } else if (ret > 0) {
4338                         goto loop;
4339                 }
4340
4341                 /* Checks */
4342                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4343                                                  fs_info->stripesize);
4344
4345                 /* move on to the next group */
4346                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4347                     block_group->start + block_group->length) {
4348                         btrfs_add_free_space_unused(block_group,
4349                                             ffe_ctl->found_offset,
4350                                             ffe_ctl->num_bytes);
4351                         goto loop;
4352                 }
4353
4354                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4355                         btrfs_add_free_space_unused(block_group,
4356                                         ffe_ctl->found_offset,
4357                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4358
4359                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4360                                                ffe_ctl->num_bytes,
4361                                                ffe_ctl->delalloc,
4362                                                ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS);
4363                 if (ret == -EAGAIN) {
4364                         btrfs_add_free_space_unused(block_group,
4365                                         ffe_ctl->found_offset,
4366                                         ffe_ctl->num_bytes);
4367                         goto loop;
4368                 }
4369                 btrfs_inc_block_group_reservations(block_group);
4370
4371                 /* we are all good, lets return */
4372                 ins->objectid = ffe_ctl->search_start;
4373                 ins->offset = ffe_ctl->num_bytes;
4374
4375                 trace_btrfs_reserve_extent(block_group, ffe_ctl);
4376                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4377                 break;
4378 loop:
4379                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4380                 cond_resched();
4381         }
4382         up_read(&space_info->groups_sem);
4383
4384         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4385         if (ret > 0)
4386                 goto search;
4387
4388         if (ret == -ENOSPC && !cache_block_group_error) {
4389                 /*
4390                  * Use ffe_ctl->total_free_space as fallback if we can't find
4391                  * any contiguous hole.
4392                  */
4393                 if (!ffe_ctl->max_extent_size)
4394                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4395                 spin_lock(&space_info->lock);
4396                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4397                 spin_unlock(&space_info->lock);
4398                 ins->offset = ffe_ctl->max_extent_size;
4399         } else if (ret == -ENOSPC) {
4400                 ret = cache_block_group_error;
4401         }
4402         return ret;
4403 }
4404
4405 /*
4406  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4407  *                        hole that is at least as big as @num_bytes.
4408  *
4409  * @root           -    The root that will contain this extent
4410  *
4411  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4412  *                      is used for accounting purposes. This value differs
4413  *                      from @num_bytes only in the case of compressed extents.
4414  *
4415  * @num_bytes      -    Number of bytes to allocate on-disk.
4416  *
4417  * @min_alloc_size -    Indicates the minimum amount of space that the
4418  *                      allocator should try to satisfy. In some cases
4419  *                      @num_bytes may be larger than what is required and if
4420  *                      the filesystem is fragmented then allocation fails.
4421  *                      However, the presence of @min_alloc_size gives a
4422  *                      chance to try and satisfy the smaller allocation.
4423  *
4424  * @empty_size     -    A hint that you plan on doing more COW. This is the
4425  *                      size in bytes the allocator should try to find free
4426  *                      next to the block it returns.  This is just a hint and
4427  *                      may be ignored by the allocator.
4428  *
4429  * @hint_byte      -    Hint to the allocator to start searching above the byte
4430  *                      address passed. It might be ignored.
4431  *
4432  * @ins            -    This key is modified to record the found hole. It will
4433  *                      have the following values:
4434  *                      ins->objectid == start position
4435  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4436  *                      ins->offset == the size of the hole.
4437  *
4438  * @is_data        -    Boolean flag indicating whether an extent is
4439  *                      allocated for data (true) or metadata (false)
4440  *
4441  * @delalloc       -    Boolean flag indicating whether this allocation is for
4442  *                      delalloc or not. If 'true' data_rwsem of block groups
4443  *                      is going to be acquired.
4444  *
4445  *
4446  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4447  * case -ENOSPC is returned then @ins->offset will contain the size of the
4448  * largest available hole the allocator managed to find.
4449  */
4450 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4451                          u64 num_bytes, u64 min_alloc_size,
4452                          u64 empty_size, u64 hint_byte,
4453                          struct btrfs_key *ins, int is_data, int delalloc)
4454 {
4455         struct btrfs_fs_info *fs_info = root->fs_info;
4456         struct find_free_extent_ctl ffe_ctl = {};
4457         bool final_tried = num_bytes == min_alloc_size;
4458         u64 flags;
4459         int ret;
4460         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4461         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4462
4463         flags = get_alloc_profile_by_root(root, is_data);
4464 again:
4465         WARN_ON(num_bytes < fs_info->sectorsize);
4466
4467         ffe_ctl.ram_bytes = ram_bytes;
4468         ffe_ctl.num_bytes = num_bytes;
4469         ffe_ctl.min_alloc_size = min_alloc_size;
4470         ffe_ctl.empty_size = empty_size;
4471         ffe_ctl.flags = flags;
4472         ffe_ctl.delalloc = delalloc;
4473         ffe_ctl.hint_byte = hint_byte;
4474         ffe_ctl.for_treelog = for_treelog;
4475         ffe_ctl.for_data_reloc = for_data_reloc;
4476
4477         ret = find_free_extent(root, ins, &ffe_ctl);
4478         if (!ret && !is_data) {
4479                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4480         } else if (ret == -ENOSPC) {
4481                 if (!final_tried && ins->offset) {
4482                         num_bytes = min(num_bytes >> 1, ins->offset);
4483                         num_bytes = round_down(num_bytes,
4484                                                fs_info->sectorsize);
4485                         num_bytes = max(num_bytes, min_alloc_size);
4486                         ram_bytes = num_bytes;
4487                         if (num_bytes == min_alloc_size)
4488                                 final_tried = true;
4489                         goto again;
4490                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4491                         struct btrfs_space_info *sinfo;
4492
4493                         sinfo = btrfs_find_space_info(fs_info, flags);
4494                         btrfs_err(fs_info,
4495         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4496                                   flags, num_bytes, for_treelog, for_data_reloc);
4497                         if (sinfo)
4498                                 btrfs_dump_space_info(fs_info, sinfo,
4499                                                       num_bytes, 1);
4500                 }
4501         }
4502
4503         return ret;
4504 }
4505
4506 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4507                                u64 start, u64 len, int delalloc)
4508 {
4509         struct btrfs_block_group *cache;
4510
4511         cache = btrfs_lookup_block_group(fs_info, start);
4512         if (!cache) {
4513                 btrfs_err(fs_info, "Unable to find block group for %llu",
4514                           start);
4515                 return -ENOSPC;
4516         }
4517
4518         btrfs_add_free_space(cache, start, len);
4519         btrfs_free_reserved_bytes(cache, len, delalloc);
4520         trace_btrfs_reserved_extent_free(fs_info, start, len);
4521
4522         btrfs_put_block_group(cache);
4523         return 0;
4524 }
4525
4526 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4527                               u64 len)
4528 {
4529         struct btrfs_block_group *cache;
4530         int ret = 0;
4531
4532         cache = btrfs_lookup_block_group(trans->fs_info, start);
4533         if (!cache) {
4534                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4535                           start);
4536                 return -ENOSPC;
4537         }
4538
4539         ret = pin_down_extent(trans, cache, start, len, 1);
4540         btrfs_put_block_group(cache);
4541         return ret;
4542 }
4543
4544 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4545                                  u64 num_bytes)
4546 {
4547         struct btrfs_fs_info *fs_info = trans->fs_info;
4548         int ret;
4549
4550         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4551         if (ret)
4552                 return ret;
4553
4554         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4555         if (ret) {
4556                 ASSERT(!ret);
4557                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4558                           bytenr, num_bytes);
4559                 return ret;
4560         }
4561
4562         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4563         return 0;
4564 }
4565
4566 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4567                                       u64 parent, u64 root_objectid,
4568                                       u64 flags, u64 owner, u64 offset,
4569                                       struct btrfs_key *ins, int ref_mod)
4570 {
4571         struct btrfs_fs_info *fs_info = trans->fs_info;
4572         struct btrfs_root *extent_root;
4573         int ret;
4574         struct btrfs_extent_item *extent_item;
4575         struct btrfs_extent_inline_ref *iref;
4576         struct btrfs_path *path;
4577         struct extent_buffer *leaf;
4578         int type;
4579         u32 size;
4580
4581         if (parent > 0)
4582                 type = BTRFS_SHARED_DATA_REF_KEY;
4583         else
4584                 type = BTRFS_EXTENT_DATA_REF_KEY;
4585
4586         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4587
4588         path = btrfs_alloc_path();
4589         if (!path)
4590                 return -ENOMEM;
4591
4592         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4593         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4594         if (ret) {
4595                 btrfs_free_path(path);
4596                 return ret;
4597         }
4598
4599         leaf = path->nodes[0];
4600         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4601                                      struct btrfs_extent_item);
4602         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4603         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4604         btrfs_set_extent_flags(leaf, extent_item,
4605                                flags | BTRFS_EXTENT_FLAG_DATA);
4606
4607         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4608         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4609         if (parent > 0) {
4610                 struct btrfs_shared_data_ref *ref;
4611                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4612                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4613                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4614         } else {
4615                 struct btrfs_extent_data_ref *ref;
4616                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4617                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4618                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4619                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4620                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4621         }
4622
4623         btrfs_mark_buffer_dirty(path->nodes[0]);
4624         btrfs_free_path(path);
4625
4626         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4627 }
4628
4629 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4630                                      struct btrfs_delayed_ref_node *node,
4631                                      struct btrfs_delayed_extent_op *extent_op)
4632 {
4633         struct btrfs_fs_info *fs_info = trans->fs_info;
4634         struct btrfs_root *extent_root;
4635         int ret;
4636         struct btrfs_extent_item *extent_item;
4637         struct btrfs_key extent_key;
4638         struct btrfs_tree_block_info *block_info;
4639         struct btrfs_extent_inline_ref *iref;
4640         struct btrfs_path *path;
4641         struct extent_buffer *leaf;
4642         struct btrfs_delayed_tree_ref *ref;
4643         u32 size = sizeof(*extent_item) + sizeof(*iref);
4644         u64 flags = extent_op->flags_to_set;
4645         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4646
4647         ref = btrfs_delayed_node_to_tree_ref(node);
4648
4649         extent_key.objectid = node->bytenr;
4650         if (skinny_metadata) {
4651                 extent_key.offset = ref->level;
4652                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4653         } else {
4654                 extent_key.offset = node->num_bytes;
4655                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4656                 size += sizeof(*block_info);
4657         }
4658
4659         path = btrfs_alloc_path();
4660         if (!path)
4661                 return -ENOMEM;
4662
4663         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4664         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4665                                       size);
4666         if (ret) {
4667                 btrfs_free_path(path);
4668                 return ret;
4669         }
4670
4671         leaf = path->nodes[0];
4672         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4673                                      struct btrfs_extent_item);
4674         btrfs_set_extent_refs(leaf, extent_item, 1);
4675         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4676         btrfs_set_extent_flags(leaf, extent_item,
4677                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4678
4679         if (skinny_metadata) {
4680                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4681         } else {
4682                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4683                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4684                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4685                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4686         }
4687
4688         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4689                 btrfs_set_extent_inline_ref_type(leaf, iref,
4690                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4691                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4692         } else {
4693                 btrfs_set_extent_inline_ref_type(leaf, iref,
4694                                                  BTRFS_TREE_BLOCK_REF_KEY);
4695                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4696         }
4697
4698         btrfs_mark_buffer_dirty(leaf);
4699         btrfs_free_path(path);
4700
4701         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4702 }
4703
4704 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4705                                      struct btrfs_root *root, u64 owner,
4706                                      u64 offset, u64 ram_bytes,
4707                                      struct btrfs_key *ins)
4708 {
4709         struct btrfs_ref generic_ref = { 0 };
4710
4711         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4712
4713         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4714                                ins->objectid, ins->offset, 0);
4715         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner,
4716                             offset, 0, false);
4717         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4718
4719         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4720 }
4721
4722 /*
4723  * this is used by the tree logging recovery code.  It records that
4724  * an extent has been allocated and makes sure to clear the free
4725  * space cache bits as well
4726  */
4727 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4728                                    u64 root_objectid, u64 owner, u64 offset,
4729                                    struct btrfs_key *ins)
4730 {
4731         struct btrfs_fs_info *fs_info = trans->fs_info;
4732         int ret;
4733         struct btrfs_block_group *block_group;
4734         struct btrfs_space_info *space_info;
4735
4736         /*
4737          * Mixed block groups will exclude before processing the log so we only
4738          * need to do the exclude dance if this fs isn't mixed.
4739          */
4740         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4741                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4742                                               ins->offset);
4743                 if (ret)
4744                         return ret;
4745         }
4746
4747         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4748         if (!block_group)
4749                 return -EINVAL;
4750
4751         space_info = block_group->space_info;
4752         spin_lock(&space_info->lock);
4753         spin_lock(&block_group->lock);
4754         space_info->bytes_reserved += ins->offset;
4755         block_group->reserved += ins->offset;
4756         spin_unlock(&block_group->lock);
4757         spin_unlock(&space_info->lock);
4758
4759         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4760                                          offset, ins, 1);
4761         if (ret)
4762                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4763         btrfs_put_block_group(block_group);
4764         return ret;
4765 }
4766
4767 static struct extent_buffer *
4768 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4769                       u64 bytenr, int level, u64 owner,
4770                       enum btrfs_lock_nesting nest)
4771 {
4772         struct btrfs_fs_info *fs_info = root->fs_info;
4773         struct extent_buffer *buf;
4774         u64 lockdep_owner = owner;
4775
4776         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
4777         if (IS_ERR(buf))
4778                 return buf;
4779
4780         /*
4781          * Extra safety check in case the extent tree is corrupted and extent
4782          * allocator chooses to use a tree block which is already used and
4783          * locked.
4784          */
4785         if (buf->lock_owner == current->pid) {
4786                 btrfs_err_rl(fs_info,
4787 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4788                         buf->start, btrfs_header_owner(buf), current->pid);
4789                 free_extent_buffer(buf);
4790                 return ERR_PTR(-EUCLEAN);
4791         }
4792
4793         /*
4794          * The reloc trees are just snapshots, so we need them to appear to be
4795          * just like any other fs tree WRT lockdep.
4796          *
4797          * The exception however is in replace_path() in relocation, where we
4798          * hold the lock on the original fs root and then search for the reloc
4799          * root.  At that point we need to make sure any reloc root buffers are
4800          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
4801          * lockdep happy.
4802          */
4803         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
4804             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
4805                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
4806
4807         /* btrfs_clean_tree_block() accesses generation field. */
4808         btrfs_set_header_generation(buf, trans->transid);
4809
4810         /*
4811          * This needs to stay, because we could allocate a freed block from an
4812          * old tree into a new tree, so we need to make sure this new block is
4813          * set to the appropriate level and owner.
4814          */
4815         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
4816
4817         __btrfs_tree_lock(buf, nest);
4818         btrfs_clear_buffer_dirty(trans, buf);
4819         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4820         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
4821
4822         set_extent_buffer_uptodate(buf);
4823
4824         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4825         btrfs_set_header_level(buf, level);
4826         btrfs_set_header_bytenr(buf, buf->start);
4827         btrfs_set_header_generation(buf, trans->transid);
4828         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4829         btrfs_set_header_owner(buf, owner);
4830         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4831         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4832         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4833                 buf->log_index = root->log_transid % 2;
4834                 /*
4835                  * we allow two log transactions at a time, use different
4836                  * EXTENT bit to differentiate dirty pages.
4837                  */
4838                 if (buf->log_index == 0)
4839                         set_extent_dirty(&root->dirty_log_pages, buf->start,
4840                                         buf->start + buf->len - 1, GFP_NOFS);
4841                 else
4842                         set_extent_new(&root->dirty_log_pages, buf->start,
4843                                         buf->start + buf->len - 1);
4844         } else {
4845                 buf->log_index = -1;
4846                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
4847                          buf->start + buf->len - 1, GFP_NOFS);
4848         }
4849         /* this returns a buffer locked for blocking */
4850         return buf;
4851 }
4852
4853 /*
4854  * finds a free extent and does all the dirty work required for allocation
4855  * returns the tree buffer or an ERR_PTR on error.
4856  */
4857 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4858                                              struct btrfs_root *root,
4859                                              u64 parent, u64 root_objectid,
4860                                              const struct btrfs_disk_key *key,
4861                                              int level, u64 hint,
4862                                              u64 empty_size,
4863                                              enum btrfs_lock_nesting nest)
4864 {
4865         struct btrfs_fs_info *fs_info = root->fs_info;
4866         struct btrfs_key ins;
4867         struct btrfs_block_rsv *block_rsv;
4868         struct extent_buffer *buf;
4869         struct btrfs_delayed_extent_op *extent_op;
4870         struct btrfs_ref generic_ref = { 0 };
4871         u64 flags = 0;
4872         int ret;
4873         u32 blocksize = fs_info->nodesize;
4874         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4875
4876 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4877         if (btrfs_is_testing(fs_info)) {
4878                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4879                                             level, root_objectid, nest);
4880                 if (!IS_ERR(buf))
4881                         root->alloc_bytenr += blocksize;
4882                 return buf;
4883         }
4884 #endif
4885
4886         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4887         if (IS_ERR(block_rsv))
4888                 return ERR_CAST(block_rsv);
4889
4890         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4891                                    empty_size, hint, &ins, 0, 0);
4892         if (ret)
4893                 goto out_unuse;
4894
4895         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4896                                     root_objectid, nest);
4897         if (IS_ERR(buf)) {
4898                 ret = PTR_ERR(buf);
4899                 goto out_free_reserved;
4900         }
4901
4902         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4903                 if (parent == 0)
4904                         parent = ins.objectid;
4905                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
4906         } else
4907                 BUG_ON(parent > 0);
4908
4909         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
4910                 extent_op = btrfs_alloc_delayed_extent_op();
4911                 if (!extent_op) {
4912                         ret = -ENOMEM;
4913                         goto out_free_buf;
4914                 }
4915                 if (key)
4916                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
4917                 else
4918                         memset(&extent_op->key, 0, sizeof(extent_op->key));
4919                 extent_op->flags_to_set = flags;
4920                 extent_op->update_key = skinny_metadata ? false : true;
4921                 extent_op->update_flags = true;
4922                 extent_op->level = level;
4923
4924                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4925                                        ins.objectid, ins.offset, parent);
4926                 btrfs_init_tree_ref(&generic_ref, level, root_objectid,
4927                                     root->root_key.objectid, false);
4928                 btrfs_ref_tree_mod(fs_info, &generic_ref);
4929                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
4930                 if (ret)
4931                         goto out_free_delayed;
4932         }
4933         return buf;
4934
4935 out_free_delayed:
4936         btrfs_free_delayed_extent_op(extent_op);
4937 out_free_buf:
4938         btrfs_tree_unlock(buf);
4939         free_extent_buffer(buf);
4940 out_free_reserved:
4941         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
4942 out_unuse:
4943         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
4944         return ERR_PTR(ret);
4945 }
4946
4947 struct walk_control {
4948         u64 refs[BTRFS_MAX_LEVEL];
4949         u64 flags[BTRFS_MAX_LEVEL];
4950         struct btrfs_key update_progress;
4951         struct btrfs_key drop_progress;
4952         int drop_level;
4953         int stage;
4954         int level;
4955         int shared_level;
4956         int update_ref;
4957         int keep_locks;
4958         int reada_slot;
4959         int reada_count;
4960         int restarted;
4961 };
4962
4963 #define DROP_REFERENCE  1
4964 #define UPDATE_BACKREF  2
4965
4966 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
4967                                      struct btrfs_root *root,
4968                                      struct walk_control *wc,
4969                                      struct btrfs_path *path)
4970 {
4971         struct btrfs_fs_info *fs_info = root->fs_info;
4972         u64 bytenr;
4973         u64 generation;
4974         u64 refs;
4975         u64 flags;
4976         u32 nritems;
4977         struct btrfs_key key;
4978         struct extent_buffer *eb;
4979         int ret;
4980         int slot;
4981         int nread = 0;
4982
4983         if (path->slots[wc->level] < wc->reada_slot) {
4984                 wc->reada_count = wc->reada_count * 2 / 3;
4985                 wc->reada_count = max(wc->reada_count, 2);
4986         } else {
4987                 wc->reada_count = wc->reada_count * 3 / 2;
4988                 wc->reada_count = min_t(int, wc->reada_count,
4989                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
4990         }
4991
4992         eb = path->nodes[wc->level];
4993         nritems = btrfs_header_nritems(eb);
4994
4995         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
4996                 if (nread >= wc->reada_count)
4997                         break;
4998
4999                 cond_resched();
5000                 bytenr = btrfs_node_blockptr(eb, slot);
5001                 generation = btrfs_node_ptr_generation(eb, slot);
5002
5003                 if (slot == path->slots[wc->level])
5004                         goto reada;
5005
5006                 if (wc->stage == UPDATE_BACKREF &&
5007                     generation <= root->root_key.offset)
5008                         continue;
5009
5010                 /* We don't lock the tree block, it's OK to be racy here */
5011                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5012                                                wc->level - 1, 1, &refs,
5013                                                &flags);
5014                 /* We don't care about errors in readahead. */
5015                 if (ret < 0)
5016                         continue;
5017                 BUG_ON(refs == 0);
5018
5019                 if (wc->stage == DROP_REFERENCE) {
5020                         if (refs == 1)
5021                                 goto reada;
5022
5023                         if (wc->level == 1 &&
5024                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5025                                 continue;
5026                         if (!wc->update_ref ||
5027                             generation <= root->root_key.offset)
5028                                 continue;
5029                         btrfs_node_key_to_cpu(eb, &key, slot);
5030                         ret = btrfs_comp_cpu_keys(&key,
5031                                                   &wc->update_progress);
5032                         if (ret < 0)
5033                                 continue;
5034                 } else {
5035                         if (wc->level == 1 &&
5036                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5037                                 continue;
5038                 }
5039 reada:
5040                 btrfs_readahead_node_child(eb, slot);
5041                 nread++;
5042         }
5043         wc->reada_slot = slot;
5044 }
5045
5046 /*
5047  * helper to process tree block while walking down the tree.
5048  *
5049  * when wc->stage == UPDATE_BACKREF, this function updates
5050  * back refs for pointers in the block.
5051  *
5052  * NOTE: return value 1 means we should stop walking down.
5053  */
5054 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5055                                    struct btrfs_root *root,
5056                                    struct btrfs_path *path,
5057                                    struct walk_control *wc, int lookup_info)
5058 {
5059         struct btrfs_fs_info *fs_info = root->fs_info;
5060         int level = wc->level;
5061         struct extent_buffer *eb = path->nodes[level];
5062         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5063         int ret;
5064
5065         if (wc->stage == UPDATE_BACKREF &&
5066             btrfs_header_owner(eb) != root->root_key.objectid)
5067                 return 1;
5068
5069         /*
5070          * when reference count of tree block is 1, it won't increase
5071          * again. once full backref flag is set, we never clear it.
5072          */
5073         if (lookup_info &&
5074             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5075              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5076                 BUG_ON(!path->locks[level]);
5077                 ret = btrfs_lookup_extent_info(trans, fs_info,
5078                                                eb->start, level, 1,
5079                                                &wc->refs[level],
5080                                                &wc->flags[level]);
5081                 BUG_ON(ret == -ENOMEM);
5082                 if (ret)
5083                         return ret;
5084                 BUG_ON(wc->refs[level] == 0);
5085         }
5086
5087         if (wc->stage == DROP_REFERENCE) {
5088                 if (wc->refs[level] > 1)
5089                         return 1;
5090
5091                 if (path->locks[level] && !wc->keep_locks) {
5092                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5093                         path->locks[level] = 0;
5094                 }
5095                 return 0;
5096         }
5097
5098         /* wc->stage == UPDATE_BACKREF */
5099         if (!(wc->flags[level] & flag)) {
5100                 BUG_ON(!path->locks[level]);
5101                 ret = btrfs_inc_ref(trans, root, eb, 1);
5102                 BUG_ON(ret); /* -ENOMEM */
5103                 ret = btrfs_dec_ref(trans, root, eb, 0);
5104                 BUG_ON(ret); /* -ENOMEM */
5105                 ret = btrfs_set_disk_extent_flags(trans, eb, flag,
5106                                                   btrfs_header_level(eb));
5107                 BUG_ON(ret); /* -ENOMEM */
5108                 wc->flags[level] |= flag;
5109         }
5110
5111         /*
5112          * the block is shared by multiple trees, so it's not good to
5113          * keep the tree lock
5114          */
5115         if (path->locks[level] && level > 0) {
5116                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5117                 path->locks[level] = 0;
5118         }
5119         return 0;
5120 }
5121
5122 /*
5123  * This is used to verify a ref exists for this root to deal with a bug where we
5124  * would have a drop_progress key that hadn't been updated properly.
5125  */
5126 static int check_ref_exists(struct btrfs_trans_handle *trans,
5127                             struct btrfs_root *root, u64 bytenr, u64 parent,
5128                             int level)
5129 {
5130         struct btrfs_path *path;
5131         struct btrfs_extent_inline_ref *iref;
5132         int ret;
5133
5134         path = btrfs_alloc_path();
5135         if (!path)
5136                 return -ENOMEM;
5137
5138         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5139                                     root->fs_info->nodesize, parent,
5140                                     root->root_key.objectid, level, 0);
5141         btrfs_free_path(path);
5142         if (ret == -ENOENT)
5143                 return 0;
5144         if (ret < 0)
5145                 return ret;
5146         return 1;
5147 }
5148
5149 /*
5150  * helper to process tree block pointer.
5151  *
5152  * when wc->stage == DROP_REFERENCE, this function checks
5153  * reference count of the block pointed to. if the block
5154  * is shared and we need update back refs for the subtree
5155  * rooted at the block, this function changes wc->stage to
5156  * UPDATE_BACKREF. if the block is shared and there is no
5157  * need to update back, this function drops the reference
5158  * to the block.
5159  *
5160  * NOTE: return value 1 means we should stop walking down.
5161  */
5162 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5163                                  struct btrfs_root *root,
5164                                  struct btrfs_path *path,
5165                                  struct walk_control *wc, int *lookup_info)
5166 {
5167         struct btrfs_fs_info *fs_info = root->fs_info;
5168         u64 bytenr;
5169         u64 generation;
5170         u64 parent;
5171         struct btrfs_tree_parent_check check = { 0 };
5172         struct btrfs_key key;
5173         struct btrfs_ref ref = { 0 };
5174         struct extent_buffer *next;
5175         int level = wc->level;
5176         int reada = 0;
5177         int ret = 0;
5178         bool need_account = false;
5179
5180         generation = btrfs_node_ptr_generation(path->nodes[level],
5181                                                path->slots[level]);
5182         /*
5183          * if the lower level block was created before the snapshot
5184          * was created, we know there is no need to update back refs
5185          * for the subtree
5186          */
5187         if (wc->stage == UPDATE_BACKREF &&
5188             generation <= root->root_key.offset) {
5189                 *lookup_info = 1;
5190                 return 1;
5191         }
5192
5193         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5194
5195         check.level = level - 1;
5196         check.transid = generation;
5197         check.owner_root = root->root_key.objectid;
5198         check.has_first_key = true;
5199         btrfs_node_key_to_cpu(path->nodes[level], &check.first_key,
5200                               path->slots[level]);
5201
5202         next = find_extent_buffer(fs_info, bytenr);
5203         if (!next) {
5204                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5205                                 root->root_key.objectid, level - 1);
5206                 if (IS_ERR(next))
5207                         return PTR_ERR(next);
5208                 reada = 1;
5209         }
5210         btrfs_tree_lock(next);
5211
5212         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5213                                        &wc->refs[level - 1],
5214                                        &wc->flags[level - 1]);
5215         if (ret < 0)
5216                 goto out_unlock;
5217
5218         if (unlikely(wc->refs[level - 1] == 0)) {
5219                 btrfs_err(fs_info, "Missing references.");
5220                 ret = -EIO;
5221                 goto out_unlock;
5222         }
5223         *lookup_info = 0;
5224
5225         if (wc->stage == DROP_REFERENCE) {
5226                 if (wc->refs[level - 1] > 1) {
5227                         need_account = true;
5228                         if (level == 1 &&
5229                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5230                                 goto skip;
5231
5232                         if (!wc->update_ref ||
5233                             generation <= root->root_key.offset)
5234                                 goto skip;
5235
5236                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5237                                               path->slots[level]);
5238                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5239                         if (ret < 0)
5240                                 goto skip;
5241
5242                         wc->stage = UPDATE_BACKREF;
5243                         wc->shared_level = level - 1;
5244                 }
5245         } else {
5246                 if (level == 1 &&
5247                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5248                         goto skip;
5249         }
5250
5251         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5252                 btrfs_tree_unlock(next);
5253                 free_extent_buffer(next);
5254                 next = NULL;
5255                 *lookup_info = 1;
5256         }
5257
5258         if (!next) {
5259                 if (reada && level == 1)
5260                         reada_walk_down(trans, root, wc, path);
5261                 next = read_tree_block(fs_info, bytenr, &check);
5262                 if (IS_ERR(next)) {
5263                         return PTR_ERR(next);
5264                 } else if (!extent_buffer_uptodate(next)) {
5265                         free_extent_buffer(next);
5266                         return -EIO;
5267                 }
5268                 btrfs_tree_lock(next);
5269         }
5270
5271         level--;
5272         ASSERT(level == btrfs_header_level(next));
5273         if (level != btrfs_header_level(next)) {
5274                 btrfs_err(root->fs_info, "mismatched level");
5275                 ret = -EIO;
5276                 goto out_unlock;
5277         }
5278         path->nodes[level] = next;
5279         path->slots[level] = 0;
5280         path->locks[level] = BTRFS_WRITE_LOCK;
5281         wc->level = level;
5282         if (wc->level == 1)
5283                 wc->reada_slot = 0;
5284         return 0;
5285 skip:
5286         wc->refs[level - 1] = 0;
5287         wc->flags[level - 1] = 0;
5288         if (wc->stage == DROP_REFERENCE) {
5289                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5290                         parent = path->nodes[level]->start;
5291                 } else {
5292                         ASSERT(root->root_key.objectid ==
5293                                btrfs_header_owner(path->nodes[level]));
5294                         if (root->root_key.objectid !=
5295                             btrfs_header_owner(path->nodes[level])) {
5296                                 btrfs_err(root->fs_info,
5297                                                 "mismatched block owner");
5298                                 ret = -EIO;
5299                                 goto out_unlock;
5300                         }
5301                         parent = 0;
5302                 }
5303
5304                 /*
5305                  * If we had a drop_progress we need to verify the refs are set
5306                  * as expected.  If we find our ref then we know that from here
5307                  * on out everything should be correct, and we can clear the
5308                  * ->restarted flag.
5309                  */
5310                 if (wc->restarted) {
5311                         ret = check_ref_exists(trans, root, bytenr, parent,
5312                                                level - 1);
5313                         if (ret < 0)
5314                                 goto out_unlock;
5315                         if (ret == 0)
5316                                 goto no_delete;
5317                         ret = 0;
5318                         wc->restarted = 0;
5319                 }
5320
5321                 /*
5322                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5323                  * already accounted them at merge time (replace_path),
5324                  * thus we could skip expensive subtree trace here.
5325                  */
5326                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5327                     need_account) {
5328                         ret = btrfs_qgroup_trace_subtree(trans, next,
5329                                                          generation, level - 1);
5330                         if (ret) {
5331                                 btrfs_err_rl(fs_info,
5332                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5333                                              ret);
5334                         }
5335                 }
5336
5337                 /*
5338                  * We need to update the next key in our walk control so we can
5339                  * update the drop_progress key accordingly.  We don't care if
5340                  * find_next_key doesn't find a key because that means we're at
5341                  * the end and are going to clean up now.
5342                  */
5343                 wc->drop_level = level;
5344                 find_next_key(path, level, &wc->drop_progress);
5345
5346                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5347                                        fs_info->nodesize, parent);
5348                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid,
5349                                     0, false);
5350                 ret = btrfs_free_extent(trans, &ref);
5351                 if (ret)
5352                         goto out_unlock;
5353         }
5354 no_delete:
5355         *lookup_info = 1;
5356         ret = 1;
5357
5358 out_unlock:
5359         btrfs_tree_unlock(next);
5360         free_extent_buffer(next);
5361
5362         return ret;
5363 }
5364
5365 /*
5366  * helper to process tree block while walking up the tree.
5367  *
5368  * when wc->stage == DROP_REFERENCE, this function drops
5369  * reference count on the block.
5370  *
5371  * when wc->stage == UPDATE_BACKREF, this function changes
5372  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5373  * to UPDATE_BACKREF previously while processing the block.
5374  *
5375  * NOTE: return value 1 means we should stop walking up.
5376  */
5377 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5378                                  struct btrfs_root *root,
5379                                  struct btrfs_path *path,
5380                                  struct walk_control *wc)
5381 {
5382         struct btrfs_fs_info *fs_info = root->fs_info;
5383         int ret;
5384         int level = wc->level;
5385         struct extent_buffer *eb = path->nodes[level];
5386         u64 parent = 0;
5387
5388         if (wc->stage == UPDATE_BACKREF) {
5389                 BUG_ON(wc->shared_level < level);
5390                 if (level < wc->shared_level)
5391                         goto out;
5392
5393                 ret = find_next_key(path, level + 1, &wc->update_progress);
5394                 if (ret > 0)
5395                         wc->update_ref = 0;
5396
5397                 wc->stage = DROP_REFERENCE;
5398                 wc->shared_level = -1;
5399                 path->slots[level] = 0;
5400
5401                 /*
5402                  * check reference count again if the block isn't locked.
5403                  * we should start walking down the tree again if reference
5404                  * count is one.
5405                  */
5406                 if (!path->locks[level]) {
5407                         BUG_ON(level == 0);
5408                         btrfs_tree_lock(eb);
5409                         path->locks[level] = BTRFS_WRITE_LOCK;
5410
5411                         ret = btrfs_lookup_extent_info(trans, fs_info,
5412                                                        eb->start, level, 1,
5413                                                        &wc->refs[level],
5414                                                        &wc->flags[level]);
5415                         if (ret < 0) {
5416                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5417                                 path->locks[level] = 0;
5418                                 return ret;
5419                         }
5420                         BUG_ON(wc->refs[level] == 0);
5421                         if (wc->refs[level] == 1) {
5422                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5423                                 path->locks[level] = 0;
5424                                 return 1;
5425                         }
5426                 }
5427         }
5428
5429         /* wc->stage == DROP_REFERENCE */
5430         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5431
5432         if (wc->refs[level] == 1) {
5433                 if (level == 0) {
5434                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5435                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5436                         else
5437                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5438                         BUG_ON(ret); /* -ENOMEM */
5439                         if (is_fstree(root->root_key.objectid)) {
5440                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5441                                 if (ret) {
5442                                         btrfs_err_rl(fs_info,
5443         "error %d accounting leaf items, quota is out of sync, rescan required",
5444                                              ret);
5445                                 }
5446                         }
5447                 }
5448                 /* Make block locked assertion in btrfs_clear_buffer_dirty happy. */
5449                 if (!path->locks[level]) {
5450                         btrfs_tree_lock(eb);
5451                         path->locks[level] = BTRFS_WRITE_LOCK;
5452                 }
5453                 btrfs_clear_buffer_dirty(trans, eb);
5454         }
5455
5456         if (eb == root->node) {
5457                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5458                         parent = eb->start;
5459                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5460                         goto owner_mismatch;
5461         } else {
5462                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5463                         parent = path->nodes[level + 1]->start;
5464                 else if (root->root_key.objectid !=
5465                          btrfs_header_owner(path->nodes[level + 1]))
5466                         goto owner_mismatch;
5467         }
5468
5469         btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5470                               wc->refs[level] == 1);
5471 out:
5472         wc->refs[level] = 0;
5473         wc->flags[level] = 0;
5474         return 0;
5475
5476 owner_mismatch:
5477         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5478                      btrfs_header_owner(eb), root->root_key.objectid);
5479         return -EUCLEAN;
5480 }
5481
5482 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5483                                    struct btrfs_root *root,
5484                                    struct btrfs_path *path,
5485                                    struct walk_control *wc)
5486 {
5487         int level = wc->level;
5488         int lookup_info = 1;
5489         int ret = 0;
5490
5491         while (level >= 0) {
5492                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5493                 if (ret)
5494                         break;
5495
5496                 if (level == 0)
5497                         break;
5498
5499                 if (path->slots[level] >=
5500                     btrfs_header_nritems(path->nodes[level]))
5501                         break;
5502
5503                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5504                 if (ret > 0) {
5505                         path->slots[level]++;
5506                         continue;
5507                 } else if (ret < 0)
5508                         break;
5509                 level = wc->level;
5510         }
5511         return (ret == 1) ? 0 : ret;
5512 }
5513
5514 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5515                                  struct btrfs_root *root,
5516                                  struct btrfs_path *path,
5517                                  struct walk_control *wc, int max_level)
5518 {
5519         int level = wc->level;
5520         int ret;
5521
5522         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5523         while (level < max_level && path->nodes[level]) {
5524                 wc->level = level;
5525                 if (path->slots[level] + 1 <
5526                     btrfs_header_nritems(path->nodes[level])) {
5527                         path->slots[level]++;
5528                         return 0;
5529                 } else {
5530                         ret = walk_up_proc(trans, root, path, wc);
5531                         if (ret > 0)
5532                                 return 0;
5533                         if (ret < 0)
5534                                 return ret;
5535
5536                         if (path->locks[level]) {
5537                                 btrfs_tree_unlock_rw(path->nodes[level],
5538                                                      path->locks[level]);
5539                                 path->locks[level] = 0;
5540                         }
5541                         free_extent_buffer(path->nodes[level]);
5542                         path->nodes[level] = NULL;
5543                         level++;
5544                 }
5545         }
5546         return 1;
5547 }
5548
5549 /*
5550  * drop a subvolume tree.
5551  *
5552  * this function traverses the tree freeing any blocks that only
5553  * referenced by the tree.
5554  *
5555  * when a shared tree block is found. this function decreases its
5556  * reference count by one. if update_ref is true, this function
5557  * also make sure backrefs for the shared block and all lower level
5558  * blocks are properly updated.
5559  *
5560  * If called with for_reloc == 0, may exit early with -EAGAIN
5561  */
5562 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5563 {
5564         const bool is_reloc_root = (root->root_key.objectid ==
5565                                     BTRFS_TREE_RELOC_OBJECTID);
5566         struct btrfs_fs_info *fs_info = root->fs_info;
5567         struct btrfs_path *path;
5568         struct btrfs_trans_handle *trans;
5569         struct btrfs_root *tree_root = fs_info->tree_root;
5570         struct btrfs_root_item *root_item = &root->root_item;
5571         struct walk_control *wc;
5572         struct btrfs_key key;
5573         int err = 0;
5574         int ret;
5575         int level;
5576         bool root_dropped = false;
5577         bool unfinished_drop = false;
5578
5579         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5580
5581         path = btrfs_alloc_path();
5582         if (!path) {
5583                 err = -ENOMEM;
5584                 goto out;
5585         }
5586
5587         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5588         if (!wc) {
5589                 btrfs_free_path(path);
5590                 err = -ENOMEM;
5591                 goto out;
5592         }
5593
5594         /*
5595          * Use join to avoid potential EINTR from transaction start. See
5596          * wait_reserve_ticket and the whole reservation callchain.
5597          */
5598         if (for_reloc)
5599                 trans = btrfs_join_transaction(tree_root);
5600         else
5601                 trans = btrfs_start_transaction(tree_root, 0);
5602         if (IS_ERR(trans)) {
5603                 err = PTR_ERR(trans);
5604                 goto out_free;
5605         }
5606
5607         err = btrfs_run_delayed_items(trans);
5608         if (err)
5609                 goto out_end_trans;
5610
5611         /*
5612          * This will help us catch people modifying the fs tree while we're
5613          * dropping it.  It is unsafe to mess with the fs tree while it's being
5614          * dropped as we unlock the root node and parent nodes as we walk down
5615          * the tree, assuming nothing will change.  If something does change
5616          * then we'll have stale information and drop references to blocks we've
5617          * already dropped.
5618          */
5619         set_bit(BTRFS_ROOT_DELETING, &root->state);
5620         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
5621
5622         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5623                 level = btrfs_header_level(root->node);
5624                 path->nodes[level] = btrfs_lock_root_node(root);
5625                 path->slots[level] = 0;
5626                 path->locks[level] = BTRFS_WRITE_LOCK;
5627                 memset(&wc->update_progress, 0,
5628                        sizeof(wc->update_progress));
5629         } else {
5630                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5631                 memcpy(&wc->update_progress, &key,
5632                        sizeof(wc->update_progress));
5633
5634                 level = btrfs_root_drop_level(root_item);
5635                 BUG_ON(level == 0);
5636                 path->lowest_level = level;
5637                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5638                 path->lowest_level = 0;
5639                 if (ret < 0) {
5640                         err = ret;
5641                         goto out_end_trans;
5642                 }
5643                 WARN_ON(ret > 0);
5644
5645                 /*
5646                  * unlock our path, this is safe because only this
5647                  * function is allowed to delete this snapshot
5648                  */
5649                 btrfs_unlock_up_safe(path, 0);
5650
5651                 level = btrfs_header_level(root->node);
5652                 while (1) {
5653                         btrfs_tree_lock(path->nodes[level]);
5654                         path->locks[level] = BTRFS_WRITE_LOCK;
5655
5656                         ret = btrfs_lookup_extent_info(trans, fs_info,
5657                                                 path->nodes[level]->start,
5658                                                 level, 1, &wc->refs[level],
5659                                                 &wc->flags[level]);
5660                         if (ret < 0) {
5661                                 err = ret;
5662                                 goto out_end_trans;
5663                         }
5664                         BUG_ON(wc->refs[level] == 0);
5665
5666                         if (level == btrfs_root_drop_level(root_item))
5667                                 break;
5668
5669                         btrfs_tree_unlock(path->nodes[level]);
5670                         path->locks[level] = 0;
5671                         WARN_ON(wc->refs[level] != 1);
5672                         level--;
5673                 }
5674         }
5675
5676         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5677         wc->level = level;
5678         wc->shared_level = -1;
5679         wc->stage = DROP_REFERENCE;
5680         wc->update_ref = update_ref;
5681         wc->keep_locks = 0;
5682         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5683
5684         while (1) {
5685
5686                 ret = walk_down_tree(trans, root, path, wc);
5687                 if (ret < 0) {
5688                         btrfs_abort_transaction(trans, ret);
5689                         err = ret;
5690                         break;
5691                 }
5692
5693                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5694                 if (ret < 0) {
5695                         btrfs_abort_transaction(trans, ret);
5696                         err = ret;
5697                         break;
5698                 }
5699
5700                 if (ret > 0) {
5701                         BUG_ON(wc->stage != DROP_REFERENCE);
5702                         break;
5703                 }
5704
5705                 if (wc->stage == DROP_REFERENCE) {
5706                         wc->drop_level = wc->level;
5707                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5708                                               &wc->drop_progress,
5709                                               path->slots[wc->drop_level]);
5710                 }
5711                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5712                                       &wc->drop_progress);
5713                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5714
5715                 BUG_ON(wc->level == 0);
5716                 if (btrfs_should_end_transaction(trans) ||
5717                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5718                         ret = btrfs_update_root(trans, tree_root,
5719                                                 &root->root_key,
5720                                                 root_item);
5721                         if (ret) {
5722                                 btrfs_abort_transaction(trans, ret);
5723                                 err = ret;
5724                                 goto out_end_trans;
5725                         }
5726
5727                         if (!is_reloc_root)
5728                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5729
5730                         btrfs_end_transaction_throttle(trans);
5731                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5732                                 btrfs_debug(fs_info,
5733                                             "drop snapshot early exit");
5734                                 err = -EAGAIN;
5735                                 goto out_free;
5736                         }
5737
5738                        /*
5739                         * Use join to avoid potential EINTR from transaction
5740                         * start. See wait_reserve_ticket and the whole
5741                         * reservation callchain.
5742                         */
5743                         if (for_reloc)
5744                                 trans = btrfs_join_transaction(tree_root);
5745                         else
5746                                 trans = btrfs_start_transaction(tree_root, 0);
5747                         if (IS_ERR(trans)) {
5748                                 err = PTR_ERR(trans);
5749                                 goto out_free;
5750                         }
5751                 }
5752         }
5753         btrfs_release_path(path);
5754         if (err)
5755                 goto out_end_trans;
5756
5757         ret = btrfs_del_root(trans, &root->root_key);
5758         if (ret) {
5759                 btrfs_abort_transaction(trans, ret);
5760                 err = ret;
5761                 goto out_end_trans;
5762         }
5763
5764         if (!is_reloc_root) {
5765                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5766                                       NULL, NULL);
5767                 if (ret < 0) {
5768                         btrfs_abort_transaction(trans, ret);
5769                         err = ret;
5770                         goto out_end_trans;
5771                 } else if (ret > 0) {
5772                         /* if we fail to delete the orphan item this time
5773                          * around, it'll get picked up the next time.
5774                          *
5775                          * The most common failure here is just -ENOENT.
5776                          */
5777                         btrfs_del_orphan_item(trans, tree_root,
5778                                               root->root_key.objectid);
5779                 }
5780         }
5781
5782         /*
5783          * This subvolume is going to be completely dropped, and won't be
5784          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5785          * commit transaction time.  So free it here manually.
5786          */
5787         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5788         btrfs_qgroup_free_meta_all_pertrans(root);
5789
5790         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5791                 btrfs_add_dropped_root(trans, root);
5792         else
5793                 btrfs_put_root(root);
5794         root_dropped = true;
5795 out_end_trans:
5796         if (!is_reloc_root)
5797                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5798
5799         btrfs_end_transaction_throttle(trans);
5800 out_free:
5801         kfree(wc);
5802         btrfs_free_path(path);
5803 out:
5804         /*
5805          * We were an unfinished drop root, check to see if there are any
5806          * pending, and if not clear and wake up any waiters.
5807          */
5808         if (!err && unfinished_drop)
5809                 btrfs_maybe_wake_unfinished_drop(fs_info);
5810
5811         /*
5812          * So if we need to stop dropping the snapshot for whatever reason we
5813          * need to make sure to add it back to the dead root list so that we
5814          * keep trying to do the work later.  This also cleans up roots if we
5815          * don't have it in the radix (like when we recover after a power fail
5816          * or unmount) so we don't leak memory.
5817          */
5818         if (!for_reloc && !root_dropped)
5819                 btrfs_add_dead_root(root);
5820         return err;
5821 }
5822
5823 /*
5824  * drop subtree rooted at tree block 'node'.
5825  *
5826  * NOTE: this function will unlock and release tree block 'node'
5827  * only used by relocation code
5828  */
5829 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5830                         struct btrfs_root *root,
5831                         struct extent_buffer *node,
5832                         struct extent_buffer *parent)
5833 {
5834         struct btrfs_fs_info *fs_info = root->fs_info;
5835         struct btrfs_path *path;
5836         struct walk_control *wc;
5837         int level;
5838         int parent_level;
5839         int ret = 0;
5840         int wret;
5841
5842         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5843
5844         path = btrfs_alloc_path();
5845         if (!path)
5846                 return -ENOMEM;
5847
5848         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5849         if (!wc) {
5850                 btrfs_free_path(path);
5851                 return -ENOMEM;
5852         }
5853
5854         btrfs_assert_tree_write_locked(parent);
5855         parent_level = btrfs_header_level(parent);
5856         atomic_inc(&parent->refs);
5857         path->nodes[parent_level] = parent;
5858         path->slots[parent_level] = btrfs_header_nritems(parent);
5859
5860         btrfs_assert_tree_write_locked(node);
5861         level = btrfs_header_level(node);
5862         path->nodes[level] = node;
5863         path->slots[level] = 0;
5864         path->locks[level] = BTRFS_WRITE_LOCK;
5865
5866         wc->refs[parent_level] = 1;
5867         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5868         wc->level = level;
5869         wc->shared_level = -1;
5870         wc->stage = DROP_REFERENCE;
5871         wc->update_ref = 0;
5872         wc->keep_locks = 1;
5873         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5874
5875         while (1) {
5876                 wret = walk_down_tree(trans, root, path, wc);
5877                 if (wret < 0) {
5878                         ret = wret;
5879                         break;
5880                 }
5881
5882                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5883                 if (wret < 0)
5884                         ret = wret;
5885                 if (wret != 0)
5886                         break;
5887         }
5888
5889         kfree(wc);
5890         btrfs_free_path(path);
5891         return ret;
5892 }
5893
5894 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5895                                    u64 start, u64 end)
5896 {
5897         return unpin_extent_range(fs_info, start, end, false);
5898 }
5899
5900 /*
5901  * It used to be that old block groups would be left around forever.
5902  * Iterating over them would be enough to trim unused space.  Since we
5903  * now automatically remove them, we also need to iterate over unallocated
5904  * space.
5905  *
5906  * We don't want a transaction for this since the discard may take a
5907  * substantial amount of time.  We don't require that a transaction be
5908  * running, but we do need to take a running transaction into account
5909  * to ensure that we're not discarding chunks that were released or
5910  * allocated in the current transaction.
5911  *
5912  * Holding the chunks lock will prevent other threads from allocating
5913  * or releasing chunks, but it won't prevent a running transaction
5914  * from committing and releasing the memory that the pending chunks
5915  * list head uses.  For that, we need to take a reference to the
5916  * transaction and hold the commit root sem.  We only need to hold
5917  * it while performing the free space search since we have already
5918  * held back allocations.
5919  */
5920 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
5921 {
5922         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
5923         int ret;
5924
5925         *trimmed = 0;
5926
5927         /* Discard not supported = nothing to do. */
5928         if (!bdev_max_discard_sectors(device->bdev))
5929                 return 0;
5930
5931         /* Not writable = nothing to do. */
5932         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
5933                 return 0;
5934
5935         /* No free space = nothing to do. */
5936         if (device->total_bytes <= device->bytes_used)
5937                 return 0;
5938
5939         ret = 0;
5940
5941         while (1) {
5942                 struct btrfs_fs_info *fs_info = device->fs_info;
5943                 u64 bytes;
5944
5945                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
5946                 if (ret)
5947                         break;
5948
5949                 find_first_clear_extent_bit(&device->alloc_state, start,
5950                                             &start, &end,
5951                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
5952
5953                 /* Check if there are any CHUNK_* bits left */
5954                 if (start > device->total_bytes) {
5955                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
5956                         btrfs_warn_in_rcu(fs_info,
5957 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
5958                                           start, end - start + 1,
5959                                           btrfs_dev_name(device),
5960                                           device->total_bytes);
5961                         mutex_unlock(&fs_info->chunk_mutex);
5962                         ret = 0;
5963                         break;
5964                 }
5965
5966                 /* Ensure we skip the reserved space on each device. */
5967                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
5968
5969                 /*
5970                  * If find_first_clear_extent_bit find a range that spans the
5971                  * end of the device it will set end to -1, in this case it's up
5972                  * to the caller to trim the value to the size of the device.
5973                  */
5974                 end = min(end, device->total_bytes - 1);
5975
5976                 len = end - start + 1;
5977
5978                 /* We didn't find any extents */
5979                 if (!len) {
5980                         mutex_unlock(&fs_info->chunk_mutex);
5981                         ret = 0;
5982                         break;
5983                 }
5984
5985                 ret = btrfs_issue_discard(device->bdev, start, len,
5986                                           &bytes);
5987                 if (!ret)
5988                         set_extent_bits(&device->alloc_state, start,
5989                                         start + bytes - 1,
5990                                         CHUNK_TRIMMED);
5991                 mutex_unlock(&fs_info->chunk_mutex);
5992
5993                 if (ret)
5994                         break;
5995
5996                 start += len;
5997                 *trimmed += bytes;
5998
5999                 if (fatal_signal_pending(current)) {
6000                         ret = -ERESTARTSYS;
6001                         break;
6002                 }
6003
6004                 cond_resched();
6005         }
6006
6007         return ret;
6008 }
6009
6010 /*
6011  * Trim the whole filesystem by:
6012  * 1) trimming the free space in each block group
6013  * 2) trimming the unallocated space on each device
6014  *
6015  * This will also continue trimming even if a block group or device encounters
6016  * an error.  The return value will be the last error, or 0 if nothing bad
6017  * happens.
6018  */
6019 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6020 {
6021         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6022         struct btrfs_block_group *cache = NULL;
6023         struct btrfs_device *device;
6024         u64 group_trimmed;
6025         u64 range_end = U64_MAX;
6026         u64 start;
6027         u64 end;
6028         u64 trimmed = 0;
6029         u64 bg_failed = 0;
6030         u64 dev_failed = 0;
6031         int bg_ret = 0;
6032         int dev_ret = 0;
6033         int ret = 0;
6034
6035         if (range->start == U64_MAX)
6036                 return -EINVAL;
6037
6038         /*
6039          * Check range overflow if range->len is set.
6040          * The default range->len is U64_MAX.
6041          */
6042         if (range->len != U64_MAX &&
6043             check_add_overflow(range->start, range->len, &range_end))
6044                 return -EINVAL;
6045
6046         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6047         for (; cache; cache = btrfs_next_block_group(cache)) {
6048                 if (cache->start >= range_end) {
6049                         btrfs_put_block_group(cache);
6050                         break;
6051                 }
6052
6053                 start = max(range->start, cache->start);
6054                 end = min(range_end, cache->start + cache->length);
6055
6056                 if (end - start >= range->minlen) {
6057                         if (!btrfs_block_group_done(cache)) {
6058                                 ret = btrfs_cache_block_group(cache, true);
6059                                 if (ret) {
6060                                         bg_failed++;
6061                                         bg_ret = ret;
6062                                         continue;
6063                                 }
6064                         }
6065                         ret = btrfs_trim_block_group(cache,
6066                                                      &group_trimmed,
6067                                                      start,
6068                                                      end,
6069                                                      range->minlen);
6070
6071                         trimmed += group_trimmed;
6072                         if (ret) {
6073                                 bg_failed++;
6074                                 bg_ret = ret;
6075                                 continue;
6076                         }
6077                 }
6078         }
6079
6080         if (bg_failed)
6081                 btrfs_warn(fs_info,
6082                         "failed to trim %llu block group(s), last error %d",
6083                         bg_failed, bg_ret);
6084
6085         mutex_lock(&fs_devices->device_list_mutex);
6086         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6087                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6088                         continue;
6089
6090                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6091                 if (ret) {
6092                         dev_failed++;
6093                         dev_ret = ret;
6094                         break;
6095                 }
6096
6097                 trimmed += group_trimmed;
6098         }
6099         mutex_unlock(&fs_devices->device_list_mutex);
6100
6101         if (dev_failed)
6102                 btrfs_warn(fs_info,
6103                         "failed to trim %llu device(s), last error %d",
6104                         dev_failed, dev_ret);
6105         range->len = trimmed;
6106         if (bg_ret)
6107                 return bg_ret;
6108         return dev_ret;
6109 }