fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "misc.h"
  20 #include "tree-log.h"
  21 #include "disk-io.h"
  22 #include "print-tree.h"
  23 #include "volumes.h"
  24 #include "raid56.h"
  25 #include "locking.h"
  26 #include "free-space-cache.h"
  27 #include "free-space-tree.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31 #include "space-info.h"
  32 #include "block-rsv.h"
  33 #include "delalloc-space.h"
  34 #include "block-group.h"
  35 #include "discard.h"
  36 #include "rcu-string.h"
  37 #include "zoned.h"
  38 #include "dev-replace.h"
  39 #include "fs.h"
  40 #include "accessors.h"
  41 #include "extent-tree.h"
  42 #include "root-tree.h"
  43 #include "file-item.h"
  44 #include "orphan.h"
  45 #include "tree-checker.h"
  46
  47 #undef SCRAMBLE_DELAYED_REFS
  48
  49
  50 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  51                                struct btrfs_delayed_ref_node *node, u64 parent,
  52                                u64 root_objectid, u64 owner_objectid,
  53                                u64 owner_offset, int refs_to_drop,
  54                                struct btrfs_delayed_extent_op *extra_op);
  55 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  56                                     struct extent_buffer *leaf,
  57                                     struct btrfs_extent_item *ei);
  58 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  59                                       u64 parent, u64 root_objectid,
  60                                       u64 flags, u64 owner, u64 offset,
  61                                       struct btrfs_key *ins, int ref_mod);
  62 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  63                                      struct btrfs_delayed_ref_node *node,
  64                                      struct btrfs_delayed_extent_op *extent_op);
  65 static int find_next_key(struct btrfs_path *path, int level,
  66                          struct btrfs_key *key);
  67
  68 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  69 {
  70         return (cache->flags & bits) == bits;
  71 }
  72
  73 int btrfs_add_excluded_extent(struct btrfs_fs_info *fs_info,
  74                               u64 start, u64 num_bytes)
  75 {
  76         u64 end = start + num_bytes - 1;
  77         set_extent_bits(&fs_info->excluded_extents, start, end,
  78                         EXTENT_UPTODATE);
  79         return 0;
  80 }
  81
  82 void btrfs_free_excluded_extents(struct btrfs_block_group *cache)
  83 {
  84         struct btrfs_fs_info *fs_info = cache->fs_info;
  85         u64 start, end;
  86
  87         start = cache->start;
  88         end = start + cache->length - 1;
  89
  90         clear_extent_bits(&fs_info->excluded_extents, start, end,
  91                           EXTENT_UPTODATE);
  92 }
  93
  94 /* simple helper to search for an existing data extent at a given offset */
  95 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  96 {
  97         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  98         int ret;
  99         struct btrfs_key key;
 100         struct btrfs_path *path;
 101
 102         path = btrfs_alloc_path();
 103         if (!path)
 104                 return -ENOMEM;
 105
 106         key.objectid = start;
 107         key.offset = len;
 108         key.type = BTRFS_EXTENT_ITEM_KEY;
 109         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 110         btrfs_free_path(path);
 111         return ret;
 112 }
 113
 114 /*
 115  * helper function to lookup reference count and flags of a tree block.
 116  *
 117  * the head node for delayed ref is used to store the sum of all the
 118  * reference count modifications queued up in the rbtree. the head
 119  * node may also store the extent flags to set. This way you can check
 120  * to see what the reference count and extent flags would be if all of
 121  * the delayed refs are not processed.
 122  */
 123 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 124                              struct btrfs_fs_info *fs_info, u64 bytenr,
 125                              u64 offset, int metadata, u64 *refs, u64 *flags)
 126 {
 127         struct btrfs_root *extent_root;
 128         struct btrfs_delayed_ref_head *head;
 129         struct btrfs_delayed_ref_root *delayed_refs;
 130         struct btrfs_path *path;
 131         struct btrfs_extent_item *ei;
 132         struct extent_buffer *leaf;
 133         struct btrfs_key key;
 134         u32 item_size;
 135         u64 num_refs;
 136         u64 extent_flags;
 137         int ret;
 138
 139         /*
 140          * If we don't have skinny metadata, don't bother doing anything
 141          * different
 142          */
 143         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 144                 offset = fs_info->nodesize;
 145                 metadata = 0;
 146         }
 147
 148         path = btrfs_alloc_path();
 149         if (!path)
 150                 return -ENOMEM;
 151
 152         if (!trans) {
 153                 path->skip_locking = 1;
 154                 path->search_commit_root = 1;
 155         }
 156
 157 search_again:
 158         key.objectid = bytenr;
 159         key.offset = offset;
 160         if (metadata)
 161                 key.type = BTRFS_METADATA_ITEM_KEY;
 162         else
 163                 key.type = BTRFS_EXTENT_ITEM_KEY;
 164
 165         extent_root = btrfs_extent_root(fs_info, bytenr);
 166         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 167         if (ret < 0)
 168                 goto out_free;
 169
 170         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 171                 if (path->slots[0]) {
 172                         path->slots[0]--;
 173                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 174                                               path->slots[0]);
 175                         if (key.objectid == bytenr &&
 176                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 177                             key.offset == fs_info->nodesize)
 178                                 ret = 0;
 179                 }
 180         }
 181
 182         if (ret == 0) {
 183                 leaf = path->nodes[0];
 184                 item_size = btrfs_item_size(leaf, path->slots[0]);
 185                 if (item_size >= sizeof(*ei)) {
 186                         ei = btrfs_item_ptr(leaf, path->slots[0],
 187                                             struct btrfs_extent_item);
 188                         num_refs = btrfs_extent_refs(leaf, ei);
 189                         extent_flags = btrfs_extent_flags(leaf, ei);
 190                 } else {
 191                         ret = -EINVAL;
 192                         btrfs_print_v0_err(fs_info);
 193                         if (trans)
 194                                 btrfs_abort_transaction(trans, ret);
 195                         else
 196                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 197
 198                         goto out_free;
 199                 }
 200
 201                 BUG_ON(num_refs == 0);
 202         } else {
 203                 num_refs = 0;
 204                 extent_flags = 0;
 205                 ret = 0;
 206         }
 207
 208         if (!trans)
 209                 goto out;
 210
 211         delayed_refs = &trans->transaction->delayed_refs;
 212         spin_lock(&delayed_refs->lock);
 213         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 214         if (head) {
 215                 if (!mutex_trylock(&head->mutex)) {
 216                         refcount_inc(&head->refs);
 217                         spin_unlock(&delayed_refs->lock);
 218
 219                         btrfs_release_path(path);
 220
 221                         /*
 222                          * Mutex was contended, block until it's released and try
 223                          * again
 224                          */
 225                         mutex_lock(&head->mutex);
 226                         mutex_unlock(&head->mutex);
 227                         btrfs_put_delayed_ref_head(head);
 228                         goto search_again;
 229                 }
 230                 spin_lock(&head->lock);
 231                 if (head->extent_op && head->extent_op->update_flags)
 232                         extent_flags |= head->extent_op->flags_to_set;
 233                 else
 234                         BUG_ON(num_refs == 0);
 235
 236                 num_refs += head->ref_mod;
 237                 spin_unlock(&head->lock);
 238                 mutex_unlock(&head->mutex);
 239         }
 240         spin_unlock(&delayed_refs->lock);
 241 out:
 242         WARN_ON(num_refs == 0);
 243         if (refs)
 244                 *refs = num_refs;
 245         if (flags)
 246                 *flags = extent_flags;
 247 out_free:
 248         btrfs_free_path(path);
 249         return ret;
 250 }
 251
 252 /*
 253  * Back reference rules.  Back refs have three main goals:
 254  *
 255  * 1) differentiate between all holders of references to an extent so that
 256  *    when a reference is dropped we can make sure it was a valid reference
 257  *    before freeing the extent.
 258  *
 259  * 2) Provide enough information to quickly find the holders of an extent
 260  *    if we notice a given block is corrupted or bad.
 261  *
 262  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 263  *    maintenance.  This is actually the same as #2, but with a slightly
 264  *    different use case.
 265  *
 266  * There are two kinds of back refs. The implicit back refs is optimized
 267  * for pointers in non-shared tree blocks. For a given pointer in a block,
 268  * back refs of this kind provide information about the block's owner tree
 269  * and the pointer's key. These information allow us to find the block by
 270  * b-tree searching. The full back refs is for pointers in tree blocks not
 271  * referenced by their owner trees. The location of tree block is recorded
 272  * in the back refs. Actually the full back refs is generic, and can be
 273  * used in all cases the implicit back refs is used. The major shortcoming
 274  * of the full back refs is its overhead. Every time a tree block gets
 275  * COWed, we have to update back refs entry for all pointers in it.
 276  *
 277  * For a newly allocated tree block, we use implicit back refs for
 278  * pointers in it. This means most tree related operations only involve
 279  * implicit back refs. For a tree block created in old transaction, the
 280  * only way to drop a reference to it is COW it. So we can detect the
 281  * event that tree block loses its owner tree's reference and do the
 282  * back refs conversion.
 283  *
 284  * When a tree block is COWed through a tree, there are four cases:
 285  *
 286  * The reference count of the block is one and the tree is the block's
 287  * owner tree. Nothing to do in this case.
 288  *
 289  * The reference count of the block is one and the tree is not the
 290  * block's owner tree. In this case, full back refs is used for pointers
 291  * in the block. Remove these full back refs, add implicit back refs for
 292  * every pointers in the new block.
 293  *
 294  * The reference count of the block is greater than one and the tree is
 295  * the block's owner tree. In this case, implicit back refs is used for
 296  * pointers in the block. Add full back refs for every pointers in the
 297  * block, increase lower level extents' reference counts. The original
 298  * implicit back refs are entailed to the new block.
 299  *
 300  * The reference count of the block is greater than one and the tree is
 301  * not the block's owner tree. Add implicit back refs for every pointer in
 302  * the new block, increase lower level extents' reference count.
 303  *
 304  * Back Reference Key composing:
 305  *
 306  * The key objectid corresponds to the first byte in the extent,
 307  * The key type is used to differentiate between types of back refs.
 308  * There are different meanings of the key offset for different types
 309  * of back refs.
 310  *
 311  * File extents can be referenced by:
 312  *
 313  * - multiple snapshots, subvolumes, or different generations in one subvol
 314  * - different files inside a single subvolume
 315  * - different offsets inside a file (bookend extents in file.c)
 316  *
 317  * The extent ref structure for the implicit back refs has fields for:
 318  *
 319  * - Objectid of the subvolume root
 320  * - objectid of the file holding the reference
 321  * - original offset in the file
 322  * - how many bookend extents
 323  *
 324  * The key offset for the implicit back refs is hash of the first
 325  * three fields.
 326  *
 327  * The extent ref structure for the full back refs has field for:
 328  *
 329  * - number of pointers in the tree leaf
 330  *
 331  * The key offset for the implicit back refs is the first byte of
 332  * the tree leaf
 333  *
 334  * When a file extent is allocated, The implicit back refs is used.
 335  * the fields are filled in:
 336  *
 337  *     (root_key.objectid, inode objectid, offset in file, 1)
 338  *
 339  * When a file extent is removed file truncation, we find the
 340  * corresponding implicit back refs and check the following fields:
 341  *
 342  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 343  *
 344  * Btree extents can be referenced by:
 345  *
 346  * - Different subvolumes
 347  *
 348  * Both the implicit back refs and the full back refs for tree blocks
 349  * only consist of key. The key offset for the implicit back refs is
 350  * objectid of block's owner tree. The key offset for the full back refs
 351  * is the first byte of parent block.
 352  *
 353  * When implicit back refs is used, information about the lowest key and
 354  * level of the tree block are required. These information are stored in
 355  * tree block info structure.
 356  */
 357
 358 /*
 359  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 360  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 361  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 362  */
 363 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 364                                      struct btrfs_extent_inline_ref *iref,
 365                                      enum btrfs_inline_ref_type is_data)
 366 {
 367         int type = btrfs_extent_inline_ref_type(eb, iref);
 368         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 369
 370         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 371             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 372             type == BTRFS_SHARED_DATA_REF_KEY ||
 373             type == BTRFS_EXTENT_DATA_REF_KEY) {
 374                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 375                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 376                                 return type;
 377                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 378                                 ASSERT(eb->fs_info);
 379                                 /*
 380                                  * Every shared one has parent tree block,
 381                                  * which must be aligned to sector size.
 382                                  */
 383                                 if (offset &&
 384                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 385                                         return type;
 386                         }
 387                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 388                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 389                                 return type;
 390                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 391                                 ASSERT(eb->fs_info);
 392                                 /*
 393                                  * Every shared one has parent tree block,
 394                                  * which must be aligned to sector size.
 395                                  */
 396                                 if (offset &&
 397                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 398                                         return type;
 399                         }
 400                 } else {
 401                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 402                         return type;
 403                 }
 404         }
 405
 406         btrfs_print_leaf((struct extent_buffer *)eb);
 407         btrfs_err(eb->fs_info,
 408                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 409                   eb->start, (unsigned long)iref, type);
 410         WARN_ON(1);
 411
 412         return BTRFS_REF_TYPE_INVALID;
 413 }
 414
 415 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 416 {
 417         u32 high_crc = ~(u32)0;
 418         u32 low_crc = ~(u32)0;
 419         __le64 lenum;
 420
 421         lenum = cpu_to_le64(root_objectid);
 422         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
 423         lenum = cpu_to_le64(owner);
 424         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 425         lenum = cpu_to_le64(offset);
 426         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 427
 428         return ((u64)high_crc << 31) ^ (u64)low_crc;
 429 }
 430
 431 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 432                                      struct btrfs_extent_data_ref *ref)
 433 {
 434         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 435                                     btrfs_extent_data_ref_objectid(leaf, ref),
 436                                     btrfs_extent_data_ref_offset(leaf, ref));
 437 }
 438
 439 static int match_extent_data_ref(struct extent_buffer *leaf,
 440                                  struct btrfs_extent_data_ref *ref,
 441                                  u64 root_objectid, u64 owner, u64 offset)
 442 {
 443         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 444             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 445             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 446                 return 0;
 447         return 1;
 448 }
 449
 450 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 451                                            struct btrfs_path *path,
 452                                            u64 bytenr, u64 parent,
 453                                            u64 root_objectid,
 454                                            u64 owner, u64 offset)
 455 {
 456         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 457         struct btrfs_key key;
 458         struct btrfs_extent_data_ref *ref;
 459         struct extent_buffer *leaf;
 460         u32 nritems;
 461         int ret;
 462         int recow;
 463         int err = -ENOENT;
 464
 465         key.objectid = bytenr;
 466         if (parent) {
 467                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 468                 key.offset = parent;
 469         } else {
 470                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 471                 key.offset = hash_extent_data_ref(root_objectid,
 472                                                   owner, offset);
 473         }
 474 again:
 475         recow = 0;
 476         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 477         if (ret < 0) {
 478                 err = ret;
 479                 goto fail;
 480         }
 481
 482         if (parent) {
 483                 if (!ret)
 484                         return 0;
 485                 goto fail;
 486         }
 487
 488         leaf = path->nodes[0];
 489         nritems = btrfs_header_nritems(leaf);
 490         while (1) {
 491                 if (path->slots[0] >= nritems) {
 492                         ret = btrfs_next_leaf(root, path);
 493                         if (ret < 0)
 494                                 err = ret;
 495                         if (ret)
 496                                 goto fail;
 497
 498                         leaf = path->nodes[0];
 499                         nritems = btrfs_header_nritems(leaf);
 500                         recow = 1;
 501                 }
 502
 503                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 504                 if (key.objectid != bytenr ||
 505                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 506                         goto fail;
 507
 508                 ref = btrfs_item_ptr(leaf, path->slots[0],
 509                                      struct btrfs_extent_data_ref);
 510
 511                 if (match_extent_data_ref(leaf, ref, root_objectid,
 512                                           owner, offset)) {
 513                         if (recow) {
 514                                 btrfs_release_path(path);
 515                                 goto again;
 516                         }
 517                         err = 0;
 518                         break;
 519                 }
 520                 path->slots[0]++;
 521         }
 522 fail:
 523         return err;
 524 }
 525
 526 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 527                                            struct btrfs_path *path,
 528                                            u64 bytenr, u64 parent,
 529                                            u64 root_objectid, u64 owner,
 530                                            u64 offset, int refs_to_add)
 531 {
 532         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 533         struct btrfs_key key;
 534         struct extent_buffer *leaf;
 535         u32 size;
 536         u32 num_refs;
 537         int ret;
 538
 539         key.objectid = bytenr;
 540         if (parent) {
 541                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 542                 key.offset = parent;
 543                 size = sizeof(struct btrfs_shared_data_ref);
 544         } else {
 545                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 546                 key.offset = hash_extent_data_ref(root_objectid,
 547                                                   owner, offset);
 548                 size = sizeof(struct btrfs_extent_data_ref);
 549         }
 550
 551         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 552         if (ret && ret != -EEXIST)
 553                 goto fail;
 554
 555         leaf = path->nodes[0];
 556         if (parent) {
 557                 struct btrfs_shared_data_ref *ref;
 558                 ref = btrfs_item_ptr(leaf, path->slots[0],
 559                                      struct btrfs_shared_data_ref);
 560                 if (ret == 0) {
 561                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 562                 } else {
 563                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 564                         num_refs += refs_to_add;
 565                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 566                 }
 567         } else {
 568                 struct btrfs_extent_data_ref *ref;
 569                 while (ret == -EEXIST) {
 570                         ref = btrfs_item_ptr(leaf, path->slots[0],
 571                                              struct btrfs_extent_data_ref);
 572                         if (match_extent_data_ref(leaf, ref, root_objectid,
 573                                                   owner, offset))
 574                                 break;
 575                         btrfs_release_path(path);
 576                         key.offset++;
 577                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 578                                                       size);
 579                         if (ret && ret != -EEXIST)
 580                                 goto fail;
 581
 582                         leaf = path->nodes[0];
 583                 }
 584                 ref = btrfs_item_ptr(leaf, path->slots[0],
 585                                      struct btrfs_extent_data_ref);
 586                 if (ret == 0) {
 587                         btrfs_set_extent_data_ref_root(leaf, ref,
 588                                                        root_objectid);
 589                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 590                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 591                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 592                 } else {
 593                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 594                         num_refs += refs_to_add;
 595                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 596                 }
 597         }
 598         btrfs_mark_buffer_dirty(leaf);
 599         ret = 0;
 600 fail:
 601         btrfs_release_path(path);
 602         return ret;
 603 }
 604
 605 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 606                                            struct btrfs_root *root,
 607                                            struct btrfs_path *path,
 608                                            int refs_to_drop)
 609 {
 610         struct btrfs_key key;
 611         struct btrfs_extent_data_ref *ref1 = NULL;
 612         struct btrfs_shared_data_ref *ref2 = NULL;
 613         struct extent_buffer *leaf;
 614         u32 num_refs = 0;
 615         int ret = 0;
 616
 617         leaf = path->nodes[0];
 618         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 619
 620         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 621                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 622                                       struct btrfs_extent_data_ref);
 623                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 624         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 625                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 626                                       struct btrfs_shared_data_ref);
 627                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 628         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 629                 btrfs_print_v0_err(trans->fs_info);
 630                 btrfs_abort_transaction(trans, -EINVAL);
 631                 return -EINVAL;
 632         } else {
 633                 BUG();
 634         }
 635
 636         BUG_ON(num_refs < refs_to_drop);
 637         num_refs -= refs_to_drop;
 638
 639         if (num_refs == 0) {
 640                 ret = btrfs_del_item(trans, root, path);
 641         } else {
 642                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 643                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 644                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 645                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 646                 btrfs_mark_buffer_dirty(leaf);
 647         }
 648         return ret;
 649 }
 650
 651 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 652                                           struct btrfs_extent_inline_ref *iref)
 653 {
 654         struct btrfs_key key;
 655         struct extent_buffer *leaf;
 656         struct btrfs_extent_data_ref *ref1;
 657         struct btrfs_shared_data_ref *ref2;
 658         u32 num_refs = 0;
 659         int type;
 660
 661         leaf = path->nodes[0];
 662         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 663
 664         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 665         if (iref) {
 666                 /*
 667                  * If type is invalid, we should have bailed out earlier than
 668                  * this call.
 669                  */
 670                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 671                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 672                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 673                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 674                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 675                 } else {
 676                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 677                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 678                 }
 679         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 680                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 681                                       struct btrfs_extent_data_ref);
 682                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 683         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 684                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 685                                       struct btrfs_shared_data_ref);
 686                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 687         } else {
 688                 WARN_ON(1);
 689         }
 690         return num_refs;
 691 }
 692
 693 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 694                                           struct btrfs_path *path,
 695                                           u64 bytenr, u64 parent,
 696                                           u64 root_objectid)
 697 {
 698         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 699         struct btrfs_key key;
 700         int ret;
 701
 702         key.objectid = bytenr;
 703         if (parent) {
 704                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 705                 key.offset = parent;
 706         } else {
 707                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 708                 key.offset = root_objectid;
 709         }
 710
 711         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 712         if (ret > 0)
 713                 ret = -ENOENT;
 714         return ret;
 715 }
 716
 717 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 718                                           struct btrfs_path *path,
 719                                           u64 bytenr, u64 parent,
 720                                           u64 root_objectid)
 721 {
 722         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 723         struct btrfs_key key;
 724         int ret;
 725
 726         key.objectid = bytenr;
 727         if (parent) {
 728                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 729                 key.offset = parent;
 730         } else {
 731                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 732                 key.offset = root_objectid;
 733         }
 734
 735         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 736         btrfs_release_path(path);
 737         return ret;
 738 }
 739
 740 static inline int extent_ref_type(u64 parent, u64 owner)
 741 {
 742         int type;
 743         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 744                 if (parent > 0)
 745                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 746                 else
 747                         type = BTRFS_TREE_BLOCK_REF_KEY;
 748         } else {
 749                 if (parent > 0)
 750                         type = BTRFS_SHARED_DATA_REF_KEY;
 751                 else
 752                         type = BTRFS_EXTENT_DATA_REF_KEY;
 753         }
 754         return type;
 755 }
 756
 757 static int find_next_key(struct btrfs_path *path, int level,
 758                          struct btrfs_key *key)
 759
 760 {
 761         for (; level < BTRFS_MAX_LEVEL; level++) {
 762                 if (!path->nodes[level])
 763                         break;
 764                 if (path->slots[level] + 1 >=
 765                     btrfs_header_nritems(path->nodes[level]))
 766                         continue;
 767                 if (level == 0)
 768                         btrfs_item_key_to_cpu(path->nodes[level], key,
 769                                               path->slots[level] + 1);
 770                 else
 771                         btrfs_node_key_to_cpu(path->nodes[level], key,
 772                                               path->slots[level] + 1);
 773                 return 0;
 774         }
 775         return 1;
 776 }
 777
 778 /*
 779  * look for inline back ref. if back ref is found, *ref_ret is set
 780  * to the address of inline back ref, and 0 is returned.
 781  *
 782  * if back ref isn't found, *ref_ret is set to the address where it
 783  * should be inserted, and -ENOENT is returned.
 784  *
 785  * if insert is true and there are too many inline back refs, the path
 786  * points to the extent item, and -EAGAIN is returned.
 787  *
 788  * NOTE: inline back refs are ordered in the same way that back ref
 789  *       items in the tree are ordered.
 790  */
 791 static noinline_for_stack
 792 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 793                                  struct btrfs_path *path,
 794                                  struct btrfs_extent_inline_ref **ref_ret,
 795                                  u64 bytenr, u64 num_bytes,
 796                                  u64 parent, u64 root_objectid,
 797                                  u64 owner, u64 offset, int insert)
 798 {
 799         struct btrfs_fs_info *fs_info = trans->fs_info;
 800         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 801         struct btrfs_key key;
 802         struct extent_buffer *leaf;
 803         struct btrfs_extent_item *ei;
 804         struct btrfs_extent_inline_ref *iref;
 805         u64 flags;
 806         u64 item_size;
 807         unsigned long ptr;
 808         unsigned long end;
 809         int extra_size;
 810         int type;
 811         int want;
 812         int ret;
 813         int err = 0;
 814         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 815         int needed;
 816
 817         key.objectid = bytenr;
 818         key.type = BTRFS_EXTENT_ITEM_KEY;
 819         key.offset = num_bytes;
 820
 821         want = extent_ref_type(parent, owner);
 822         if (insert) {
 823                 extra_size = btrfs_extent_inline_ref_size(want);
 824                 path->search_for_extension = 1;
 825                 path->keep_locks = 1;
 826         } else
 827                 extra_size = -1;
 828
 829         /*
 830          * Owner is our level, so we can just add one to get the level for the
 831          * block we are interested in.
 832          */
 833         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 834                 key.type = BTRFS_METADATA_ITEM_KEY;
 835                 key.offset = owner;
 836         }
 837
 838 again:
 839         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 840         if (ret < 0) {
 841                 err = ret;
 842                 goto out;
 843         }
 844
 845         /*
 846          * We may be a newly converted file system which still has the old fat
 847          * extent entries for metadata, so try and see if we have one of those.
 848          */
 849         if (ret > 0 && skinny_metadata) {
 850                 skinny_metadata = false;
 851                 if (path->slots[0]) {
 852                         path->slots[0]--;
 853                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 854                                               path->slots[0]);
 855                         if (key.objectid == bytenr &&
 856                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 857                             key.offset == num_bytes)
 858                                 ret = 0;
 859                 }
 860                 if (ret) {
 861                         key.objectid = bytenr;
 862                         key.type = BTRFS_EXTENT_ITEM_KEY;
 863                         key.offset = num_bytes;
 864                         btrfs_release_path(path);
 865                         goto again;
 866                 }
 867         }
 868
 869         if (ret && !insert) {
 870                 err = -ENOENT;
 871                 goto out;
 872         } else if (WARN_ON(ret)) {
 873                 err = -EIO;
 874                 goto out;
 875         }
 876
 877         leaf = path->nodes[0];
 878         item_size = btrfs_item_size(leaf, path->slots[0]);
 879         if (unlikely(item_size < sizeof(*ei))) {
 880                 err = -EINVAL;
 881                 btrfs_print_v0_err(fs_info);
 882                 btrfs_abort_transaction(trans, err);
 883                 goto out;
 884         }
 885
 886         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 887         flags = btrfs_extent_flags(leaf, ei);
 888
 889         ptr = (unsigned long)(ei + 1);
 890         end = (unsigned long)ei + item_size;
 891
 892         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 893                 ptr += sizeof(struct btrfs_tree_block_info);
 894                 BUG_ON(ptr > end);
 895         }
 896
 897         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 898                 needed = BTRFS_REF_TYPE_DATA;
 899         else
 900                 needed = BTRFS_REF_TYPE_BLOCK;
 901
 902         err = -ENOENT;
 903         while (1) {
 904                 if (ptr >= end) {
 905                         if (ptr > end) {
 906                                 err = -EUCLEAN;
 907                                 btrfs_print_leaf(path->nodes[0]);
 908                                 btrfs_crit(fs_info,
 909 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 910                                         path->slots[0], root_objectid, owner, offset, parent);
 911                         }
 912                         break;
 913                 }
 914                 iref = (struct btrfs_extent_inline_ref *)ptr;
 915                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 916                 if (type == BTRFS_REF_TYPE_INVALID) {
 917                         err = -EUCLEAN;
 918                         goto out;
 919                 }
 920
 921                 if (want < type)
 922                         break;
 923                 if (want > type) {
 924                         ptr += btrfs_extent_inline_ref_size(type);
 925                         continue;
 926                 }
 927
 928                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 929                         struct btrfs_extent_data_ref *dref;
 930                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 931                         if (match_extent_data_ref(leaf, dref, root_objectid,
 932                                                   owner, offset)) {
 933                                 err = 0;
 934                                 break;
 935                         }
 936                         if (hash_extent_data_ref_item(leaf, dref) <
 937                             hash_extent_data_ref(root_objectid, owner, offset))
 938                                 break;
 939                 } else {
 940                         u64 ref_offset;
 941                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 942                         if (parent > 0) {
 943                                 if (parent == ref_offset) {
 944                                         err = 0;
 945                                         break;
 946                                 }
 947                                 if (ref_offset < parent)
 948                                         break;
 949                         } else {
 950                                 if (root_objectid == ref_offset) {
 951                                         err = 0;
 952                                         break;
 953                                 }
 954                                 if (ref_offset < root_objectid)
 955                                         break;
 956                         }
 957                 }
 958                 ptr += btrfs_extent_inline_ref_size(type);
 959         }
 960         if (err == -ENOENT && insert) {
 961                 if (item_size + extra_size >=
 962                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 963                         err = -EAGAIN;
 964                         goto out;
 965                 }
 966                 /*
 967                  * To add new inline back ref, we have to make sure
 968                  * there is no corresponding back ref item.
 969                  * For simplicity, we just do not add new inline back
 970                  * ref if there is any kind of item for this block
 971                  */
 972                 if (find_next_key(path, 0, &key) == 0 &&
 973                     key.objectid == bytenr &&
 974                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 975                         err = -EAGAIN;
 976                         goto out;
 977                 }
 978         }
 979         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 980 out:
 981         if (insert) {
 982                 path->keep_locks = 0;
 983                 path->search_for_extension = 0;
 984                 btrfs_unlock_up_safe(path, 1);
 985         }
 986         return err;
 987 }
 988
 989 /*
 990  * helper to add new inline back ref
 991  */
 992 static noinline_for_stack
 993 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
 994                                  struct btrfs_path *path,
 995                                  struct btrfs_extent_inline_ref *iref,
 996                                  u64 parent, u64 root_objectid,
 997                                  u64 owner, u64 offset, int refs_to_add,
 998                                  struct btrfs_delayed_extent_op *extent_op)
 999 {
1000         struct extent_buffer *leaf;
1001         struct btrfs_extent_item *ei;
1002         unsigned long ptr;
1003         unsigned long end;
1004         unsigned long item_offset;
1005         u64 refs;
1006         int size;
1007         int type;
1008
1009         leaf = path->nodes[0];
1010         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1011         item_offset = (unsigned long)iref - (unsigned long)ei;
1012
1013         type = extent_ref_type(parent, owner);
1014         size = btrfs_extent_inline_ref_size(type);
1015
1016         btrfs_extend_item(path, size);
1017
1018         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1019         refs = btrfs_extent_refs(leaf, ei);
1020         refs += refs_to_add;
1021         btrfs_set_extent_refs(leaf, ei, refs);
1022         if (extent_op)
1023                 __run_delayed_extent_op(extent_op, leaf, ei);
1024
1025         ptr = (unsigned long)ei + item_offset;
1026         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1027         if (ptr < end - size)
1028                 memmove_extent_buffer(leaf, ptr + size, ptr,
1029                                       end - size - ptr);
1030
1031         iref = (struct btrfs_extent_inline_ref *)ptr;
1032         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1033         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1034                 struct btrfs_extent_data_ref *dref;
1035                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1036                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1037                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1038                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1039                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1040         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1041                 struct btrfs_shared_data_ref *sref;
1042                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1043                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1044                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1045         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1046                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1047         } else {
1048                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1049         }
1050         btrfs_mark_buffer_dirty(leaf);
1051 }
1052
1053 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1054                                  struct btrfs_path *path,
1055                                  struct btrfs_extent_inline_ref **ref_ret,
1056                                  u64 bytenr, u64 num_bytes, u64 parent,
1057                                  u64 root_objectid, u64 owner, u64 offset)
1058 {
1059         int ret;
1060
1061         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1062                                            num_bytes, parent, root_objectid,
1063                                            owner, offset, 0);
1064         if (ret != -ENOENT)
1065                 return ret;
1066
1067         btrfs_release_path(path);
1068         *ref_ret = NULL;
1069
1070         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1071                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1072                                             root_objectid);
1073         } else {
1074                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1075                                              root_objectid, owner, offset);
1076         }
1077         return ret;
1078 }
1079
1080 /*
1081  * helper to update/remove inline back ref
1082  */
1083 static noinline_for_stack
1084 void update_inline_extent_backref(struct btrfs_path *path,
1085                                   struct btrfs_extent_inline_ref *iref,
1086                                   int refs_to_mod,
1087                                   struct btrfs_delayed_extent_op *extent_op)
1088 {
1089         struct extent_buffer *leaf = path->nodes[0];
1090         struct btrfs_extent_item *ei;
1091         struct btrfs_extent_data_ref *dref = NULL;
1092         struct btrfs_shared_data_ref *sref = NULL;
1093         unsigned long ptr;
1094         unsigned long end;
1095         u32 item_size;
1096         int size;
1097         int type;
1098         u64 refs;
1099
1100         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1101         refs = btrfs_extent_refs(leaf, ei);
1102         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1103         refs += refs_to_mod;
1104         btrfs_set_extent_refs(leaf, ei, refs);
1105         if (extent_op)
1106                 __run_delayed_extent_op(extent_op, leaf, ei);
1107
1108         /*
1109          * If type is invalid, we should have bailed out after
1110          * lookup_inline_extent_backref().
1111          */
1112         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1113         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1114
1115         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1116                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1117                 refs = btrfs_extent_data_ref_count(leaf, dref);
1118         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1119                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1120                 refs = btrfs_shared_data_ref_count(leaf, sref);
1121         } else {
1122                 refs = 1;
1123                 BUG_ON(refs_to_mod != -1);
1124         }
1125
1126         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1127         refs += refs_to_mod;
1128
1129         if (refs > 0) {
1130                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1131                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1132                 else
1133                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1134         } else {
1135                 size =  btrfs_extent_inline_ref_size(type);
1136                 item_size = btrfs_item_size(leaf, path->slots[0]);
1137                 ptr = (unsigned long)iref;
1138                 end = (unsigned long)ei + item_size;
1139                 if (ptr + size < end)
1140                         memmove_extent_buffer(leaf, ptr, ptr + size,
1141                                               end - ptr - size);
1142                 item_size -= size;
1143                 btrfs_truncate_item(path, item_size, 1);
1144         }
1145         btrfs_mark_buffer_dirty(leaf);
1146 }
1147
1148 static noinline_for_stack
1149 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1150                                  struct btrfs_path *path,
1151                                  u64 bytenr, u64 num_bytes, u64 parent,
1152                                  u64 root_objectid, u64 owner,
1153                                  u64 offset, int refs_to_add,
1154                                  struct btrfs_delayed_extent_op *extent_op)
1155 {
1156         struct btrfs_extent_inline_ref *iref;
1157         int ret;
1158
1159         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1160                                            num_bytes, parent, root_objectid,
1161                                            owner, offset, 1);
1162         if (ret == 0) {
1163                 /*
1164                  * We're adding refs to a tree block we already own, this
1165                  * should not happen at all.
1166                  */
1167                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1168                         btrfs_crit(trans->fs_info,
1169 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu",
1170                                    bytenr, num_bytes, root_objectid);
1171                         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
1172                                 WARN_ON(1);
1173                                 btrfs_crit(trans->fs_info,
1174                         "path->slots[0]=%d path->nodes[0]:", path->slots[0]);
1175                                 btrfs_print_leaf(path->nodes[0]);
1176                         }
1177                         return -EUCLEAN;
1178                 }
1179                 update_inline_extent_backref(path, iref, refs_to_add, extent_op);
1180         } else if (ret == -ENOENT) {
1181                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1182                                             root_objectid, owner, offset,
1183                                             refs_to_add, extent_op);
1184                 ret = 0;
1185         }
1186         return ret;
1187 }
1188
1189 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1190                                  struct btrfs_root *root,
1191                                  struct btrfs_path *path,
1192                                  struct btrfs_extent_inline_ref *iref,
1193                                  int refs_to_drop, int is_data)
1194 {
1195         int ret = 0;
1196
1197         BUG_ON(!is_data && refs_to_drop != 1);
1198         if (iref)
1199                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL);
1200         else if (is_data)
1201                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1202         else
1203                 ret = btrfs_del_item(trans, root, path);
1204         return ret;
1205 }
1206
1207 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1208                                u64 *discarded_bytes)
1209 {
1210         int j, ret = 0;
1211         u64 bytes_left, end;
1212         u64 aligned_start = ALIGN(start, 1 << 9);
1213
1214         if (WARN_ON(start != aligned_start)) {
1215                 len -= aligned_start - start;
1216                 len = round_down(len, 1 << 9);
1217                 start = aligned_start;
1218         }
1219
1220         *discarded_bytes = 0;
1221
1222         if (!len)
1223                 return 0;
1224
1225         end = start + len;
1226         bytes_left = len;
1227
1228         /* Skip any superblocks on this device. */
1229         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1230                 u64 sb_start = btrfs_sb_offset(j);
1231                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1232                 u64 size = sb_start - start;
1233
1234                 if (!in_range(sb_start, start, bytes_left) &&
1235                     !in_range(sb_end, start, bytes_left) &&
1236                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1237                         continue;
1238
1239                 /*
1240                  * Superblock spans beginning of range.  Adjust start and
1241                  * try again.
1242                  */
1243                 if (sb_start <= start) {
1244                         start += sb_end - start;
1245                         if (start > end) {
1246                                 bytes_left = 0;
1247                                 break;
1248                         }
1249                         bytes_left = end - start;
1250                         continue;
1251                 }
1252
1253                 if (size) {
1254                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1255                                                    GFP_NOFS);
1256                         if (!ret)
1257                                 *discarded_bytes += size;
1258                         else if (ret != -EOPNOTSUPP)
1259                                 return ret;
1260                 }
1261
1262                 start = sb_end;
1263                 if (start > end) {
1264                         bytes_left = 0;
1265                         break;
1266                 }
1267                 bytes_left = end - start;
1268         }
1269
1270         if (bytes_left) {
1271                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1272                                            GFP_NOFS);
1273                 if (!ret)
1274                         *discarded_bytes += bytes_left;
1275         }
1276         return ret;
1277 }
1278
1279 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1280 {
1281         struct btrfs_device *dev = stripe->dev;
1282         struct btrfs_fs_info *fs_info = dev->fs_info;
1283         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1284         u64 phys = stripe->physical;
1285         u64 len = stripe->length;
1286         u64 discarded = 0;
1287         int ret = 0;
1288
1289         /* Zone reset on a zoned filesystem */
1290         if (btrfs_can_zone_reset(dev, phys, len)) {
1291                 u64 src_disc;
1292
1293                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1294                 if (ret)
1295                         goto out;
1296
1297                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1298                     dev != dev_replace->srcdev)
1299                         goto out;
1300
1301                 src_disc = discarded;
1302
1303                 /* Send to replace target as well */
1304                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1305                                               &discarded);
1306                 discarded += src_disc;
1307         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1308                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1309         } else {
1310                 ret = 0;
1311                 *bytes = 0;
1312         }
1313
1314 out:
1315         *bytes = discarded;
1316         return ret;
1317 }
1318
1319 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1320                          u64 num_bytes, u64 *actual_bytes)
1321 {
1322         int ret = 0;
1323         u64 discarded_bytes = 0;
1324         u64 end = bytenr + num_bytes;
1325         u64 cur = bytenr;
1326
1327         /*
1328          * Avoid races with device replace and make sure the devices in the
1329          * stripes don't go away while we are discarding.
1330          */
1331         btrfs_bio_counter_inc_blocked(fs_info);
1332         while (cur < end) {
1333                 struct btrfs_discard_stripe *stripes;
1334                 unsigned int num_stripes;
1335                 int i;
1336
1337                 num_bytes = end - cur;
1338                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1339                 if (IS_ERR(stripes)) {
1340                         ret = PTR_ERR(stripes);
1341                         if (ret == -EOPNOTSUPP)
1342                                 ret = 0;
1343                         break;
1344                 }
1345
1346                 for (i = 0; i < num_stripes; i++) {
1347                         struct btrfs_discard_stripe *stripe = stripes + i;
1348                         u64 bytes;
1349
1350                         if (!stripe->dev->bdev) {
1351                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1352                                 continue;
1353                         }
1354
1355                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1356                                         &stripe->dev->dev_state))
1357                                 continue;
1358
1359                         ret = do_discard_extent(stripe, &bytes);
1360                         if (ret) {
1361                                 /*
1362                                  * Keep going if discard is not supported by the
1363                                  * device.
1364                                  */
1365                                 if (ret != -EOPNOTSUPP)
1366                                         break;
1367                                 ret = 0;
1368                         } else {
1369                                 discarded_bytes += bytes;
1370                         }
1371                 }
1372                 kfree(stripes);
1373                 if (ret)
1374                         break;
1375                 cur += num_bytes;
1376         }
1377         btrfs_bio_counter_dec(fs_info);
1378         if (actual_bytes)
1379                 *actual_bytes = discarded_bytes;
1380         return ret;
1381 }
1382
1383 /* Can return -ENOMEM */
1384 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1385                          struct btrfs_ref *generic_ref)
1386 {
1387         struct btrfs_fs_info *fs_info = trans->fs_info;
1388         int ret;
1389
1390         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1391                generic_ref->action);
1392         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1393                generic_ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID);
1394
1395         if (generic_ref->type == BTRFS_REF_METADATA)
1396                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1397         else
1398                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1399
1400         btrfs_ref_tree_mod(fs_info, generic_ref);
1401
1402         return ret;
1403 }
1404
1405 /*
1406  * __btrfs_inc_extent_ref - insert backreference for a given extent
1407  *
1408  * The counterpart is in __btrfs_free_extent(), with examples and more details
1409  * how it works.
1410  *
1411  * @trans:          Handle of transaction
1412  *
1413  * @node:           The delayed ref node used to get the bytenr/length for
1414  *                  extent whose references are incremented.
1415  *
1416  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1417  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1418  *                  bytenr of the parent block. Since new extents are always
1419  *                  created with indirect references, this will only be the case
1420  *                  when relocating a shared extent. In that case, root_objectid
1421  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1422  *                  be 0
1423  *
1424  * @root_objectid:  The id of the root where this modification has originated,
1425  *                  this can be either one of the well-known metadata trees or
1426  *                  the subvolume id which references this extent.
1427  *
1428  * @owner:          For data extents it is the inode number of the owning file.
1429  *                  For metadata extents this parameter holds the level in the
1430  *                  tree of the extent.
1431  *
1432  * @offset:         For metadata extents the offset is ignored and is currently
1433  *                  always passed as 0. For data extents it is the fileoffset
1434  *                  this extent belongs to.
1435  *
1436  * @refs_to_add     Number of references to add
1437  *
1438  * @extent_op       Pointer to a structure, holding information necessary when
1439  *                  updating a tree block's flags
1440  *
1441  */
1442 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1443                                   struct btrfs_delayed_ref_node *node,
1444                                   u64 parent, u64 root_objectid,
1445                                   u64 owner, u64 offset, int refs_to_add,
1446                                   struct btrfs_delayed_extent_op *extent_op)
1447 {
1448         struct btrfs_path *path;
1449         struct extent_buffer *leaf;
1450         struct btrfs_extent_item *item;
1451         struct btrfs_key key;
1452         u64 bytenr = node->bytenr;
1453         u64 num_bytes = node->num_bytes;
1454         u64 refs;
1455         int ret;
1456
1457         path = btrfs_alloc_path();
1458         if (!path)
1459                 return -ENOMEM;
1460
1461         /* this will setup the path even if it fails to insert the back ref */
1462         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1463                                            parent, root_objectid, owner,
1464                                            offset, refs_to_add, extent_op);
1465         if ((ret < 0 && ret != -EAGAIN) || !ret)
1466                 goto out;
1467
1468         /*
1469          * Ok we had -EAGAIN which means we didn't have space to insert and
1470          * inline extent ref, so just update the reference count and add a
1471          * normal backref.
1472          */
1473         leaf = path->nodes[0];
1474         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1475         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1476         refs = btrfs_extent_refs(leaf, item);
1477         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1478         if (extent_op)
1479                 __run_delayed_extent_op(extent_op, leaf, item);
1480
1481         btrfs_mark_buffer_dirty(leaf);
1482         btrfs_release_path(path);
1483
1484         /* now insert the actual backref */
1485         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1486                 BUG_ON(refs_to_add != 1);
1487                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1488                                             root_objectid);
1489         } else {
1490                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1491                                              root_objectid, owner, offset,
1492                                              refs_to_add);
1493         }
1494         if (ret)
1495                 btrfs_abort_transaction(trans, ret);
1496 out:
1497         btrfs_free_path(path);
1498         return ret;
1499 }
1500
1501 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1502                                 struct btrfs_delayed_ref_node *node,
1503                                 struct btrfs_delayed_extent_op *extent_op,
1504                                 int insert_reserved)
1505 {
1506         int ret = 0;
1507         struct btrfs_delayed_data_ref *ref;
1508         struct btrfs_key ins;
1509         u64 parent = 0;
1510         u64 ref_root = 0;
1511         u64 flags = 0;
1512
1513         ins.objectid = node->bytenr;
1514         ins.offset = node->num_bytes;
1515         ins.type = BTRFS_EXTENT_ITEM_KEY;
1516
1517         ref = btrfs_delayed_node_to_data_ref(node);
1518         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1519
1520         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1521                 parent = ref->parent;
1522         ref_root = ref->root;
1523
1524         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1525                 if (extent_op)
1526                         flags |= extent_op->flags_to_set;
1527                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
1528                                                  flags, ref->objectid,
1529                                                  ref->offset, &ins,
1530                                                  node->ref_mod);
1531         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1532                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1533                                              ref->objectid, ref->offset,
1534                                              node->ref_mod, extent_op);
1535         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1536                 ret = __btrfs_free_extent(trans, node, parent,
1537                                           ref_root, ref->objectid,
1538                                           ref->offset, node->ref_mod,
1539                                           extent_op);
1540         } else {
1541                 BUG();
1542         }
1543         return ret;
1544 }
1545
1546 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1547                                     struct extent_buffer *leaf,
1548                                     struct btrfs_extent_item *ei)
1549 {
1550         u64 flags = btrfs_extent_flags(leaf, ei);
1551         if (extent_op->update_flags) {
1552                 flags |= extent_op->flags_to_set;
1553                 btrfs_set_extent_flags(leaf, ei, flags);
1554         }
1555
1556         if (extent_op->update_key) {
1557                 struct btrfs_tree_block_info *bi;
1558                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1559                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1560                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1561         }
1562 }
1563
1564 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1565                                  struct btrfs_delayed_ref_head *head,
1566                                  struct btrfs_delayed_extent_op *extent_op)
1567 {
1568         struct btrfs_fs_info *fs_info = trans->fs_info;
1569         struct btrfs_root *root;
1570         struct btrfs_key key;
1571         struct btrfs_path *path;
1572         struct btrfs_extent_item *ei;
1573         struct extent_buffer *leaf;
1574         u32 item_size;
1575         int ret;
1576         int err = 0;
1577         int metadata = 1;
1578
1579         if (TRANS_ABORTED(trans))
1580                 return 0;
1581
1582         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1583                 metadata = 0;
1584
1585         path = btrfs_alloc_path();
1586         if (!path)
1587                 return -ENOMEM;
1588
1589         key.objectid = head->bytenr;
1590
1591         if (metadata) {
1592                 key.type = BTRFS_METADATA_ITEM_KEY;
1593                 key.offset = extent_op->level;
1594         } else {
1595                 key.type = BTRFS_EXTENT_ITEM_KEY;
1596                 key.offset = head->num_bytes;
1597         }
1598
1599         root = btrfs_extent_root(fs_info, key.objectid);
1600 again:
1601         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1602         if (ret < 0) {
1603                 err = ret;
1604                 goto out;
1605         }
1606         if (ret > 0) {
1607                 if (metadata) {
1608                         if (path->slots[0] > 0) {
1609                                 path->slots[0]--;
1610                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1611                                                       path->slots[0]);
1612                                 if (key.objectid == head->bytenr &&
1613                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1614                                     key.offset == head->num_bytes)
1615                                         ret = 0;
1616                         }
1617                         if (ret > 0) {
1618                                 btrfs_release_path(path);
1619                                 metadata = 0;
1620
1621                                 key.objectid = head->bytenr;
1622                                 key.offset = head->num_bytes;
1623                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1624                                 goto again;
1625                         }
1626                 } else {
1627                         err = -EIO;
1628                         goto out;
1629                 }
1630         }
1631
1632         leaf = path->nodes[0];
1633         item_size = btrfs_item_size(leaf, path->slots[0]);
1634
1635         if (unlikely(item_size < sizeof(*ei))) {
1636                 err = -EINVAL;
1637                 btrfs_print_v0_err(fs_info);
1638                 btrfs_abort_transaction(trans, err);
1639                 goto out;
1640         }
1641
1642         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1643         __run_delayed_extent_op(extent_op, leaf, ei);
1644
1645         btrfs_mark_buffer_dirty(leaf);
1646 out:
1647         btrfs_free_path(path);
1648         return err;
1649 }
1650
1651 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1652                                 struct btrfs_delayed_ref_node *node,
1653                                 struct btrfs_delayed_extent_op *extent_op,
1654                                 int insert_reserved)
1655 {
1656         int ret = 0;
1657         struct btrfs_delayed_tree_ref *ref;
1658         u64 parent = 0;
1659         u64 ref_root = 0;
1660
1661         ref = btrfs_delayed_node_to_tree_ref(node);
1662         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1663
1664         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1665                 parent = ref->parent;
1666         ref_root = ref->root;
1667
1668         if (node->ref_mod != 1) {
1669                 btrfs_err(trans->fs_info,
1670         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
1671                           node->bytenr, node->ref_mod, node->action, ref_root,
1672                           parent);
1673                 return -EIO;
1674         }
1675         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1676                 BUG_ON(!extent_op || !extent_op->update_flags);
1677                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1678         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1679                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1680                                              ref->level, 0, 1, extent_op);
1681         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1682                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1683                                           ref->level, 0, 1, extent_op);
1684         } else {
1685                 BUG();
1686         }
1687         return ret;
1688 }
1689
1690 /* helper function to actually process a single delayed ref entry */
1691 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1692                                struct btrfs_delayed_ref_node *node,
1693                                struct btrfs_delayed_extent_op *extent_op,
1694                                int insert_reserved)
1695 {
1696         int ret = 0;
1697
1698         if (TRANS_ABORTED(trans)) {
1699                 if (insert_reserved)
1700                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1701                 return 0;
1702         }
1703
1704         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1705             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1706                 ret = run_delayed_tree_ref(trans, node, extent_op,
1707                                            insert_reserved);
1708         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1709                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1710                 ret = run_delayed_data_ref(trans, node, extent_op,
1711                                            insert_reserved);
1712         else
1713                 BUG();
1714         if (ret && insert_reserved)
1715                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1716         if (ret < 0)
1717                 btrfs_err(trans->fs_info,
1718 "failed to run delayed ref for logical %llu num_bytes %llu type %u action %u ref_mod %d: %d",
1719                           node->bytenr, node->num_bytes, node->type,
1720                           node->action, node->ref_mod, ret);
1721         return ret;
1722 }
1723
1724 static inline struct btrfs_delayed_ref_node *
1725 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1726 {
1727         struct btrfs_delayed_ref_node *ref;
1728
1729         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1730                 return NULL;
1731
1732         /*
1733          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1734          * This is to prevent a ref count from going down to zero, which deletes
1735          * the extent item from the extent tree, when there still are references
1736          * to add, which would fail because they would not find the extent item.
1737          */
1738         if (!list_empty(&head->ref_add_list))
1739                 return list_first_entry(&head->ref_add_list,
1740                                 struct btrfs_delayed_ref_node, add_list);
1741
1742         ref = rb_entry(rb_first_cached(&head->ref_tree),
1743                        struct btrfs_delayed_ref_node, ref_node);
1744         ASSERT(list_empty(&ref->add_list));
1745         return ref;
1746 }
1747
1748 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1749                                       struct btrfs_delayed_ref_head *head)
1750 {
1751         spin_lock(&delayed_refs->lock);
1752         head->processing = 0;
1753         delayed_refs->num_heads_ready++;
1754         spin_unlock(&delayed_refs->lock);
1755         btrfs_delayed_ref_unlock(head);
1756 }
1757
1758 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1759                                 struct btrfs_delayed_ref_head *head)
1760 {
1761         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1762
1763         if (!extent_op)
1764                 return NULL;
1765
1766         if (head->must_insert_reserved) {
1767                 head->extent_op = NULL;
1768                 btrfs_free_delayed_extent_op(extent_op);
1769                 return NULL;
1770         }
1771         return extent_op;
1772 }
1773
1774 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1775                                      struct btrfs_delayed_ref_head *head)
1776 {
1777         struct btrfs_delayed_extent_op *extent_op;
1778         int ret;
1779
1780         extent_op = cleanup_extent_op(head);
1781         if (!extent_op)
1782                 return 0;
1783         head->extent_op = NULL;
1784         spin_unlock(&head->lock);
1785         ret = run_delayed_extent_op(trans, head, extent_op);
1786         btrfs_free_delayed_extent_op(extent_op);
1787         return ret ? ret : 1;
1788 }
1789
1790 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1791                                   struct btrfs_delayed_ref_root *delayed_refs,
1792                                   struct btrfs_delayed_ref_head *head)
1793 {
1794         int nr_items = 1;       /* Dropping this ref head update. */
1795
1796         /*
1797          * We had csum deletions accounted for in our delayed refs rsv, we need
1798          * to drop the csum leaves for this update from our delayed_refs_rsv.
1799          */
1800         if (head->total_ref_mod < 0 && head->is_data) {
1801                 spin_lock(&delayed_refs->lock);
1802                 delayed_refs->pending_csums -= head->num_bytes;
1803                 spin_unlock(&delayed_refs->lock);
1804                 nr_items += btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1805         }
1806
1807         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
1808 }
1809
1810 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1811                             struct btrfs_delayed_ref_head *head)
1812 {
1813
1814         struct btrfs_fs_info *fs_info = trans->fs_info;
1815         struct btrfs_delayed_ref_root *delayed_refs;
1816         int ret;
1817
1818         delayed_refs = &trans->transaction->delayed_refs;
1819
1820         ret = run_and_cleanup_extent_op(trans, head);
1821         if (ret < 0) {
1822                 unselect_delayed_ref_head(delayed_refs, head);
1823                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1824                 return ret;
1825         } else if (ret) {
1826                 return ret;
1827         }
1828
1829         /*
1830          * Need to drop our head ref lock and re-acquire the delayed ref lock
1831          * and then re-check to make sure nobody got added.
1832          */
1833         spin_unlock(&head->lock);
1834         spin_lock(&delayed_refs->lock);
1835         spin_lock(&head->lock);
1836         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1837                 spin_unlock(&head->lock);
1838                 spin_unlock(&delayed_refs->lock);
1839                 return 1;
1840         }
1841         btrfs_delete_ref_head(delayed_refs, head);
1842         spin_unlock(&head->lock);
1843         spin_unlock(&delayed_refs->lock);
1844
1845         if (head->must_insert_reserved) {
1846                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1847                 if (head->is_data) {
1848                         struct btrfs_root *csum_root;
1849
1850                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1851                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1852                                               head->num_bytes);
1853                 }
1854         }
1855
1856         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1857
1858         trace_run_delayed_ref_head(fs_info, head, 0);
1859         btrfs_delayed_ref_unlock(head);
1860         btrfs_put_delayed_ref_head(head);
1861         return ret;
1862 }
1863
1864 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1865                                         struct btrfs_trans_handle *trans)
1866 {
1867         struct btrfs_delayed_ref_root *delayed_refs =
1868                 &trans->transaction->delayed_refs;
1869         struct btrfs_delayed_ref_head *head = NULL;
1870         int ret;
1871
1872         spin_lock(&delayed_refs->lock);
1873         head = btrfs_select_ref_head(delayed_refs);
1874         if (!head) {
1875                 spin_unlock(&delayed_refs->lock);
1876                 return head;
1877         }
1878
1879         /*
1880          * Grab the lock that says we are going to process all the refs for
1881          * this head
1882          */
1883         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1884         spin_unlock(&delayed_refs->lock);
1885
1886         /*
1887          * We may have dropped the spin lock to get the head mutex lock, and
1888          * that might have given someone else time to free the head.  If that's
1889          * true, it has been removed from our list and we can move on.
1890          */
1891         if (ret == -EAGAIN)
1892                 head = ERR_PTR(-EAGAIN);
1893
1894         return head;
1895 }
1896
1897 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1898                                     struct btrfs_delayed_ref_head *locked_ref,
1899                                     unsigned long *run_refs)
1900 {
1901         struct btrfs_fs_info *fs_info = trans->fs_info;
1902         struct btrfs_delayed_ref_root *delayed_refs;
1903         struct btrfs_delayed_extent_op *extent_op;
1904         struct btrfs_delayed_ref_node *ref;
1905         int must_insert_reserved = 0;
1906         int ret;
1907
1908         delayed_refs = &trans->transaction->delayed_refs;
1909
1910         lockdep_assert_held(&locked_ref->mutex);
1911         lockdep_assert_held(&locked_ref->lock);
1912
1913         while ((ref = select_delayed_ref(locked_ref))) {
1914                 if (ref->seq &&
1915                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1916                         spin_unlock(&locked_ref->lock);
1917                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1918                         return -EAGAIN;
1919                 }
1920
1921                 (*run_refs)++;
1922                 ref->in_tree = 0;
1923                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1924                 RB_CLEAR_NODE(&ref->ref_node);
1925                 if (!list_empty(&ref->add_list))
1926                         list_del(&ref->add_list);
1927                 /*
1928                  * When we play the delayed ref, also correct the ref_mod on
1929                  * head
1930                  */
1931                 switch (ref->action) {
1932                 case BTRFS_ADD_DELAYED_REF:
1933                 case BTRFS_ADD_DELAYED_EXTENT:
1934                         locked_ref->ref_mod -= ref->ref_mod;
1935                         break;
1936                 case BTRFS_DROP_DELAYED_REF:
1937                         locked_ref->ref_mod += ref->ref_mod;
1938                         break;
1939                 default:
1940                         WARN_ON(1);
1941                 }
1942                 atomic_dec(&delayed_refs->num_entries);
1943
1944                 /*
1945                  * Record the must_insert_reserved flag before we drop the
1946                  * spin lock.
1947                  */
1948                 must_insert_reserved = locked_ref->must_insert_reserved;
1949                 locked_ref->must_insert_reserved = 0;
1950
1951                 extent_op = locked_ref->extent_op;
1952                 locked_ref->extent_op = NULL;
1953                 spin_unlock(&locked_ref->lock);
1954
1955                 ret = run_one_delayed_ref(trans, ref, extent_op,
1956                                           must_insert_reserved);
1957
1958                 btrfs_free_delayed_extent_op(extent_op);
1959                 if (ret) {
1960                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1961                         btrfs_put_delayed_ref(ref);
1962                         return ret;
1963                 }
1964
1965                 btrfs_put_delayed_ref(ref);
1966                 cond_resched();
1967
1968                 spin_lock(&locked_ref->lock);
1969                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
1970         }
1971
1972         return 0;
1973 }
1974
1975 /*
1976  * Returns 0 on success or if called with an already aborted transaction.
1977  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
1978  */
1979 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
1980                                              unsigned long nr)
1981 {
1982         struct btrfs_fs_info *fs_info = trans->fs_info;
1983         struct btrfs_delayed_ref_root *delayed_refs;
1984         struct btrfs_delayed_ref_head *locked_ref = NULL;
1985         ktime_t start = ktime_get();
1986         int ret;
1987         unsigned long count = 0;
1988         unsigned long actual_count = 0;
1989
1990         delayed_refs = &trans->transaction->delayed_refs;
1991         do {
1992                 if (!locked_ref) {
1993                         locked_ref = btrfs_obtain_ref_head(trans);
1994                         if (IS_ERR_OR_NULL(locked_ref)) {
1995                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
1996                                         continue;
1997                                 } else {
1998                                         break;
1999                                 }
2000                         }
2001                         count++;
2002                 }
2003                 /*
2004                  * We need to try and merge add/drops of the same ref since we
2005                  * can run into issues with relocate dropping the implicit ref
2006                  * and then it being added back again before the drop can
2007                  * finish.  If we merged anything we need to re-loop so we can
2008                  * get a good ref.
2009                  * Or we can get node references of the same type that weren't
2010                  * merged when created due to bumps in the tree mod seq, and
2011                  * we need to merge them to prevent adding an inline extent
2012                  * backref before dropping it (triggering a BUG_ON at
2013                  * insert_inline_extent_backref()).
2014                  */
2015                 spin_lock(&locked_ref->lock);
2016                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2017
2018                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2019                                                       &actual_count);
2020                 if (ret < 0 && ret != -EAGAIN) {
2021                         /*
2022                          * Error, btrfs_run_delayed_refs_for_head already
2023                          * unlocked everything so just bail out
2024                          */
2025                         return ret;
2026                 } else if (!ret) {
2027                         /*
2028                          * Success, perform the usual cleanup of a processed
2029                          * head
2030                          */
2031                         ret = cleanup_ref_head(trans, locked_ref);
2032                         if (ret > 0 ) {
2033                                 /* We dropped our lock, we need to loop. */
2034                                 ret = 0;
2035                                 continue;
2036                         } else if (ret) {
2037                                 return ret;
2038                         }
2039                 }
2040
2041                 /*
2042                  * Either success case or btrfs_run_delayed_refs_for_head
2043                  * returned -EAGAIN, meaning we need to select another head
2044                  */
2045
2046                 locked_ref = NULL;
2047                 cond_resched();
2048         } while ((nr != -1 && count < nr) || locked_ref);
2049
2050         /*
2051          * We don't want to include ref heads since we can have empty ref heads
2052          * and those will drastically skew our runtime down since we just do
2053          * accounting, no actual extent tree updates.
2054          */
2055         if (actual_count > 0) {
2056                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2057                 u64 avg;
2058
2059                 /*
2060                  * We weigh the current average higher than our current runtime
2061                  * to avoid large swings in the average.
2062                  */
2063                 spin_lock(&delayed_refs->lock);
2064                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2065                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2066                 spin_unlock(&delayed_refs->lock);
2067         }
2068         return 0;
2069 }
2070
2071 #ifdef SCRAMBLE_DELAYED_REFS
2072 /*
2073  * Normally delayed refs get processed in ascending bytenr order. This
2074  * correlates in most cases to the order added. To expose dependencies on this
2075  * order, we start to process the tree in the middle instead of the beginning
2076  */
2077 static u64 find_middle(struct rb_root *root)
2078 {
2079         struct rb_node *n = root->rb_node;
2080         struct btrfs_delayed_ref_node *entry;
2081         int alt = 1;
2082         u64 middle;
2083         u64 first = 0, last = 0;
2084
2085         n = rb_first(root);
2086         if (n) {
2087                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2088                 first = entry->bytenr;
2089         }
2090         n = rb_last(root);
2091         if (n) {
2092                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2093                 last = entry->bytenr;
2094         }
2095         n = root->rb_node;
2096
2097         while (n) {
2098                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2099                 WARN_ON(!entry->in_tree);
2100
2101                 middle = entry->bytenr;
2102
2103                 if (alt)
2104                         n = n->rb_left;
2105                 else
2106                         n = n->rb_right;
2107
2108                 alt = 1 - alt;
2109         }
2110         return middle;
2111 }
2112 #endif
2113
2114 /*
2115  * this starts processing the delayed reference count updates and
2116  * extent insertions we have queued up so far.  count can be
2117  * 0, which means to process everything in the tree at the start
2118  * of the run (but not newly added entries), or it can be some target
2119  * number you'd like to process.
2120  *
2121  * Returns 0 on success or if called with an aborted transaction
2122  * Returns <0 on error and aborts the transaction
2123  */
2124 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2125                            unsigned long count)
2126 {
2127         struct btrfs_fs_info *fs_info = trans->fs_info;
2128         struct rb_node *node;
2129         struct btrfs_delayed_ref_root *delayed_refs;
2130         struct btrfs_delayed_ref_head *head;
2131         int ret;
2132         int run_all = count == (unsigned long)-1;
2133
2134         /* We'll clean this up in btrfs_cleanup_transaction */
2135         if (TRANS_ABORTED(trans))
2136                 return 0;
2137
2138         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2139                 return 0;
2140
2141         delayed_refs = &trans->transaction->delayed_refs;
2142         if (count == 0)
2143                 count = delayed_refs->num_heads_ready;
2144
2145 again:
2146 #ifdef SCRAMBLE_DELAYED_REFS
2147         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2148 #endif
2149         ret = __btrfs_run_delayed_refs(trans, count);
2150         if (ret < 0) {
2151                 btrfs_abort_transaction(trans, ret);
2152                 return ret;
2153         }
2154
2155         if (run_all) {
2156                 btrfs_create_pending_block_groups(trans);
2157
2158                 spin_lock(&delayed_refs->lock);
2159                 node = rb_first_cached(&delayed_refs->href_root);
2160                 if (!node) {
2161                         spin_unlock(&delayed_refs->lock);
2162                         goto out;
2163                 }
2164                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2165                                 href_node);
2166                 refcount_inc(&head->refs);
2167                 spin_unlock(&delayed_refs->lock);
2168
2169                 /* Mutex was contended, block until it's released and retry. */
2170                 mutex_lock(&head->mutex);
2171                 mutex_unlock(&head->mutex);
2172
2173                 btrfs_put_delayed_ref_head(head);
2174                 cond_resched();
2175                 goto again;
2176         }
2177 out:
2178         return 0;
2179 }
2180
2181 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2182                                 struct extent_buffer *eb, u64 flags,
2183                                 int level)
2184 {
2185         struct btrfs_delayed_extent_op *extent_op;
2186         int ret;
2187
2188         extent_op = btrfs_alloc_delayed_extent_op();
2189         if (!extent_op)
2190                 return -ENOMEM;
2191
2192         extent_op->flags_to_set = flags;
2193         extent_op->update_flags = true;
2194         extent_op->update_key = false;
2195         extent_op->level = level;
2196
2197         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2198         if (ret)
2199                 btrfs_free_delayed_extent_op(extent_op);
2200         return ret;
2201 }
2202
2203 static noinline int check_delayed_ref(struct btrfs_root *root,
2204                                       struct btrfs_path *path,
2205                                       u64 objectid, u64 offset, u64 bytenr)
2206 {
2207         struct btrfs_delayed_ref_head *head;
2208         struct btrfs_delayed_ref_node *ref;
2209         struct btrfs_delayed_data_ref *data_ref;
2210         struct btrfs_delayed_ref_root *delayed_refs;
2211         struct btrfs_transaction *cur_trans;
2212         struct rb_node *node;
2213         int ret = 0;
2214
2215         spin_lock(&root->fs_info->trans_lock);
2216         cur_trans = root->fs_info->running_transaction;
2217         if (cur_trans)
2218                 refcount_inc(&cur_trans->use_count);
2219         spin_unlock(&root->fs_info->trans_lock);
2220         if (!cur_trans)
2221                 return 0;
2222
2223         delayed_refs = &cur_trans->delayed_refs;
2224         spin_lock(&delayed_refs->lock);
2225         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2226         if (!head) {
2227                 spin_unlock(&delayed_refs->lock);
2228                 btrfs_put_transaction(cur_trans);
2229                 return 0;
2230         }
2231
2232         if (!mutex_trylock(&head->mutex)) {
2233                 if (path->nowait) {
2234                         spin_unlock(&delayed_refs->lock);
2235                         btrfs_put_transaction(cur_trans);
2236                         return -EAGAIN;
2237                 }
2238
2239                 refcount_inc(&head->refs);
2240                 spin_unlock(&delayed_refs->lock);
2241
2242                 btrfs_release_path(path);
2243
2244                 /*
2245                  * Mutex was contended, block until it's released and let
2246                  * caller try again
2247                  */
2248                 mutex_lock(&head->mutex);
2249                 mutex_unlock(&head->mutex);
2250                 btrfs_put_delayed_ref_head(head);
2251                 btrfs_put_transaction(cur_trans);
2252                 return -EAGAIN;
2253         }
2254         spin_unlock(&delayed_refs->lock);
2255
2256         spin_lock(&head->lock);
2257         /*
2258          * XXX: We should replace this with a proper search function in the
2259          * future.
2260          */
2261         for (node = rb_first_cached(&head->ref_tree); node;
2262              node = rb_next(node)) {
2263                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2264                 /* If it's a shared ref we know a cross reference exists */
2265                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2266                         ret = 1;
2267                         break;
2268                 }
2269
2270                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2271
2272                 /*
2273                  * If our ref doesn't match the one we're currently looking at
2274                  * then we have a cross reference.
2275                  */
2276                 if (data_ref->root != root->root_key.objectid ||
2277                     data_ref->objectid != objectid ||
2278                     data_ref->offset != offset) {
2279                         ret = 1;
2280                         break;
2281                 }
2282         }
2283         spin_unlock(&head->lock);
2284         mutex_unlock(&head->mutex);
2285         btrfs_put_transaction(cur_trans);
2286         return ret;
2287 }
2288
2289 static noinline int check_committed_ref(struct btrfs_root *root,
2290                                         struct btrfs_path *path,
2291                                         u64 objectid, u64 offset, u64 bytenr,
2292                                         bool strict)
2293 {
2294         struct btrfs_fs_info *fs_info = root->fs_info;
2295         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2296         struct extent_buffer *leaf;
2297         struct btrfs_extent_data_ref *ref;
2298         struct btrfs_extent_inline_ref *iref;
2299         struct btrfs_extent_item *ei;
2300         struct btrfs_key key;
2301         u32 item_size;
2302         int type;
2303         int ret;
2304
2305         key.objectid = bytenr;
2306         key.offset = (u64)-1;
2307         key.type = BTRFS_EXTENT_ITEM_KEY;
2308
2309         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2310         if (ret < 0)
2311                 goto out;
2312         BUG_ON(ret == 0); /* Corruption */
2313
2314         ret = -ENOENT;
2315         if (path->slots[0] == 0)
2316                 goto out;
2317
2318         path->slots[0]--;
2319         leaf = path->nodes[0];
2320         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2321
2322         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2323                 goto out;
2324
2325         ret = 1;
2326         item_size = btrfs_item_size(leaf, path->slots[0]);
2327         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2328
2329         /* If extent item has more than 1 inline ref then it's shared */
2330         if (item_size != sizeof(*ei) +
2331             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2332                 goto out;
2333
2334         /*
2335          * If extent created before last snapshot => it's shared unless the
2336          * snapshot has been deleted. Use the heuristic if strict is false.
2337          */
2338         if (!strict &&
2339             (btrfs_extent_generation(leaf, ei) <=
2340              btrfs_root_last_snapshot(&root->root_item)))
2341                 goto out;
2342
2343         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2344
2345         /* If this extent has SHARED_DATA_REF then it's shared */
2346         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2347         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2348                 goto out;
2349
2350         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2351         if (btrfs_extent_refs(leaf, ei) !=
2352             btrfs_extent_data_ref_count(leaf, ref) ||
2353             btrfs_extent_data_ref_root(leaf, ref) !=
2354             root->root_key.objectid ||
2355             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2356             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2357                 goto out;
2358
2359         ret = 0;
2360 out:
2361         return ret;
2362 }
2363
2364 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2365                           u64 bytenr, bool strict, struct btrfs_path *path)
2366 {
2367         int ret;
2368
2369         do {
2370                 ret = check_committed_ref(root, path, objectid,
2371                                           offset, bytenr, strict);
2372                 if (ret && ret != -ENOENT)
2373                         goto out;
2374
2375                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2376         } while (ret == -EAGAIN);
2377
2378 out:
2379         btrfs_release_path(path);
2380         if (btrfs_is_data_reloc_root(root))
2381                 WARN_ON(ret > 0);
2382         return ret;
2383 }
2384
2385 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2386                            struct btrfs_root *root,
2387                            struct extent_buffer *buf,
2388                            int full_backref, int inc)
2389 {
2390         struct btrfs_fs_info *fs_info = root->fs_info;
2391         u64 bytenr;
2392         u64 num_bytes;
2393         u64 parent;
2394         u64 ref_root;
2395         u32 nritems;
2396         struct btrfs_key key;
2397         struct btrfs_file_extent_item *fi;
2398         struct btrfs_ref generic_ref = { 0 };
2399         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2400         int i;
2401         int action;
2402         int level;
2403         int ret = 0;
2404
2405         if (btrfs_is_testing(fs_info))
2406                 return 0;
2407
2408         ref_root = btrfs_header_owner(buf);
2409         nritems = btrfs_header_nritems(buf);
2410         level = btrfs_header_level(buf);
2411
2412         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2413                 return 0;
2414
2415         if (full_backref)
2416                 parent = buf->start;
2417         else
2418                 parent = 0;
2419         if (inc)
2420                 action = BTRFS_ADD_DELAYED_REF;
2421         else
2422                 action = BTRFS_DROP_DELAYED_REF;
2423
2424         for (i = 0; i < nritems; i++) {
2425                 if (level == 0) {
2426                         btrfs_item_key_to_cpu(buf, &key, i);
2427                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2428                                 continue;
2429                         fi = btrfs_item_ptr(buf, i,
2430                                             struct btrfs_file_extent_item);
2431                         if (btrfs_file_extent_type(buf, fi) ==
2432                             BTRFS_FILE_EXTENT_INLINE)
2433                                 continue;
2434                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2435                         if (bytenr == 0)
2436                                 continue;
2437
2438                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2439                         key.offset -= btrfs_file_extent_offset(buf, fi);
2440                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2441                                                num_bytes, parent);
2442                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2443                                             key.offset, root->root_key.objectid,
2444                                             for_reloc);
2445                         if (inc)
2446                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2447                         else
2448                                 ret = btrfs_free_extent(trans, &generic_ref);
2449                         if (ret)
2450                                 goto fail;
2451                 } else {
2452                         bytenr = btrfs_node_blockptr(buf, i);
2453                         num_bytes = fs_info->nodesize;
2454                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2455                                                num_bytes, parent);
2456                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root,
2457                                             root->root_key.objectid, for_reloc);
2458                         if (inc)
2459                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2460                         else
2461                                 ret = btrfs_free_extent(trans, &generic_ref);
2462                         if (ret)
2463                                 goto fail;
2464                 }
2465         }
2466         return 0;
2467 fail:
2468         return ret;
2469 }
2470
2471 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2472                   struct extent_buffer *buf, int full_backref)
2473 {
2474         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2475 }
2476
2477 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2478                   struct extent_buffer *buf, int full_backref)
2479 {
2480         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2481 }
2482
2483 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2484 {
2485         struct btrfs_fs_info *fs_info = root->fs_info;
2486         u64 flags;
2487         u64 ret;
2488
2489         if (data)
2490                 flags = BTRFS_BLOCK_GROUP_DATA;
2491         else if (root == fs_info->chunk_root)
2492                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2493         else
2494                 flags = BTRFS_BLOCK_GROUP_METADATA;
2495
2496         ret = btrfs_get_alloc_profile(fs_info, flags);
2497         return ret;
2498 }
2499
2500 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2501 {
2502         struct rb_node *leftmost;
2503         u64 bytenr = 0;
2504
2505         read_lock(&fs_info->block_group_cache_lock);
2506         /* Get the block group with the lowest logical start address. */
2507         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2508         if (leftmost) {
2509                 struct btrfs_block_group *bg;
2510
2511                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2512                 bytenr = bg->start;
2513         }
2514         read_unlock(&fs_info->block_group_cache_lock);
2515
2516         return bytenr;
2517 }
2518
2519 static int pin_down_extent(struct btrfs_trans_handle *trans,
2520                            struct btrfs_block_group *cache,
2521                            u64 bytenr, u64 num_bytes, int reserved)
2522 {
2523         struct btrfs_fs_info *fs_info = cache->fs_info;
2524
2525         spin_lock(&cache->space_info->lock);
2526         spin_lock(&cache->lock);
2527         cache->pinned += num_bytes;
2528         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2529                                              num_bytes);
2530         if (reserved) {
2531                 cache->reserved -= num_bytes;
2532                 cache->space_info->bytes_reserved -= num_bytes;
2533         }
2534         spin_unlock(&cache->lock);
2535         spin_unlock(&cache->space_info->lock);
2536
2537         set_extent_dirty(&trans->transaction->pinned_extents, bytenr,
2538                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
2539         return 0;
2540 }
2541
2542 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2543                      u64 bytenr, u64 num_bytes, int reserved)
2544 {
2545         struct btrfs_block_group *cache;
2546
2547         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2548         BUG_ON(!cache); /* Logic error */
2549
2550         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2551
2552         btrfs_put_block_group(cache);
2553         return 0;
2554 }
2555
2556 /*
2557  * this function must be called within transaction
2558  */
2559 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2560                                     u64 bytenr, u64 num_bytes)
2561 {
2562         struct btrfs_block_group *cache;
2563         int ret;
2564
2565         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2566         if (!cache)
2567                 return -EINVAL;
2568
2569         /*
2570          * Fully cache the free space first so that our pin removes the free space
2571          * from the cache.
2572          */
2573         ret = btrfs_cache_block_group(cache, true);
2574         if (ret)
2575                 goto out;
2576
2577         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2578
2579         /* remove us from the free space cache (if we're there at all) */
2580         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2581 out:
2582         btrfs_put_block_group(cache);
2583         return ret;
2584 }
2585
2586 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2587                                    u64 start, u64 num_bytes)
2588 {
2589         int ret;
2590         struct btrfs_block_group *block_group;
2591
2592         block_group = btrfs_lookup_block_group(fs_info, start);
2593         if (!block_group)
2594                 return -EINVAL;
2595
2596         ret = btrfs_cache_block_group(block_group, true);
2597         if (ret)
2598                 goto out;
2599
2600         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2601 out:
2602         btrfs_put_block_group(block_group);
2603         return ret;
2604 }
2605
2606 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2607 {
2608         struct btrfs_fs_info *fs_info = eb->fs_info;
2609         struct btrfs_file_extent_item *item;
2610         struct btrfs_key key;
2611         int found_type;
2612         int i;
2613         int ret = 0;
2614
2615         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2616                 return 0;
2617
2618         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2619                 btrfs_item_key_to_cpu(eb, &key, i);
2620                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2621                         continue;
2622                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2623                 found_type = btrfs_file_extent_type(eb, item);
2624                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2625                         continue;
2626                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2627                         continue;
2628                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2629                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2630                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2631                 if (ret)
2632                         break;
2633         }
2634
2635         return ret;
2636 }
2637
2638 static void
2639 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2640 {
2641         atomic_inc(&bg->reservations);
2642 }
2643
2644 /*
2645  * Returns the free cluster for the given space info and sets empty_cluster to
2646  * what it should be based on the mount options.
2647  */
2648 static struct btrfs_free_cluster *
2649 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2650                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2651 {
2652         struct btrfs_free_cluster *ret = NULL;
2653
2654         *empty_cluster = 0;
2655         if (btrfs_mixed_space_info(space_info))
2656                 return ret;
2657
2658         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2659                 ret = &fs_info->meta_alloc_cluster;
2660                 if (btrfs_test_opt(fs_info, SSD))
2661                         *empty_cluster = SZ_2M;
2662                 else
2663                         *empty_cluster = SZ_64K;
2664         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2665                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2666                 *empty_cluster = SZ_2M;
2667                 ret = &fs_info->data_alloc_cluster;
2668         }
2669
2670         return ret;
2671 }
2672
2673 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2674                               u64 start, u64 end,
2675                               const bool return_free_space)
2676 {
2677         struct btrfs_block_group *cache = NULL;
2678         struct btrfs_space_info *space_info;
2679         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2680         struct btrfs_free_cluster *cluster = NULL;
2681         u64 len;
2682         u64 total_unpinned = 0;
2683         u64 empty_cluster = 0;
2684         bool readonly;
2685
2686         while (start <= end) {
2687                 readonly = false;
2688                 if (!cache ||
2689                     start >= cache->start + cache->length) {
2690                         if (cache)
2691                                 btrfs_put_block_group(cache);
2692                         total_unpinned = 0;
2693                         cache = btrfs_lookup_block_group(fs_info, start);
2694                         BUG_ON(!cache); /* Logic error */
2695
2696                         cluster = fetch_cluster_info(fs_info,
2697                                                      cache->space_info,
2698                                                      &empty_cluster);
2699                         empty_cluster <<= 1;
2700                 }
2701
2702                 len = cache->start + cache->length - start;
2703                 len = min(len, end + 1 - start);
2704
2705                 if (return_free_space)
2706                         btrfs_add_free_space(cache, start, len);
2707
2708                 start += len;
2709                 total_unpinned += len;
2710                 space_info = cache->space_info;
2711
2712                 /*
2713                  * If this space cluster has been marked as fragmented and we've
2714                  * unpinned enough in this block group to potentially allow a
2715                  * cluster to be created inside of it go ahead and clear the
2716                  * fragmented check.
2717                  */
2718                 if (cluster && cluster->fragmented &&
2719                     total_unpinned > empty_cluster) {
2720                         spin_lock(&cluster->lock);
2721                         cluster->fragmented = 0;
2722                         spin_unlock(&cluster->lock);
2723                 }
2724
2725                 spin_lock(&space_info->lock);
2726                 spin_lock(&cache->lock);
2727                 cache->pinned -= len;
2728                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2729                 space_info->max_extent_size = 0;
2730                 if (cache->ro) {
2731                         space_info->bytes_readonly += len;
2732                         readonly = true;
2733                 } else if (btrfs_is_zoned(fs_info)) {
2734                         /* Need reset before reusing in a zoned block group */
2735                         space_info->bytes_zone_unusable += len;
2736                         readonly = true;
2737                 }
2738                 spin_unlock(&cache->lock);
2739                 if (!readonly && return_free_space &&
2740                     global_rsv->space_info == space_info) {
2741                         spin_lock(&global_rsv->lock);
2742                         if (!global_rsv->full) {
2743                                 u64 to_add = min(len, global_rsv->size -
2744                                                       global_rsv->reserved);
2745
2746                                 global_rsv->reserved += to_add;
2747                                 btrfs_space_info_update_bytes_may_use(fs_info,
2748                                                 space_info, to_add);
2749                                 if (global_rsv->reserved >= global_rsv->size)
2750                                         global_rsv->full = 1;
2751                                 len -= to_add;
2752                         }
2753                         spin_unlock(&global_rsv->lock);
2754                 }
2755                 /* Add to any tickets we may have */
2756                 if (!readonly && return_free_space && len)
2757                         btrfs_try_granting_tickets(fs_info, space_info);
2758                 spin_unlock(&space_info->lock);
2759         }
2760
2761         if (cache)
2762                 btrfs_put_block_group(cache);
2763         return 0;
2764 }
2765
2766 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2767 {
2768         struct btrfs_fs_info *fs_info = trans->fs_info;
2769         struct btrfs_block_group *block_group, *tmp;
2770         struct list_head *deleted_bgs;
2771         struct extent_io_tree *unpin;
2772         u64 start;
2773         u64 end;
2774         int ret;
2775
2776         unpin = &trans->transaction->pinned_extents;
2777
2778         while (!TRANS_ABORTED(trans)) {
2779                 struct extent_state *cached_state = NULL;
2780
2781                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2782                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2783                                             EXTENT_DIRTY, &cached_state);
2784                 if (ret) {
2785                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2786                         break;
2787                 }
2788
2789                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2790                         ret = btrfs_discard_extent(fs_info, start,
2791                                                    end + 1 - start, NULL);
2792
2793                 clear_extent_dirty(unpin, start, end, &cached_state);
2794                 unpin_extent_range(fs_info, start, end, true);
2795                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2796                 free_extent_state(cached_state);
2797                 cond_resched();
2798         }
2799
2800         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2801                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2802                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2803         }
2804
2805         /*
2806          * Transaction is finished.  We don't need the lock anymore.  We
2807          * do need to clean up the block groups in case of a transaction
2808          * abort.
2809          */
2810         deleted_bgs = &trans->transaction->deleted_bgs;
2811         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2812                 u64 trimmed = 0;
2813
2814                 ret = -EROFS;
2815                 if (!TRANS_ABORTED(trans))
2816                         ret = btrfs_discard_extent(fs_info,
2817                                                    block_group->start,
2818                                                    block_group->length,
2819                                                    &trimmed);
2820
2821                 list_del_init(&block_group->bg_list);
2822                 btrfs_unfreeze_block_group(block_group);
2823                 btrfs_put_block_group(block_group);
2824
2825                 if (ret) {
2826                         const char *errstr = btrfs_decode_error(ret);
2827                         btrfs_warn(fs_info,
2828                            "discard failed while removing blockgroup: errno=%d %s",
2829                                    ret, errstr);
2830                 }
2831         }
2832
2833         return 0;
2834 }
2835
2836 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2837                                      u64 bytenr, u64 num_bytes, bool is_data)
2838 {
2839         int ret;
2840
2841         if (is_data) {
2842                 struct btrfs_root *csum_root;
2843
2844                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
2845                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
2846                 if (ret) {
2847                         btrfs_abort_transaction(trans, ret);
2848                         return ret;
2849                 }
2850         }
2851
2852         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
2853         if (ret) {
2854                 btrfs_abort_transaction(trans, ret);
2855                 return ret;
2856         }
2857
2858         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
2859         if (ret)
2860                 btrfs_abort_transaction(trans, ret);
2861
2862         return ret;
2863 }
2864
2865 /*
2866  * Drop one or more refs of @node.
2867  *
2868  * 1. Locate the extent refs.
2869  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2870  *    Locate it, then reduce the refs number or remove the ref line completely.
2871  *
2872  * 2. Update the refs count in EXTENT/METADATA_ITEM
2873  *
2874  * Inline backref case:
2875  *
2876  * in extent tree we have:
2877  *
2878  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2879  *              refs 2 gen 6 flags DATA
2880  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2881  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
2882  *
2883  * This function gets called with:
2884  *
2885  *    node->bytenr = 13631488
2886  *    node->num_bytes = 1048576
2887  *    root_objectid = FS_TREE
2888  *    owner_objectid = 257
2889  *    owner_offset = 0
2890  *    refs_to_drop = 1
2891  *
2892  * Then we should get some like:
2893  *
2894  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2895  *              refs 1 gen 6 flags DATA
2896  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2897  *
2898  * Keyed backref case:
2899  *
2900  * in extent tree we have:
2901  *
2902  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2903  *              refs 754 gen 6 flags DATA
2904  *      [...]
2905  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
2906  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
2907  *
2908  * This function get called with:
2909  *
2910  *    node->bytenr = 13631488
2911  *    node->num_bytes = 1048576
2912  *    root_objectid = FS_TREE
2913  *    owner_objectid = 866
2914  *    owner_offset = 0
2915  *    refs_to_drop = 1
2916  *
2917  * Then we should get some like:
2918  *
2919  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2920  *              refs 753 gen 6 flags DATA
2921  *
2922  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
2923  */
2924 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2925                                struct btrfs_delayed_ref_node *node, u64 parent,
2926                                u64 root_objectid, u64 owner_objectid,
2927                                u64 owner_offset, int refs_to_drop,
2928                                struct btrfs_delayed_extent_op *extent_op)
2929 {
2930         struct btrfs_fs_info *info = trans->fs_info;
2931         struct btrfs_key key;
2932         struct btrfs_path *path;
2933         struct btrfs_root *extent_root;
2934         struct extent_buffer *leaf;
2935         struct btrfs_extent_item *ei;
2936         struct btrfs_extent_inline_ref *iref;
2937         int ret;
2938         int is_data;
2939         int extent_slot = 0;
2940         int found_extent = 0;
2941         int num_to_del = 1;
2942         u32 item_size;
2943         u64 refs;
2944         u64 bytenr = node->bytenr;
2945         u64 num_bytes = node->num_bytes;
2946         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2947
2948         extent_root = btrfs_extent_root(info, bytenr);
2949         ASSERT(extent_root);
2950
2951         path = btrfs_alloc_path();
2952         if (!path)
2953                 return -ENOMEM;
2954
2955         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2956
2957         if (!is_data && refs_to_drop != 1) {
2958                 btrfs_crit(info,
2959 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
2960                            node->bytenr, refs_to_drop);
2961                 ret = -EINVAL;
2962                 btrfs_abort_transaction(trans, ret);
2963                 goto out;
2964         }
2965
2966         if (is_data)
2967                 skinny_metadata = false;
2968
2969         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2970                                     parent, root_objectid, owner_objectid,
2971                                     owner_offset);
2972         if (ret == 0) {
2973                 /*
2974                  * Either the inline backref or the SHARED_DATA_REF/
2975                  * SHARED_BLOCK_REF is found
2976                  *
2977                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
2978                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
2979                  */
2980                 extent_slot = path->slots[0];
2981                 while (extent_slot >= 0) {
2982                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2983                                               extent_slot);
2984                         if (key.objectid != bytenr)
2985                                 break;
2986                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
2987                             key.offset == num_bytes) {
2988                                 found_extent = 1;
2989                                 break;
2990                         }
2991                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
2992                             key.offset == owner_objectid) {
2993                                 found_extent = 1;
2994                                 break;
2995                         }
2996
2997                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
2998                         if (path->slots[0] - extent_slot > 5)
2999                                 break;
3000                         extent_slot--;
3001                 }
3002
3003                 if (!found_extent) {
3004                         if (iref) {
3005                                 btrfs_crit(info,
3006 "invalid iref, no EXTENT/METADATA_ITEM found but has inline extent ref");
3007                                 btrfs_abort_transaction(trans, -EUCLEAN);
3008                                 goto err_dump;
3009                         }
3010                         /* Must be SHARED_* item, remove the backref first */
3011                         ret = remove_extent_backref(trans, extent_root, path,
3012                                                     NULL, refs_to_drop, is_data);
3013                         if (ret) {
3014                                 btrfs_abort_transaction(trans, ret);
3015                                 goto out;
3016                         }
3017                         btrfs_release_path(path);
3018
3019                         /* Slow path to locate EXTENT/METADATA_ITEM */
3020                         key.objectid = bytenr;
3021                         key.type = BTRFS_EXTENT_ITEM_KEY;
3022                         key.offset = num_bytes;
3023
3024                         if (!is_data && skinny_metadata) {
3025                                 key.type = BTRFS_METADATA_ITEM_KEY;
3026                                 key.offset = owner_objectid;
3027                         }
3028
3029                         ret = btrfs_search_slot(trans, extent_root,
3030                                                 &key, path, -1, 1);
3031                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3032                                 /*
3033                                  * Couldn't find our skinny metadata item,
3034                                  * see if we have ye olde extent item.
3035                                  */
3036                                 path->slots[0]--;
3037                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3038                                                       path->slots[0]);
3039                                 if (key.objectid == bytenr &&
3040                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3041                                     key.offset == num_bytes)
3042                                         ret = 0;
3043                         }
3044
3045                         if (ret > 0 && skinny_metadata) {
3046                                 skinny_metadata = false;
3047                                 key.objectid = bytenr;
3048                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3049                                 key.offset = num_bytes;
3050                                 btrfs_release_path(path);
3051                                 ret = btrfs_search_slot(trans, extent_root,
3052                                                         &key, path, -1, 1);
3053                         }
3054
3055                         if (ret) {
3056                                 btrfs_err(info,
3057                                           "umm, got %d back from search, was looking for %llu",
3058                                           ret, bytenr);
3059                                 if (ret > 0)
3060                                         btrfs_print_leaf(path->nodes[0]);
3061                         }
3062                         if (ret < 0) {
3063                                 btrfs_abort_transaction(trans, ret);
3064                                 goto out;
3065                         }
3066                         extent_slot = path->slots[0];
3067                 }
3068         } else if (WARN_ON(ret == -ENOENT)) {
3069                 btrfs_print_leaf(path->nodes[0]);
3070                 btrfs_err(info,
3071                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
3072                         bytenr, parent, root_objectid, owner_objectid,
3073                         owner_offset);
3074                 btrfs_abort_transaction(trans, ret);
3075                 goto out;
3076         } else {
3077                 btrfs_abort_transaction(trans, ret);
3078                 goto out;
3079         }
3080
3081         leaf = path->nodes[0];
3082         item_size = btrfs_item_size(leaf, extent_slot);
3083         if (unlikely(item_size < sizeof(*ei))) {
3084                 ret = -EINVAL;
3085                 btrfs_print_v0_err(info);
3086                 btrfs_abort_transaction(trans, ret);
3087                 goto out;
3088         }
3089         ei = btrfs_item_ptr(leaf, extent_slot,
3090                             struct btrfs_extent_item);
3091         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3092             key.type == BTRFS_EXTENT_ITEM_KEY) {
3093                 struct btrfs_tree_block_info *bi;
3094                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3095                         btrfs_crit(info,
3096 "invalid extent item size for key (%llu, %u, %llu) owner %llu, has %u expect >= %zu",
3097                                    key.objectid, key.type, key.offset,
3098                                    owner_objectid, item_size,
3099                                    sizeof(*ei) + sizeof(*bi));
3100                         btrfs_abort_transaction(trans, -EUCLEAN);
3101                         goto err_dump;
3102                 }
3103                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3104                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3105         }
3106
3107         refs = btrfs_extent_refs(leaf, ei);
3108         if (refs < refs_to_drop) {
3109                 btrfs_crit(info,
3110                 "trying to drop %d refs but we only have %llu for bytenr %llu",
3111                           refs_to_drop, refs, bytenr);
3112                 btrfs_abort_transaction(trans, -EUCLEAN);
3113                 goto err_dump;
3114         }
3115         refs -= refs_to_drop;
3116
3117         if (refs > 0) {
3118                 if (extent_op)
3119                         __run_delayed_extent_op(extent_op, leaf, ei);
3120                 /*
3121                  * In the case of inline back ref, reference count will
3122                  * be updated by remove_extent_backref
3123                  */
3124                 if (iref) {
3125                         if (!found_extent) {
3126                                 btrfs_crit(info,
3127 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found");
3128                                 btrfs_abort_transaction(trans, -EUCLEAN);
3129                                 goto err_dump;
3130                         }
3131                 } else {
3132                         btrfs_set_extent_refs(leaf, ei, refs);
3133                         btrfs_mark_buffer_dirty(leaf);
3134                 }
3135                 if (found_extent) {
3136                         ret = remove_extent_backref(trans, extent_root, path,
3137                                                     iref, refs_to_drop, is_data);
3138                         if (ret) {
3139                                 btrfs_abort_transaction(trans, ret);
3140                                 goto out;
3141                         }
3142                 }
3143         } else {
3144                 /* In this branch refs == 1 */
3145                 if (found_extent) {
3146                         if (is_data && refs_to_drop !=
3147                             extent_data_ref_count(path, iref)) {
3148                                 btrfs_crit(info,
3149                 "invalid refs_to_drop, current refs %u refs_to_drop %u",
3150                                            extent_data_ref_count(path, iref),
3151                                            refs_to_drop);
3152                                 btrfs_abort_transaction(trans, -EUCLEAN);
3153                                 goto err_dump;
3154                         }
3155                         if (iref) {
3156                                 if (path->slots[0] != extent_slot) {
3157                                         btrfs_crit(info,
3158 "invalid iref, extent item key (%llu %u %llu) doesn't have wanted iref",
3159                                                    key.objectid, key.type,
3160                                                    key.offset);
3161                                         btrfs_abort_transaction(trans, -EUCLEAN);
3162                                         goto err_dump;
3163                                 }
3164                         } else {
3165                                 /*
3166                                  * No inline ref, we must be at SHARED_* item,
3167                                  * And it's single ref, it must be:
3168                                  * |    extent_slot       ||extent_slot + 1|
3169                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3170                                  */
3171                                 if (path->slots[0] != extent_slot + 1) {
3172                                         btrfs_crit(info,
3173         "invalid SHARED_* item, previous item is not EXTENT/METADATA_ITEM");
3174                                         btrfs_abort_transaction(trans, -EUCLEAN);
3175                                         goto err_dump;
3176                                 }
3177                                 path->slots[0] = extent_slot;
3178                                 num_to_del = 2;
3179                         }
3180                 }
3181
3182                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3183                                       num_to_del);
3184                 if (ret) {
3185                         btrfs_abort_transaction(trans, ret);
3186                         goto out;
3187                 }
3188                 btrfs_release_path(path);
3189
3190                 ret = do_free_extent_accounting(trans, bytenr, num_bytes, is_data);
3191         }
3192         btrfs_release_path(path);
3193
3194 out:
3195         btrfs_free_path(path);
3196         return ret;
3197 err_dump:
3198         /*
3199          * Leaf dump can take up a lot of log buffer, so we only do full leaf
3200          * dump for debug build.
3201          */
3202         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
3203                 btrfs_crit(info, "path->slots[0]=%d extent_slot=%d",
3204                            path->slots[0], extent_slot);
3205                 btrfs_print_leaf(path->nodes[0]);
3206         }
3207
3208         btrfs_free_path(path);
3209         return -EUCLEAN;
3210 }
3211
3212 /*
3213  * when we free an block, it is possible (and likely) that we free the last
3214  * delayed ref for that extent as well.  This searches the delayed ref tree for
3215  * a given extent, and if there are no other delayed refs to be processed, it
3216  * removes it from the tree.
3217  */
3218 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3219                                       u64 bytenr)
3220 {
3221         struct btrfs_delayed_ref_head *head;
3222         struct btrfs_delayed_ref_root *delayed_refs;
3223         int ret = 0;
3224
3225         delayed_refs = &trans->transaction->delayed_refs;
3226         spin_lock(&delayed_refs->lock);
3227         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3228         if (!head)
3229                 goto out_delayed_unlock;
3230
3231         spin_lock(&head->lock);
3232         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3233                 goto out;
3234
3235         if (cleanup_extent_op(head) != NULL)
3236                 goto out;
3237
3238         /*
3239          * waiting for the lock here would deadlock.  If someone else has it
3240          * locked they are already in the process of dropping it anyway
3241          */
3242         if (!mutex_trylock(&head->mutex))
3243                 goto out;
3244
3245         btrfs_delete_ref_head(delayed_refs, head);
3246         head->processing = 0;
3247
3248         spin_unlock(&head->lock);
3249         spin_unlock(&delayed_refs->lock);
3250
3251         BUG_ON(head->extent_op);
3252         if (head->must_insert_reserved)
3253                 ret = 1;
3254
3255         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3256         mutex_unlock(&head->mutex);
3257         btrfs_put_delayed_ref_head(head);
3258         return ret;
3259 out:
3260         spin_unlock(&head->lock);
3261
3262 out_delayed_unlock:
3263         spin_unlock(&delayed_refs->lock);
3264         return 0;
3265 }
3266
3267 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3268                            u64 root_id,
3269                            struct extent_buffer *buf,
3270                            u64 parent, int last_ref)
3271 {
3272         struct btrfs_fs_info *fs_info = trans->fs_info;
3273         struct btrfs_ref generic_ref = { 0 };
3274         int ret;
3275
3276         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3277                                buf->start, buf->len, parent);
3278         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3279                             root_id, 0, false);
3280
3281         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3282                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3283                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3284                 BUG_ON(ret); /* -ENOMEM */
3285         }
3286
3287         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3288                 struct btrfs_block_group *cache;
3289                 bool must_pin = false;
3290
3291                 if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3292                         ret = check_ref_cleanup(trans, buf->start);
3293                         if (!ret) {
3294                                 btrfs_redirty_list_add(trans->transaction, buf);
3295                                 goto out;
3296                         }
3297                 }
3298
3299                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3300
3301                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3302                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3303                         btrfs_put_block_group(cache);
3304                         goto out;
3305                 }
3306
3307                 /*
3308                  * If there are tree mod log users we may have recorded mod log
3309                  * operations for this node.  If we re-allocate this node we
3310                  * could replay operations on this node that happened when it
3311                  * existed in a completely different root.  For example if it
3312                  * was part of root A, then was reallocated to root B, and we
3313                  * are doing a btrfs_old_search_slot(root b), we could replay
3314                  * operations that happened when the block was part of root A,
3315                  * giving us an inconsistent view of the btree.
3316                  *
3317                  * We are safe from races here because at this point no other
3318                  * node or root points to this extent buffer, so if after this
3319                  * check a new tree mod log user joins we will not have an
3320                  * existing log of operations on this node that we have to
3321                  * contend with.
3322                  */
3323                 if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
3324                         must_pin = true;
3325
3326                 if (must_pin || btrfs_is_zoned(fs_info)) {
3327                         btrfs_redirty_list_add(trans->transaction, buf);
3328                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3329                         btrfs_put_block_group(cache);
3330                         goto out;
3331                 }
3332
3333                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3334
3335                 btrfs_add_free_space(cache, buf->start, buf->len);
3336                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3337                 btrfs_put_block_group(cache);
3338                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3339         }
3340 out:
3341         if (last_ref) {
3342                 /*
3343                  * Deleting the buffer, clear the corrupt flag since it doesn't
3344                  * matter anymore.
3345                  */
3346                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3347         }
3348 }
3349
3350 /* Can return -ENOMEM */
3351 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3352 {
3353         struct btrfs_fs_info *fs_info = trans->fs_info;
3354         int ret;
3355
3356         if (btrfs_is_testing(fs_info))
3357                 return 0;
3358
3359         /*
3360          * tree log blocks never actually go into the extent allocation
3361          * tree, just update pinning info and exit early.
3362          */
3363         if ((ref->type == BTRFS_REF_METADATA &&
3364              ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3365             (ref->type == BTRFS_REF_DATA &&
3366              ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)) {
3367                 /* unlocks the pinned mutex */
3368                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3369                 ret = 0;
3370         } else if (ref->type == BTRFS_REF_METADATA) {
3371                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3372         } else {
3373                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3374         }
3375
3376         if (!((ref->type == BTRFS_REF_METADATA &&
3377                ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3378               (ref->type == BTRFS_REF_DATA &&
3379                ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)))
3380                 btrfs_ref_tree_mod(fs_info, ref);
3381
3382         return ret;
3383 }
3384
3385 enum btrfs_loop_type {
3386         LOOP_CACHING_NOWAIT,
3387         LOOP_CACHING_WAIT,
3388         LOOP_ALLOC_CHUNK,
3389         LOOP_NO_EMPTY_SIZE,
3390 };
3391
3392 static inline void
3393 btrfs_lock_block_group(struct btrfs_block_group *cache,
3394                        int delalloc)
3395 {
3396         if (delalloc)
3397                 down_read(&cache->data_rwsem);
3398 }
3399
3400 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3401                        int delalloc)
3402 {
3403         btrfs_get_block_group(cache);
3404         if (delalloc)
3405                 down_read(&cache->data_rwsem);
3406 }
3407
3408 static struct btrfs_block_group *btrfs_lock_cluster(
3409                    struct btrfs_block_group *block_group,
3410                    struct btrfs_free_cluster *cluster,
3411                    int delalloc)
3412         __acquires(&cluster->refill_lock)
3413 {
3414         struct btrfs_block_group *used_bg = NULL;
3415
3416         spin_lock(&cluster->refill_lock);
3417         while (1) {
3418                 used_bg = cluster->block_group;
3419                 if (!used_bg)
3420                         return NULL;
3421
3422                 if (used_bg == block_group)
3423                         return used_bg;
3424
3425                 btrfs_get_block_group(used_bg);
3426
3427                 if (!delalloc)
3428                         return used_bg;
3429
3430                 if (down_read_trylock(&used_bg->data_rwsem))
3431                         return used_bg;
3432
3433                 spin_unlock(&cluster->refill_lock);
3434
3435                 /* We should only have one-level nested. */
3436                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3437
3438                 spin_lock(&cluster->refill_lock);
3439                 if (used_bg == cluster->block_group)
3440                         return used_bg;
3441
3442                 up_read(&used_bg->data_rwsem);
3443                 btrfs_put_block_group(used_bg);
3444         }
3445 }
3446
3447 static inline void
3448 btrfs_release_block_group(struct btrfs_block_group *cache,
3449                          int delalloc)
3450 {
3451         if (delalloc)
3452                 up_read(&cache->data_rwsem);
3453         btrfs_put_block_group(cache);
3454 }
3455
3456 enum btrfs_extent_allocation_policy {
3457         BTRFS_EXTENT_ALLOC_CLUSTERED,
3458         BTRFS_EXTENT_ALLOC_ZONED,
3459 };
3460
3461 /*
3462  * Structure used internally for find_free_extent() function.  Wraps needed
3463  * parameters.
3464  */
3465 struct find_free_extent_ctl {
3466         /* Basic allocation info */
3467         u64 ram_bytes;
3468         u64 num_bytes;
3469         u64 min_alloc_size;
3470         u64 empty_size;
3471         u64 flags;
3472         int delalloc;
3473
3474         /* Where to start the search inside the bg */
3475         u64 search_start;
3476
3477         /* For clustered allocation */
3478         u64 empty_cluster;
3479         struct btrfs_free_cluster *last_ptr;
3480         bool use_cluster;
3481
3482         bool have_caching_bg;
3483         bool orig_have_caching_bg;
3484
3485         /* Allocation is called for tree-log */
3486         bool for_treelog;
3487
3488         /* Allocation is called for data relocation */
3489         bool for_data_reloc;
3490
3491         /* RAID index, converted from flags */
3492         int index;
3493
3494         /*
3495          * Current loop number, check find_free_extent_update_loop() for details
3496          */
3497         int loop;
3498
3499         /*
3500          * Whether we're refilling a cluster, if true we need to re-search
3501          * current block group but don't try to refill the cluster again.
3502          */
3503         bool retry_clustered;
3504
3505         /*
3506          * Whether we're updating free space cache, if true we need to re-search
3507          * current block group but don't try updating free space cache again.
3508          */
3509         bool retry_unclustered;
3510
3511         /* If current block group is cached */
3512         int cached;
3513
3514         /* Max contiguous hole found */
3515         u64 max_extent_size;
3516
3517         /* Total free space from free space cache, not always contiguous */
3518         u64 total_free_space;
3519
3520         /* Found result */
3521         u64 found_offset;
3522
3523         /* Hint where to start looking for an empty space */
3524         u64 hint_byte;
3525
3526         /* Allocation policy */
3527         enum btrfs_extent_allocation_policy policy;
3528 };
3529
3530
3531 /*
3532  * Helper function for find_free_extent().
3533  *
3534  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3535  * Return -EAGAIN to inform caller that we need to re-search this block group
3536  * Return >0 to inform caller that we find nothing
3537  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3538  */
3539 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3540                                       struct find_free_extent_ctl *ffe_ctl,
3541                                       struct btrfs_block_group **cluster_bg_ret)
3542 {
3543         struct btrfs_block_group *cluster_bg;
3544         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3545         u64 aligned_cluster;
3546         u64 offset;
3547         int ret;
3548
3549         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3550         if (!cluster_bg)
3551                 goto refill_cluster;
3552         if (cluster_bg != bg && (cluster_bg->ro ||
3553             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3554                 goto release_cluster;
3555
3556         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3557                         ffe_ctl->num_bytes, cluster_bg->start,
3558                         &ffe_ctl->max_extent_size);
3559         if (offset) {
3560                 /* We have a block, we're done */
3561                 spin_unlock(&last_ptr->refill_lock);
3562                 trace_btrfs_reserve_extent_cluster(cluster_bg,
3563                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
3564                 *cluster_bg_ret = cluster_bg;
3565                 ffe_ctl->found_offset = offset;
3566                 return 0;
3567         }
3568         WARN_ON(last_ptr->block_group != cluster_bg);
3569
3570 release_cluster:
3571         /*
3572          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3573          * lets just skip it and let the allocator find whatever block it can
3574          * find. If we reach this point, we will have tried the cluster
3575          * allocator plenty of times and not have found anything, so we are
3576          * likely way too fragmented for the clustering stuff to find anything.
3577          *
3578          * However, if the cluster is taken from the current block group,
3579          * release the cluster first, so that we stand a better chance of
3580          * succeeding in the unclustered allocation.
3581          */
3582         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3583                 spin_unlock(&last_ptr->refill_lock);
3584                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3585                 return -ENOENT;
3586         }
3587
3588         /* This cluster didn't work out, free it and start over */
3589         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3590
3591         if (cluster_bg != bg)
3592                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3593
3594 refill_cluster:
3595         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3596                 spin_unlock(&last_ptr->refill_lock);
3597                 return -ENOENT;
3598         }
3599
3600         aligned_cluster = max_t(u64,
3601                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3602                         bg->full_stripe_len);
3603         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3604                         ffe_ctl->num_bytes, aligned_cluster);
3605         if (ret == 0) {
3606                 /* Now pull our allocation out of this cluster */
3607                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3608                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3609                                 &ffe_ctl->max_extent_size);
3610                 if (offset) {
3611                         /* We found one, proceed */
3612                         spin_unlock(&last_ptr->refill_lock);
3613                         trace_btrfs_reserve_extent_cluster(bg,
3614                                         ffe_ctl->search_start,
3615                                         ffe_ctl->num_bytes);
3616                         ffe_ctl->found_offset = offset;
3617                         return 0;
3618                 }
3619         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
3620                    !ffe_ctl->retry_clustered) {
3621                 spin_unlock(&last_ptr->refill_lock);
3622
3623                 ffe_ctl->retry_clustered = true;
3624                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3625                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
3626                 return -EAGAIN;
3627         }
3628         /*
3629          * At this point we either didn't find a cluster or we weren't able to
3630          * allocate a block from our cluster.  Free the cluster we've been
3631          * trying to use, and go to the next block group.
3632          */
3633         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3634         spin_unlock(&last_ptr->refill_lock);
3635         return 1;
3636 }
3637
3638 /*
3639  * Return >0 to inform caller that we find nothing
3640  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3641  * Return -EAGAIN to inform caller that we need to re-search this block group
3642  */
3643 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3644                                         struct find_free_extent_ctl *ffe_ctl)
3645 {
3646         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3647         u64 offset;
3648
3649         /*
3650          * We are doing an unclustered allocation, set the fragmented flag so
3651          * we don't bother trying to setup a cluster again until we get more
3652          * space.
3653          */
3654         if (unlikely(last_ptr)) {
3655                 spin_lock(&last_ptr->lock);
3656                 last_ptr->fragmented = 1;
3657                 spin_unlock(&last_ptr->lock);
3658         }
3659         if (ffe_ctl->cached) {
3660                 struct btrfs_free_space_ctl *free_space_ctl;
3661
3662                 free_space_ctl = bg->free_space_ctl;
3663                 spin_lock(&free_space_ctl->tree_lock);
3664                 if (free_space_ctl->free_space <
3665                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3666                     ffe_ctl->empty_size) {
3667                         ffe_ctl->total_free_space = max_t(u64,
3668                                         ffe_ctl->total_free_space,
3669                                         free_space_ctl->free_space);
3670                         spin_unlock(&free_space_ctl->tree_lock);
3671                         return 1;
3672                 }
3673                 spin_unlock(&free_space_ctl->tree_lock);
3674         }
3675
3676         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3677                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3678                         &ffe_ctl->max_extent_size);
3679
3680         /*
3681          * If we didn't find a chunk, and we haven't failed on this block group
3682          * before, and this block group is in the middle of caching and we are
3683          * ok with waiting, then go ahead and wait for progress to be made, and
3684          * set @retry_unclustered to true.
3685          *
3686          * If @retry_unclustered is true then we've already waited on this
3687          * block group once and should move on to the next block group.
3688          */
3689         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
3690             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
3691                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3692                                                       ffe_ctl->empty_size);
3693                 ffe_ctl->retry_unclustered = true;
3694                 return -EAGAIN;
3695         } else if (!offset) {
3696                 return 1;
3697         }
3698         ffe_ctl->found_offset = offset;
3699         return 0;
3700 }
3701
3702 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3703                                    struct find_free_extent_ctl *ffe_ctl,
3704                                    struct btrfs_block_group **bg_ret)
3705 {
3706         int ret;
3707
3708         /* We want to try and use the cluster allocator, so lets look there */
3709         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3710                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3711                 if (ret >= 0 || ret == -EAGAIN)
3712                         return ret;
3713                 /* ret == -ENOENT case falls through */
3714         }
3715
3716         return find_free_extent_unclustered(block_group, ffe_ctl);
3717 }
3718
3719 /*
3720  * Tree-log block group locking
3721  * ============================
3722  *
3723  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3724  * indicates the starting address of a block group, which is reserved only
3725  * for tree-log metadata.
3726  *
3727  * Lock nesting
3728  * ============
3729  *
3730  * space_info::lock
3731  *   block_group::lock
3732  *     fs_info::treelog_bg_lock
3733  */
3734
3735 /*
3736  * Simple allocator for sequential-only block group. It only allows sequential
3737  * allocation. No need to play with trees. This function also reserves the
3738  * bytes as in btrfs_add_reserved_bytes.
3739  */
3740 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3741                                struct find_free_extent_ctl *ffe_ctl,
3742                                struct btrfs_block_group **bg_ret)
3743 {
3744         struct btrfs_fs_info *fs_info = block_group->fs_info;
3745         struct btrfs_space_info *space_info = block_group->space_info;
3746         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3747         u64 start = block_group->start;
3748         u64 num_bytes = ffe_ctl->num_bytes;
3749         u64 avail;
3750         u64 bytenr = block_group->start;
3751         u64 log_bytenr;
3752         u64 data_reloc_bytenr;
3753         int ret = 0;
3754         bool skip = false;
3755
3756         ASSERT(btrfs_is_zoned(block_group->fs_info));
3757
3758         /*
3759          * Do not allow non-tree-log blocks in the dedicated tree-log block
3760          * group, and vice versa.
3761          */
3762         spin_lock(&fs_info->treelog_bg_lock);
3763         log_bytenr = fs_info->treelog_bg;
3764         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3765                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3766                 skip = true;
3767         spin_unlock(&fs_info->treelog_bg_lock);
3768         if (skip)
3769                 return 1;
3770
3771         /*
3772          * Do not allow non-relocation blocks in the dedicated relocation block
3773          * group, and vice versa.
3774          */
3775         spin_lock(&fs_info->relocation_bg_lock);
3776         data_reloc_bytenr = fs_info->data_reloc_bg;
3777         if (data_reloc_bytenr &&
3778             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3779              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3780                 skip = true;
3781         spin_unlock(&fs_info->relocation_bg_lock);
3782         if (skip)
3783                 return 1;
3784
3785         /* Check RO and no space case before trying to activate it */
3786         spin_lock(&block_group->lock);
3787         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3788                 ret = 1;
3789                 /*
3790                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3791                  * Return the error after taking the locks.
3792                  */
3793         }
3794         spin_unlock(&block_group->lock);
3795
3796         if (!ret && !btrfs_zone_activate(block_group)) {
3797                 ret = 1;
3798                 /*
3799                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3800                  * Return the error after taking the locks.
3801                  */
3802         }
3803
3804         spin_lock(&space_info->lock);
3805         spin_lock(&block_group->lock);
3806         spin_lock(&fs_info->treelog_bg_lock);
3807         spin_lock(&fs_info->relocation_bg_lock);
3808
3809         if (ret)
3810                 goto out;
3811
3812         ASSERT(!ffe_ctl->for_treelog ||
3813                block_group->start == fs_info->treelog_bg ||
3814                fs_info->treelog_bg == 0);
3815         ASSERT(!ffe_ctl->for_data_reloc ||
3816                block_group->start == fs_info->data_reloc_bg ||
3817                fs_info->data_reloc_bg == 0);
3818
3819         if (block_group->ro ||
3820             test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
3821                 ret = 1;
3822                 goto out;
3823         }
3824
3825         /*
3826          * Do not allow currently using block group to be tree-log dedicated
3827          * block group.
3828          */
3829         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3830             (block_group->used || block_group->reserved)) {
3831                 ret = 1;
3832                 goto out;
3833         }
3834
3835         /*
3836          * Do not allow currently used block group to be the data relocation
3837          * dedicated block group.
3838          */
3839         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3840             (block_group->used || block_group->reserved)) {
3841                 ret = 1;
3842                 goto out;
3843         }
3844
3845         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3846         avail = block_group->zone_capacity - block_group->alloc_offset;
3847         if (avail < num_bytes) {
3848                 if (ffe_ctl->max_extent_size < avail) {
3849                         /*
3850                          * With sequential allocator, free space is always
3851                          * contiguous
3852                          */
3853                         ffe_ctl->max_extent_size = avail;
3854                         ffe_ctl->total_free_space = avail;
3855                 }
3856                 ret = 1;
3857                 goto out;
3858         }
3859
3860         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3861                 fs_info->treelog_bg = block_group->start;
3862
3863         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg)
3864                 fs_info->data_reloc_bg = block_group->start;
3865
3866         ffe_ctl->found_offset = start + block_group->alloc_offset;
3867         block_group->alloc_offset += num_bytes;
3868         spin_lock(&ctl->tree_lock);
3869         ctl->free_space -= num_bytes;
3870         spin_unlock(&ctl->tree_lock);
3871
3872         /*
3873          * We do not check if found_offset is aligned to stripesize. The
3874          * address is anyway rewritten when using zone append writing.
3875          */
3876
3877         ffe_ctl->search_start = ffe_ctl->found_offset;
3878
3879 out:
3880         if (ret && ffe_ctl->for_treelog)
3881                 fs_info->treelog_bg = 0;
3882         if (ret && ffe_ctl->for_data_reloc &&
3883             fs_info->data_reloc_bg == block_group->start) {
3884                 /*
3885                  * Do not allow further allocations from this block group.
3886                  * Compared to increasing the ->ro, setting the
3887                  * ->zoned_data_reloc_ongoing flag still allows nocow
3888                  *  writers to come in. See btrfs_inc_nocow_writers().
3889                  *
3890                  * We need to disable an allocation to avoid an allocation of
3891                  * regular (non-relocation data) extent. With mix of relocation
3892                  * extents and regular extents, we can dispatch WRITE commands
3893                  * (for relocation extents) and ZONE APPEND commands (for
3894                  * regular extents) at the same time to the same zone, which
3895                  * easily break the write pointer.
3896                  */
3897                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
3898                 fs_info->data_reloc_bg = 0;
3899         }
3900         spin_unlock(&fs_info->relocation_bg_lock);
3901         spin_unlock(&fs_info->treelog_bg_lock);
3902         spin_unlock(&block_group->lock);
3903         spin_unlock(&space_info->lock);
3904         return ret;
3905 }
3906
3907 static int do_allocation(struct btrfs_block_group *block_group,
3908                          struct find_free_extent_ctl *ffe_ctl,
3909                          struct btrfs_block_group **bg_ret)
3910 {
3911         switch (ffe_ctl->policy) {
3912         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3913                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3914         case BTRFS_EXTENT_ALLOC_ZONED:
3915                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3916         default:
3917                 BUG();
3918         }
3919 }
3920
3921 static void release_block_group(struct btrfs_block_group *block_group,
3922                                 struct find_free_extent_ctl *ffe_ctl,
3923                                 int delalloc)
3924 {
3925         switch (ffe_ctl->policy) {
3926         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3927                 ffe_ctl->retry_clustered = false;
3928                 ffe_ctl->retry_unclustered = false;
3929                 break;
3930         case BTRFS_EXTENT_ALLOC_ZONED:
3931                 /* Nothing to do */
3932                 break;
3933         default:
3934                 BUG();
3935         }
3936
3937         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3938                ffe_ctl->index);
3939         btrfs_release_block_group(block_group, delalloc);
3940 }
3941
3942 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3943                                    struct btrfs_key *ins)
3944 {
3945         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3946
3947         if (!ffe_ctl->use_cluster && last_ptr) {
3948                 spin_lock(&last_ptr->lock);
3949                 last_ptr->window_start = ins->objectid;
3950                 spin_unlock(&last_ptr->lock);
3951         }
3952 }
3953
3954 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3955                          struct btrfs_key *ins)
3956 {
3957         switch (ffe_ctl->policy) {
3958         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3959                 found_extent_clustered(ffe_ctl, ins);
3960                 break;
3961         case BTRFS_EXTENT_ALLOC_ZONED:
3962                 /* Nothing to do */
3963                 break;
3964         default:
3965                 BUG();
3966         }
3967 }
3968
3969 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
3970                                     struct find_free_extent_ctl *ffe_ctl)
3971 {
3972         /* If we can activate new zone, just allocate a chunk and use it */
3973         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
3974                 return 0;
3975
3976         /*
3977          * We already reached the max active zones. Try to finish one block
3978          * group to make a room for a new block group. This is only possible
3979          * for a data block group because btrfs_zone_finish() may need to wait
3980          * for a running transaction which can cause a deadlock for metadata
3981          * allocation.
3982          */
3983         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
3984                 int ret = btrfs_zone_finish_one_bg(fs_info);
3985
3986                 if (ret == 1)
3987                         return 0;
3988                 else if (ret < 0)
3989                         return ret;
3990         }
3991
3992         /*
3993          * If we have enough free space left in an already active block group
3994          * and we can't activate any other zone now, do not allow allocating a
3995          * new chunk and let find_free_extent() retry with a smaller size.
3996          */
3997         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
3998                 return -ENOSPC;
3999
4000         /*
4001          * Even min_alloc_size is not left in any block groups. Since we cannot
4002          * activate a new block group, allocating it may not help. Let's tell a
4003          * caller to try again and hope it progress something by writing some
4004          * parts of the region. That is only possible for data block groups,
4005          * where a part of the region can be written.
4006          */
4007         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
4008                 return -EAGAIN;
4009
4010         /*
4011          * We cannot activate a new block group and no enough space left in any
4012          * block groups. So, allocating a new block group may not help. But,
4013          * there is nothing to do anyway, so let's go with it.
4014          */
4015         return 0;
4016 }
4017
4018 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
4019                               struct find_free_extent_ctl *ffe_ctl)
4020 {
4021         switch (ffe_ctl->policy) {
4022         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4023                 return 0;
4024         case BTRFS_EXTENT_ALLOC_ZONED:
4025                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
4026         default:
4027                 BUG();
4028         }
4029 }
4030
4031 static int chunk_allocation_failed(struct find_free_extent_ctl *ffe_ctl)
4032 {
4033         switch (ffe_ctl->policy) {
4034         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4035                 /*
4036                  * If we can't allocate a new chunk we've already looped through
4037                  * at least once, move on to the NO_EMPTY_SIZE case.
4038                  */
4039                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
4040                 return 0;
4041         case BTRFS_EXTENT_ALLOC_ZONED:
4042                 /* Give up here */
4043                 return -ENOSPC;
4044         default:
4045                 BUG();
4046         }
4047 }
4048
4049 /*
4050  * Return >0 means caller needs to re-search for free extent
4051  * Return 0 means we have the needed free extent.
4052  * Return <0 means we failed to locate any free extent.
4053  */
4054 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
4055                                         struct btrfs_key *ins,
4056                                         struct find_free_extent_ctl *ffe_ctl,
4057                                         bool full_search)
4058 {
4059         struct btrfs_root *root = fs_info->chunk_root;
4060         int ret;
4061
4062         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
4063             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
4064                 ffe_ctl->orig_have_caching_bg = true;
4065
4066         if (ins->objectid) {
4067                 found_extent(ffe_ctl, ins);
4068                 return 0;
4069         }
4070
4071         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
4072                 return 1;
4073
4074         ffe_ctl->index++;
4075         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
4076                 return 1;
4077
4078         /*
4079          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
4080          *                      caching kthreads as we move along
4081          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
4082          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
4083          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
4084          *                     again
4085          */
4086         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4087                 ffe_ctl->index = 0;
4088                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
4089                         /*
4090                          * We want to skip the LOOP_CACHING_WAIT step if we
4091                          * don't have any uncached bgs and we've already done a
4092                          * full search through.
4093                          */
4094                         if (ffe_ctl->orig_have_caching_bg || !full_search)
4095                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
4096                         else
4097                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
4098                 } else {
4099                         ffe_ctl->loop++;
4100                 }
4101
4102                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4103                         struct btrfs_trans_handle *trans;
4104                         int exist = 0;
4105
4106                         /*Check if allocation policy allows to create a new chunk */
4107                         ret = can_allocate_chunk(fs_info, ffe_ctl);
4108                         if (ret)
4109                                 return ret;
4110
4111                         trans = current->journal_info;
4112                         if (trans)
4113                                 exist = 1;
4114                         else
4115                                 trans = btrfs_join_transaction(root);
4116
4117                         if (IS_ERR(trans)) {
4118                                 ret = PTR_ERR(trans);
4119                                 return ret;
4120                         }
4121
4122                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4123                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4124
4125                         /* Do not bail out on ENOSPC since we can do more. */
4126                         if (ret == -ENOSPC)
4127                                 ret = chunk_allocation_failed(ffe_ctl);
4128                         else if (ret < 0)
4129                                 btrfs_abort_transaction(trans, ret);
4130                         else
4131                                 ret = 0;
4132                         if (!exist)
4133                                 btrfs_end_transaction(trans);
4134                         if (ret)
4135                                 return ret;
4136                 }
4137
4138                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4139                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4140                                 return -ENOSPC;
4141
4142                         /*
4143                          * Don't loop again if we already have no empty_size and
4144                          * no empty_cluster.
4145                          */
4146                         if (ffe_ctl->empty_size == 0 &&
4147                             ffe_ctl->empty_cluster == 0)
4148                                 return -ENOSPC;
4149                         ffe_ctl->empty_size = 0;
4150                         ffe_ctl->empty_cluster = 0;
4151                 }
4152                 return 1;
4153         }
4154         return -ENOSPC;
4155 }
4156
4157 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4158                                         struct find_free_extent_ctl *ffe_ctl,
4159                                         struct btrfs_space_info *space_info,
4160                                         struct btrfs_key *ins)
4161 {
4162         /*
4163          * If our free space is heavily fragmented we may not be able to make
4164          * big contiguous allocations, so instead of doing the expensive search
4165          * for free space, simply return ENOSPC with our max_extent_size so we
4166          * can go ahead and search for a more manageable chunk.
4167          *
4168          * If our max_extent_size is large enough for our allocation simply
4169          * disable clustering since we will likely not be able to find enough
4170          * space to create a cluster and induce latency trying.
4171          */
4172         if (space_info->max_extent_size) {
4173                 spin_lock(&space_info->lock);
4174                 if (space_info->max_extent_size &&
4175                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4176                         ins->offset = space_info->max_extent_size;
4177                         spin_unlock(&space_info->lock);
4178                         return -ENOSPC;
4179                 } else if (space_info->max_extent_size) {
4180                         ffe_ctl->use_cluster = false;
4181                 }
4182                 spin_unlock(&space_info->lock);
4183         }
4184
4185         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4186                                                &ffe_ctl->empty_cluster);
4187         if (ffe_ctl->last_ptr) {
4188                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4189
4190                 spin_lock(&last_ptr->lock);
4191                 if (last_ptr->block_group)
4192                         ffe_ctl->hint_byte = last_ptr->window_start;
4193                 if (last_ptr->fragmented) {
4194                         /*
4195                          * We still set window_start so we can keep track of the
4196                          * last place we found an allocation to try and save
4197                          * some time.
4198                          */
4199                         ffe_ctl->hint_byte = last_ptr->window_start;
4200                         ffe_ctl->use_cluster = false;
4201                 }
4202                 spin_unlock(&last_ptr->lock);
4203         }
4204
4205         return 0;
4206 }
4207
4208 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4209                               struct find_free_extent_ctl *ffe_ctl,
4210                               struct btrfs_space_info *space_info,
4211                               struct btrfs_key *ins)
4212 {
4213         switch (ffe_ctl->policy) {
4214         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4215                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4216                                                     space_info, ins);
4217         case BTRFS_EXTENT_ALLOC_ZONED:
4218                 if (ffe_ctl->for_treelog) {
4219                         spin_lock(&fs_info->treelog_bg_lock);
4220                         if (fs_info->treelog_bg)
4221                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4222                         spin_unlock(&fs_info->treelog_bg_lock);
4223                 }
4224                 if (ffe_ctl->for_data_reloc) {
4225                         spin_lock(&fs_info->relocation_bg_lock);
4226                         if (fs_info->data_reloc_bg)
4227                                 ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4228                         spin_unlock(&fs_info->relocation_bg_lock);
4229                 }
4230                 return 0;
4231         default:
4232                 BUG();
4233         }
4234 }
4235
4236 /*
4237  * walks the btree of allocated extents and find a hole of a given size.
4238  * The key ins is changed to record the hole:
4239  * ins->objectid == start position
4240  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4241  * ins->offset == the size of the hole.
4242  * Any available blocks before search_start are skipped.
4243  *
4244  * If there is no suitable free space, we will record the max size of
4245  * the free space extent currently.
4246  *
4247  * The overall logic and call chain:
4248  *
4249  * find_free_extent()
4250  * |- Iterate through all block groups
4251  * |  |- Get a valid block group
4252  * |  |- Try to do clustered allocation in that block group
4253  * |  |- Try to do unclustered allocation in that block group
4254  * |  |- Check if the result is valid
4255  * |  |  |- If valid, then exit
4256  * |  |- Jump to next block group
4257  * |
4258  * |- Push harder to find free extents
4259  *    |- If not found, re-iterate all block groups
4260  */
4261 static noinline int find_free_extent(struct btrfs_root *root,
4262                                      struct btrfs_key *ins,
4263                                      struct find_free_extent_ctl *ffe_ctl)
4264 {
4265         struct btrfs_fs_info *fs_info = root->fs_info;
4266         int ret = 0;
4267         int cache_block_group_error = 0;
4268         struct btrfs_block_group *block_group = NULL;
4269         struct btrfs_space_info *space_info;
4270         bool full_search = false;
4271
4272         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4273
4274         ffe_ctl->search_start = 0;
4275         /* For clustered allocation */
4276         ffe_ctl->empty_cluster = 0;
4277         ffe_ctl->last_ptr = NULL;
4278         ffe_ctl->use_cluster = true;
4279         ffe_ctl->have_caching_bg = false;
4280         ffe_ctl->orig_have_caching_bg = false;
4281         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4282         ffe_ctl->loop = 0;
4283         /* For clustered allocation */
4284         ffe_ctl->retry_clustered = false;
4285         ffe_ctl->retry_unclustered = false;
4286         ffe_ctl->cached = 0;
4287         ffe_ctl->max_extent_size = 0;
4288         ffe_ctl->total_free_space = 0;
4289         ffe_ctl->found_offset = 0;
4290         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4291
4292         if (btrfs_is_zoned(fs_info))
4293                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4294
4295         ins->type = BTRFS_EXTENT_ITEM_KEY;
4296         ins->objectid = 0;
4297         ins->offset = 0;
4298
4299         trace_find_free_extent(root, ffe_ctl->num_bytes, ffe_ctl->empty_size,
4300                                ffe_ctl->flags);
4301
4302         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4303         if (!space_info) {
4304                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4305                 return -ENOSPC;
4306         }
4307
4308         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4309         if (ret < 0)
4310                 return ret;
4311
4312         ffe_ctl->search_start = max(ffe_ctl->search_start,
4313                                     first_logical_byte(fs_info));
4314         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4315         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4316                 block_group = btrfs_lookup_block_group(fs_info,
4317                                                        ffe_ctl->search_start);
4318                 /*
4319                  * we don't want to use the block group if it doesn't match our
4320                  * allocation bits, or if its not cached.
4321                  *
4322                  * However if we are re-searching with an ideal block group
4323                  * picked out then we don't care that the block group is cached.
4324                  */
4325                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4326                     block_group->cached != BTRFS_CACHE_NO) {
4327                         down_read(&space_info->groups_sem);
4328                         if (list_empty(&block_group->list) ||
4329                             block_group->ro) {
4330                                 /*
4331                                  * someone is removing this block group,
4332                                  * we can't jump into the have_block_group
4333                                  * target because our list pointers are not
4334                                  * valid
4335                                  */
4336                                 btrfs_put_block_group(block_group);
4337                                 up_read(&space_info->groups_sem);
4338                         } else {
4339                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4340                                                         block_group->flags);
4341                                 btrfs_lock_block_group(block_group,
4342                                                        ffe_ctl->delalloc);
4343                                 goto have_block_group;
4344                         }
4345                 } else if (block_group) {
4346                         btrfs_put_block_group(block_group);
4347                 }
4348         }
4349 search:
4350         ffe_ctl->have_caching_bg = false;
4351         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4352             ffe_ctl->index == 0)
4353                 full_search = true;
4354         down_read(&space_info->groups_sem);
4355         list_for_each_entry(block_group,
4356                             &space_info->block_groups[ffe_ctl->index], list) {
4357                 struct btrfs_block_group *bg_ret;
4358
4359                 /* If the block group is read-only, we can skip it entirely. */
4360                 if (unlikely(block_group->ro)) {
4361                         if (ffe_ctl->for_treelog)
4362                                 btrfs_clear_treelog_bg(block_group);
4363                         if (ffe_ctl->for_data_reloc)
4364                                 btrfs_clear_data_reloc_bg(block_group);
4365                         continue;
4366                 }
4367
4368                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4369                 ffe_ctl->search_start = block_group->start;
4370
4371                 /*
4372                  * this can happen if we end up cycling through all the
4373                  * raid types, but we want to make sure we only allocate
4374                  * for the proper type.
4375                  */
4376                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4377                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4378                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4379                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4380                                 BTRFS_BLOCK_GROUP_RAID10;
4381
4382                         /*
4383                          * if they asked for extra copies and this block group
4384                          * doesn't provide them, bail.  This does allow us to
4385                          * fill raid0 from raid1.
4386                          */
4387                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4388                                 goto loop;
4389
4390                         /*
4391                          * This block group has different flags than we want.
4392                          * It's possible that we have MIXED_GROUP flag but no
4393                          * block group is mixed.  Just skip such block group.
4394                          */
4395                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4396                         continue;
4397                 }
4398
4399 have_block_group:
4400                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4401                 if (unlikely(!ffe_ctl->cached)) {
4402                         ffe_ctl->have_caching_bg = true;
4403                         ret = btrfs_cache_block_group(block_group, false);
4404
4405                         /*
4406                          * If we get ENOMEM here or something else we want to
4407                          * try other block groups, because it may not be fatal.
4408                          * However if we can't find anything else we need to
4409                          * save our return here so that we return the actual
4410                          * error that caused problems, not ENOSPC.
4411                          */
4412                         if (ret < 0) {
4413                                 if (!cache_block_group_error)
4414                                         cache_block_group_error = ret;
4415                                 ret = 0;
4416                                 goto loop;
4417                         }
4418                         ret = 0;
4419                 }
4420
4421                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
4422                         goto loop;
4423
4424                 bg_ret = NULL;
4425                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4426                 if (ret == 0) {
4427                         if (bg_ret && bg_ret != block_group) {
4428                                 btrfs_release_block_group(block_group,
4429                                                           ffe_ctl->delalloc);
4430                                 block_group = bg_ret;
4431                         }
4432                 } else if (ret == -EAGAIN) {
4433                         goto have_block_group;
4434                 } else if (ret > 0) {
4435                         goto loop;
4436                 }
4437
4438                 /* Checks */
4439                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4440                                                  fs_info->stripesize);
4441
4442                 /* move on to the next group */
4443                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4444                     block_group->start + block_group->length) {
4445                         btrfs_add_free_space_unused(block_group,
4446                                             ffe_ctl->found_offset,
4447                                             ffe_ctl->num_bytes);
4448                         goto loop;
4449                 }
4450
4451                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4452                         btrfs_add_free_space_unused(block_group,
4453                                         ffe_ctl->found_offset,
4454                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4455
4456                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4457                                                ffe_ctl->num_bytes,
4458                                                ffe_ctl->delalloc);
4459                 if (ret == -EAGAIN) {
4460                         btrfs_add_free_space_unused(block_group,
4461                                         ffe_ctl->found_offset,
4462                                         ffe_ctl->num_bytes);
4463                         goto loop;
4464                 }
4465                 btrfs_inc_block_group_reservations(block_group);
4466
4467                 /* we are all good, lets return */
4468                 ins->objectid = ffe_ctl->search_start;
4469                 ins->offset = ffe_ctl->num_bytes;
4470
4471                 trace_btrfs_reserve_extent(block_group, ffe_ctl->search_start,
4472                                            ffe_ctl->num_bytes);
4473                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4474                 break;
4475 loop:
4476                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4477                 cond_resched();
4478         }
4479         up_read(&space_info->groups_sem);
4480
4481         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4482         if (ret > 0)
4483                 goto search;
4484
4485         if (ret == -ENOSPC && !cache_block_group_error) {
4486                 /*
4487                  * Use ffe_ctl->total_free_space as fallback if we can't find
4488                  * any contiguous hole.
4489                  */
4490                 if (!ffe_ctl->max_extent_size)
4491                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4492                 spin_lock(&space_info->lock);
4493                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4494                 spin_unlock(&space_info->lock);
4495                 ins->offset = ffe_ctl->max_extent_size;
4496         } else if (ret == -ENOSPC) {
4497                 ret = cache_block_group_error;
4498         }
4499         return ret;
4500 }
4501
4502 /*
4503  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4504  *                        hole that is at least as big as @num_bytes.
4505  *
4506  * @root           -    The root that will contain this extent
4507  *
4508  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4509  *                      is used for accounting purposes. This value differs
4510  *                      from @num_bytes only in the case of compressed extents.
4511  *
4512  * @num_bytes      -    Number of bytes to allocate on-disk.
4513  *
4514  * @min_alloc_size -    Indicates the minimum amount of space that the
4515  *                      allocator should try to satisfy. In some cases
4516  *                      @num_bytes may be larger than what is required and if
4517  *                      the filesystem is fragmented then allocation fails.
4518  *                      However, the presence of @min_alloc_size gives a
4519  *                      chance to try and satisfy the smaller allocation.
4520  *
4521  * @empty_size     -    A hint that you plan on doing more COW. This is the
4522  *                      size in bytes the allocator should try to find free
4523  *                      next to the block it returns.  This is just a hint and
4524  *                      may be ignored by the allocator.
4525  *
4526  * @hint_byte      -    Hint to the allocator to start searching above the byte
4527  *                      address passed. It might be ignored.
4528  *
4529  * @ins            -    This key is modified to record the found hole. It will
4530  *                      have the following values:
4531  *                      ins->objectid == start position
4532  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4533  *                      ins->offset == the size of the hole.
4534  *
4535  * @is_data        -    Boolean flag indicating whether an extent is
4536  *                      allocated for data (true) or metadata (false)
4537  *
4538  * @delalloc       -    Boolean flag indicating whether this allocation is for
4539  *                      delalloc or not. If 'true' data_rwsem of block groups
4540  *                      is going to be acquired.
4541  *
4542  *
4543  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4544  * case -ENOSPC is returned then @ins->offset will contain the size of the
4545  * largest available hole the allocator managed to find.
4546  */
4547 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4548                          u64 num_bytes, u64 min_alloc_size,
4549                          u64 empty_size, u64 hint_byte,
4550                          struct btrfs_key *ins, int is_data, int delalloc)
4551 {
4552         struct btrfs_fs_info *fs_info = root->fs_info;
4553         struct find_free_extent_ctl ffe_ctl = {};
4554         bool final_tried = num_bytes == min_alloc_size;
4555         u64 flags;
4556         int ret;
4557         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4558         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4559
4560         flags = get_alloc_profile_by_root(root, is_data);
4561 again:
4562         WARN_ON(num_bytes < fs_info->sectorsize);
4563
4564         ffe_ctl.ram_bytes = ram_bytes;
4565         ffe_ctl.num_bytes = num_bytes;
4566         ffe_ctl.min_alloc_size = min_alloc_size;
4567         ffe_ctl.empty_size = empty_size;
4568         ffe_ctl.flags = flags;
4569         ffe_ctl.delalloc = delalloc;
4570         ffe_ctl.hint_byte = hint_byte;
4571         ffe_ctl.for_treelog = for_treelog;
4572         ffe_ctl.for_data_reloc = for_data_reloc;
4573
4574         ret = find_free_extent(root, ins, &ffe_ctl);
4575         if (!ret && !is_data) {
4576                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4577         } else if (ret == -ENOSPC) {
4578                 if (!final_tried && ins->offset) {
4579                         num_bytes = min(num_bytes >> 1, ins->offset);
4580                         num_bytes = round_down(num_bytes,
4581                                                fs_info->sectorsize);
4582                         num_bytes = max(num_bytes, min_alloc_size);
4583                         ram_bytes = num_bytes;
4584                         if (num_bytes == min_alloc_size)
4585                                 final_tried = true;
4586                         goto again;
4587                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4588                         struct btrfs_space_info *sinfo;
4589
4590                         sinfo = btrfs_find_space_info(fs_info, flags);
4591                         btrfs_err(fs_info,
4592         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4593                                   flags, num_bytes, for_treelog, for_data_reloc);
4594                         if (sinfo)
4595                                 btrfs_dump_space_info(fs_info, sinfo,
4596                                                       num_bytes, 1);
4597                 }
4598         }
4599
4600         return ret;
4601 }
4602
4603 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4604                                u64 start, u64 len, int delalloc)
4605 {
4606         struct btrfs_block_group *cache;
4607
4608         cache = btrfs_lookup_block_group(fs_info, start);
4609         if (!cache) {
4610                 btrfs_err(fs_info, "Unable to find block group for %llu",
4611                           start);
4612                 return -ENOSPC;
4613         }
4614
4615         btrfs_add_free_space(cache, start, len);
4616         btrfs_free_reserved_bytes(cache, len, delalloc);
4617         trace_btrfs_reserved_extent_free(fs_info, start, len);
4618
4619         btrfs_put_block_group(cache);
4620         return 0;
4621 }
4622
4623 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4624                               u64 len)
4625 {
4626         struct btrfs_block_group *cache;
4627         int ret = 0;
4628
4629         cache = btrfs_lookup_block_group(trans->fs_info, start);
4630         if (!cache) {
4631                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4632                           start);
4633                 return -ENOSPC;
4634         }
4635
4636         ret = pin_down_extent(trans, cache, start, len, 1);
4637         btrfs_put_block_group(cache);
4638         return ret;
4639 }
4640
4641 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4642                                  u64 num_bytes)
4643 {
4644         struct btrfs_fs_info *fs_info = trans->fs_info;
4645         int ret;
4646
4647         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4648         if (ret)
4649                 return ret;
4650
4651         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4652         if (ret) {
4653                 ASSERT(!ret);
4654                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4655                           bytenr, num_bytes);
4656                 return ret;
4657         }
4658
4659         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4660         return 0;
4661 }
4662
4663 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4664                                       u64 parent, u64 root_objectid,
4665                                       u64 flags, u64 owner, u64 offset,
4666                                       struct btrfs_key *ins, int ref_mod)
4667 {
4668         struct btrfs_fs_info *fs_info = trans->fs_info;
4669         struct btrfs_root *extent_root;
4670         int ret;
4671         struct btrfs_extent_item *extent_item;
4672         struct btrfs_extent_inline_ref *iref;
4673         struct btrfs_path *path;
4674         struct extent_buffer *leaf;
4675         int type;
4676         u32 size;
4677
4678         if (parent > 0)
4679                 type = BTRFS_SHARED_DATA_REF_KEY;
4680         else
4681                 type = BTRFS_EXTENT_DATA_REF_KEY;
4682
4683         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4684
4685         path = btrfs_alloc_path();
4686         if (!path)
4687                 return -ENOMEM;
4688
4689         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4690         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4691         if (ret) {
4692                 btrfs_free_path(path);
4693                 return ret;
4694         }
4695
4696         leaf = path->nodes[0];
4697         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4698                                      struct btrfs_extent_item);
4699         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4700         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4701         btrfs_set_extent_flags(leaf, extent_item,
4702                                flags | BTRFS_EXTENT_FLAG_DATA);
4703
4704         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4705         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4706         if (parent > 0) {
4707                 struct btrfs_shared_data_ref *ref;
4708                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4709                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4710                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4711         } else {
4712                 struct btrfs_extent_data_ref *ref;
4713                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4714                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4715                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4716                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4717                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4718         }
4719
4720         btrfs_mark_buffer_dirty(path->nodes[0]);
4721         btrfs_free_path(path);
4722
4723         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4724 }
4725
4726 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4727                                      struct btrfs_delayed_ref_node *node,
4728                                      struct btrfs_delayed_extent_op *extent_op)
4729 {
4730         struct btrfs_fs_info *fs_info = trans->fs_info;
4731         struct btrfs_root *extent_root;
4732         int ret;
4733         struct btrfs_extent_item *extent_item;
4734         struct btrfs_key extent_key;
4735         struct btrfs_tree_block_info *block_info;
4736         struct btrfs_extent_inline_ref *iref;
4737         struct btrfs_path *path;
4738         struct extent_buffer *leaf;
4739         struct btrfs_delayed_tree_ref *ref;
4740         u32 size = sizeof(*extent_item) + sizeof(*iref);
4741         u64 flags = extent_op->flags_to_set;
4742         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4743
4744         ref = btrfs_delayed_node_to_tree_ref(node);
4745
4746         extent_key.objectid = node->bytenr;
4747         if (skinny_metadata) {
4748                 extent_key.offset = ref->level;
4749                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4750         } else {
4751                 extent_key.offset = node->num_bytes;
4752                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4753                 size += sizeof(*block_info);
4754         }
4755
4756         path = btrfs_alloc_path();
4757         if (!path)
4758                 return -ENOMEM;
4759
4760         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4761         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4762                                       size);
4763         if (ret) {
4764                 btrfs_free_path(path);
4765                 return ret;
4766         }
4767
4768         leaf = path->nodes[0];
4769         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4770                                      struct btrfs_extent_item);
4771         btrfs_set_extent_refs(leaf, extent_item, 1);
4772         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4773         btrfs_set_extent_flags(leaf, extent_item,
4774                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4775
4776         if (skinny_metadata) {
4777                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4778         } else {
4779                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4780                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4781                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4782                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4783         }
4784
4785         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4786                 btrfs_set_extent_inline_ref_type(leaf, iref,
4787                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4788                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4789         } else {
4790                 btrfs_set_extent_inline_ref_type(leaf, iref,
4791                                                  BTRFS_TREE_BLOCK_REF_KEY);
4792                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4793         }
4794
4795         btrfs_mark_buffer_dirty(leaf);
4796         btrfs_free_path(path);
4797
4798         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4799 }
4800
4801 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4802                                      struct btrfs_root *root, u64 owner,
4803                                      u64 offset, u64 ram_bytes,
4804                                      struct btrfs_key *ins)
4805 {
4806         struct btrfs_ref generic_ref = { 0 };
4807
4808         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4809
4810         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4811                                ins->objectid, ins->offset, 0);
4812         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner,
4813                             offset, 0, false);
4814         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4815
4816         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4817 }
4818
4819 /*
4820  * this is used by the tree logging recovery code.  It records that
4821  * an extent has been allocated and makes sure to clear the free
4822  * space cache bits as well
4823  */
4824 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4825                                    u64 root_objectid, u64 owner, u64 offset,
4826                                    struct btrfs_key *ins)
4827 {
4828         struct btrfs_fs_info *fs_info = trans->fs_info;
4829         int ret;
4830         struct btrfs_block_group *block_group;
4831         struct btrfs_space_info *space_info;
4832
4833         /*
4834          * Mixed block groups will exclude before processing the log so we only
4835          * need to do the exclude dance if this fs isn't mixed.
4836          */
4837         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4838                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4839                                               ins->offset);
4840                 if (ret)
4841                         return ret;
4842         }
4843
4844         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4845         if (!block_group)
4846                 return -EINVAL;
4847
4848         space_info = block_group->space_info;
4849         spin_lock(&space_info->lock);
4850         spin_lock(&block_group->lock);
4851         space_info->bytes_reserved += ins->offset;
4852         block_group->reserved += ins->offset;
4853         spin_unlock(&block_group->lock);
4854         spin_unlock(&space_info->lock);
4855
4856         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4857                                          offset, ins, 1);
4858         if (ret)
4859                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4860         btrfs_put_block_group(block_group);
4861         return ret;
4862 }
4863
4864 static struct extent_buffer *
4865 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4866                       u64 bytenr, int level, u64 owner,
4867                       enum btrfs_lock_nesting nest)
4868 {
4869         struct btrfs_fs_info *fs_info = root->fs_info;
4870         struct extent_buffer *buf;
4871         u64 lockdep_owner = owner;
4872
4873         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
4874         if (IS_ERR(buf))
4875                 return buf;
4876
4877         /*
4878          * Extra safety check in case the extent tree is corrupted and extent
4879          * allocator chooses to use a tree block which is already used and
4880          * locked.
4881          */
4882         if (buf->lock_owner == current->pid) {
4883                 btrfs_err_rl(fs_info,
4884 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4885                         buf->start, btrfs_header_owner(buf), current->pid);
4886                 free_extent_buffer(buf);
4887                 return ERR_PTR(-EUCLEAN);
4888         }
4889
4890         /*
4891          * The reloc trees are just snapshots, so we need them to appear to be
4892          * just like any other fs tree WRT lockdep.
4893          *
4894          * The exception however is in replace_path() in relocation, where we
4895          * hold the lock on the original fs root and then search for the reloc
4896          * root.  At that point we need to make sure any reloc root buffers are
4897          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
4898          * lockdep happy.
4899          */
4900         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
4901             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
4902                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
4903
4904         /* btrfs_clean_tree_block() accesses generation field. */
4905         btrfs_set_header_generation(buf, trans->transid);
4906
4907         /*
4908          * This needs to stay, because we could allocate a freed block from an
4909          * old tree into a new tree, so we need to make sure this new block is
4910          * set to the appropriate level and owner.
4911          */
4912         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
4913
4914         __btrfs_tree_lock(buf, nest);
4915         btrfs_clean_tree_block(buf);
4916         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4917         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
4918
4919         set_extent_buffer_uptodate(buf);
4920
4921         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4922         btrfs_set_header_level(buf, level);
4923         btrfs_set_header_bytenr(buf, buf->start);
4924         btrfs_set_header_generation(buf, trans->transid);
4925         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4926         btrfs_set_header_owner(buf, owner);
4927         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4928         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4929         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4930                 buf->log_index = root->log_transid % 2;
4931                 /*
4932                  * we allow two log transactions at a time, use different
4933                  * EXTENT bit to differentiate dirty pages.
4934                  */
4935                 if (buf->log_index == 0)
4936                         set_extent_dirty(&root->dirty_log_pages, buf->start,
4937                                         buf->start + buf->len - 1, GFP_NOFS);
4938                 else
4939                         set_extent_new(&root->dirty_log_pages, buf->start,
4940                                         buf->start + buf->len - 1);
4941         } else {
4942                 buf->log_index = -1;
4943                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
4944                          buf->start + buf->len - 1, GFP_NOFS);
4945         }
4946         /* this returns a buffer locked for blocking */
4947         return buf;
4948 }
4949
4950 /*
4951  * finds a free extent and does all the dirty work required for allocation
4952  * returns the tree buffer or an ERR_PTR on error.
4953  */
4954 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4955                                              struct btrfs_root *root,
4956                                              u64 parent, u64 root_objectid,
4957                                              const struct btrfs_disk_key *key,
4958                                              int level, u64 hint,
4959                                              u64 empty_size,
4960                                              enum btrfs_lock_nesting nest)
4961 {
4962         struct btrfs_fs_info *fs_info = root->fs_info;
4963         struct btrfs_key ins;
4964         struct btrfs_block_rsv *block_rsv;
4965         struct extent_buffer *buf;
4966         struct btrfs_delayed_extent_op *extent_op;
4967         struct btrfs_ref generic_ref = { 0 };
4968         u64 flags = 0;
4969         int ret;
4970         u32 blocksize = fs_info->nodesize;
4971         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4972
4973 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4974         if (btrfs_is_testing(fs_info)) {
4975                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4976                                             level, root_objectid, nest);
4977                 if (!IS_ERR(buf))
4978                         root->alloc_bytenr += blocksize;
4979                 return buf;
4980         }
4981 #endif
4982
4983         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4984         if (IS_ERR(block_rsv))
4985                 return ERR_CAST(block_rsv);
4986
4987         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4988                                    empty_size, hint, &ins, 0, 0);
4989         if (ret)
4990                 goto out_unuse;
4991
4992         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4993                                     root_objectid, nest);
4994         if (IS_ERR(buf)) {
4995                 ret = PTR_ERR(buf);
4996                 goto out_free_reserved;
4997         }
4998
4999         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5000                 if (parent == 0)
5001                         parent = ins.objectid;
5002                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5003         } else
5004                 BUG_ON(parent > 0);
5005
5006         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5007                 extent_op = btrfs_alloc_delayed_extent_op();
5008                 if (!extent_op) {
5009                         ret = -ENOMEM;
5010                         goto out_free_buf;
5011                 }
5012                 if (key)
5013                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5014                 else
5015                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5016                 extent_op->flags_to_set = flags;
5017                 extent_op->update_key = skinny_metadata ? false : true;
5018                 extent_op->update_flags = true;
5019                 extent_op->level = level;
5020
5021                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
5022                                        ins.objectid, ins.offset, parent);
5023                 btrfs_init_tree_ref(&generic_ref, level, root_objectid,
5024                                     root->root_key.objectid, false);
5025                 btrfs_ref_tree_mod(fs_info, &generic_ref);
5026                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
5027                 if (ret)
5028                         goto out_free_delayed;
5029         }
5030         return buf;
5031
5032 out_free_delayed:
5033         btrfs_free_delayed_extent_op(extent_op);
5034 out_free_buf:
5035         btrfs_tree_unlock(buf);
5036         free_extent_buffer(buf);
5037 out_free_reserved:
5038         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
5039 out_unuse:
5040         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
5041         return ERR_PTR(ret);
5042 }
5043
5044 struct walk_control {
5045         u64 refs[BTRFS_MAX_LEVEL];
5046         u64 flags[BTRFS_MAX_LEVEL];
5047         struct btrfs_key update_progress;
5048         struct btrfs_key drop_progress;
5049         int drop_level;
5050         int stage;
5051         int level;
5052         int shared_level;
5053         int update_ref;
5054         int keep_locks;
5055         int reada_slot;
5056         int reada_count;
5057         int restarted;
5058 };
5059
5060 #define DROP_REFERENCE  1
5061 #define UPDATE_BACKREF  2
5062
5063 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5064                                      struct btrfs_root *root,
5065                                      struct walk_control *wc,
5066                                      struct btrfs_path *path)
5067 {
5068         struct btrfs_fs_info *fs_info = root->fs_info;
5069         u64 bytenr;
5070         u64 generation;
5071         u64 refs;
5072         u64 flags;
5073         u32 nritems;
5074         struct btrfs_key key;
5075         struct extent_buffer *eb;
5076         int ret;
5077         int slot;
5078         int nread = 0;
5079
5080         if (path->slots[wc->level] < wc->reada_slot) {
5081                 wc->reada_count = wc->reada_count * 2 / 3;
5082                 wc->reada_count = max(wc->reada_count, 2);
5083         } else {
5084                 wc->reada_count = wc->reada_count * 3 / 2;
5085                 wc->reada_count = min_t(int, wc->reada_count,
5086                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5087         }
5088
5089         eb = path->nodes[wc->level];
5090         nritems = btrfs_header_nritems(eb);
5091
5092         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5093                 if (nread >= wc->reada_count)
5094                         break;
5095
5096                 cond_resched();
5097                 bytenr = btrfs_node_blockptr(eb, slot);
5098                 generation = btrfs_node_ptr_generation(eb, slot);
5099
5100                 if (slot == path->slots[wc->level])
5101                         goto reada;
5102
5103                 if (wc->stage == UPDATE_BACKREF &&
5104                     generation <= root->root_key.offset)
5105                         continue;
5106
5107                 /* We don't lock the tree block, it's OK to be racy here */
5108                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5109                                                wc->level - 1, 1, &refs,
5110                                                &flags);
5111                 /* We don't care about errors in readahead. */
5112                 if (ret < 0)
5113                         continue;
5114                 BUG_ON(refs == 0);
5115
5116                 if (wc->stage == DROP_REFERENCE) {
5117                         if (refs == 1)
5118                                 goto reada;
5119
5120                         if (wc->level == 1 &&
5121                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5122                                 continue;
5123                         if (!wc->update_ref ||
5124                             generation <= root->root_key.offset)
5125                                 continue;
5126                         btrfs_node_key_to_cpu(eb, &key, slot);
5127                         ret = btrfs_comp_cpu_keys(&key,
5128                                                   &wc->update_progress);
5129                         if (ret < 0)
5130                                 continue;
5131                 } else {
5132                         if (wc->level == 1 &&
5133                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5134                                 continue;
5135                 }
5136 reada:
5137                 btrfs_readahead_node_child(eb, slot);
5138                 nread++;
5139         }
5140         wc->reada_slot = slot;
5141 }
5142
5143 /*
5144  * helper to process tree block while walking down the tree.
5145  *
5146  * when wc->stage == UPDATE_BACKREF, this function updates
5147  * back refs for pointers in the block.
5148  *
5149  * NOTE: return value 1 means we should stop walking down.
5150  */
5151 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5152                                    struct btrfs_root *root,
5153                                    struct btrfs_path *path,
5154                                    struct walk_control *wc, int lookup_info)
5155 {
5156         struct btrfs_fs_info *fs_info = root->fs_info;
5157         int level = wc->level;
5158         struct extent_buffer *eb = path->nodes[level];
5159         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5160         int ret;
5161
5162         if (wc->stage == UPDATE_BACKREF &&
5163             btrfs_header_owner(eb) != root->root_key.objectid)
5164                 return 1;
5165
5166         /*
5167          * when reference count of tree block is 1, it won't increase
5168          * again. once full backref flag is set, we never clear it.
5169          */
5170         if (lookup_info &&
5171             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5172              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5173                 BUG_ON(!path->locks[level]);
5174                 ret = btrfs_lookup_extent_info(trans, fs_info,
5175                                                eb->start, level, 1,
5176                                                &wc->refs[level],
5177                                                &wc->flags[level]);
5178                 BUG_ON(ret == -ENOMEM);
5179                 if (ret)
5180                         return ret;
5181                 BUG_ON(wc->refs[level] == 0);
5182         }
5183
5184         if (wc->stage == DROP_REFERENCE) {
5185                 if (wc->refs[level] > 1)
5186                         return 1;
5187
5188                 if (path->locks[level] && !wc->keep_locks) {
5189                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5190                         path->locks[level] = 0;
5191                 }
5192                 return 0;
5193         }
5194
5195         /* wc->stage == UPDATE_BACKREF */
5196         if (!(wc->flags[level] & flag)) {
5197                 BUG_ON(!path->locks[level]);
5198                 ret = btrfs_inc_ref(trans, root, eb, 1);
5199                 BUG_ON(ret); /* -ENOMEM */
5200                 ret = btrfs_dec_ref(trans, root, eb, 0);
5201                 BUG_ON(ret); /* -ENOMEM */
5202                 ret = btrfs_set_disk_extent_flags(trans, eb, flag,
5203                                                   btrfs_header_level(eb));
5204                 BUG_ON(ret); /* -ENOMEM */
5205                 wc->flags[level] |= flag;
5206         }
5207
5208         /*
5209          * the block is shared by multiple trees, so it's not good to
5210          * keep the tree lock
5211          */
5212         if (path->locks[level] && level > 0) {
5213                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5214                 path->locks[level] = 0;
5215         }
5216         return 0;
5217 }
5218
5219 /*
5220  * This is used to verify a ref exists for this root to deal with a bug where we
5221  * would have a drop_progress key that hadn't been updated properly.
5222  */
5223 static int check_ref_exists(struct btrfs_trans_handle *trans,
5224                             struct btrfs_root *root, u64 bytenr, u64 parent,
5225                             int level)
5226 {
5227         struct btrfs_path *path;
5228         struct btrfs_extent_inline_ref *iref;
5229         int ret;
5230
5231         path = btrfs_alloc_path();
5232         if (!path)
5233                 return -ENOMEM;
5234
5235         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5236                                     root->fs_info->nodesize, parent,
5237                                     root->root_key.objectid, level, 0);
5238         btrfs_free_path(path);
5239         if (ret == -ENOENT)
5240                 return 0;
5241         if (ret < 0)
5242                 return ret;
5243         return 1;
5244 }
5245
5246 /*
5247  * helper to process tree block pointer.
5248  *
5249  * when wc->stage == DROP_REFERENCE, this function checks
5250  * reference count of the block pointed to. if the block
5251  * is shared and we need update back refs for the subtree
5252  * rooted at the block, this function changes wc->stage to
5253  * UPDATE_BACKREF. if the block is shared and there is no
5254  * need to update back, this function drops the reference
5255  * to the block.
5256  *
5257  * NOTE: return value 1 means we should stop walking down.
5258  */
5259 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5260                                  struct btrfs_root *root,
5261                                  struct btrfs_path *path,
5262                                  struct walk_control *wc, int *lookup_info)
5263 {
5264         struct btrfs_fs_info *fs_info = root->fs_info;
5265         u64 bytenr;
5266         u64 generation;
5267         u64 parent;
5268         struct btrfs_tree_parent_check check = { 0 };
5269         struct btrfs_key key;
5270         struct btrfs_ref ref = { 0 };
5271         struct extent_buffer *next;
5272         int level = wc->level;
5273         int reada = 0;
5274         int ret = 0;
5275         bool need_account = false;
5276
5277         generation = btrfs_node_ptr_generation(path->nodes[level],
5278                                                path->slots[level]);
5279         /*
5280          * if the lower level block was created before the snapshot
5281          * was created, we know there is no need to update back refs
5282          * for the subtree
5283          */
5284         if (wc->stage == UPDATE_BACKREF &&
5285             generation <= root->root_key.offset) {
5286                 *lookup_info = 1;
5287                 return 1;
5288         }
5289
5290         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5291
5292         check.level = level - 1;
5293         check.transid = generation;
5294         check.owner_root = root->root_key.objectid;
5295         check.has_first_key = true;
5296         btrfs_node_key_to_cpu(path->nodes[level], &check.first_key,
5297                               path->slots[level]);
5298
5299         next = find_extent_buffer(fs_info, bytenr);
5300         if (!next) {
5301                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5302                                 root->root_key.objectid, level - 1);
5303                 if (IS_ERR(next))
5304                         return PTR_ERR(next);
5305                 reada = 1;
5306         }
5307         btrfs_tree_lock(next);
5308
5309         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5310                                        &wc->refs[level - 1],
5311                                        &wc->flags[level - 1]);
5312         if (ret < 0)
5313                 goto out_unlock;
5314
5315         if (unlikely(wc->refs[level - 1] == 0)) {
5316                 btrfs_err(fs_info, "Missing references.");
5317                 ret = -EIO;
5318                 goto out_unlock;
5319         }
5320         *lookup_info = 0;
5321
5322         if (wc->stage == DROP_REFERENCE) {
5323                 if (wc->refs[level - 1] > 1) {
5324                         need_account = true;
5325                         if (level == 1 &&
5326                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5327                                 goto skip;
5328
5329                         if (!wc->update_ref ||
5330                             generation <= root->root_key.offset)
5331                                 goto skip;
5332
5333                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5334                                               path->slots[level]);
5335                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5336                         if (ret < 0)
5337                                 goto skip;
5338
5339                         wc->stage = UPDATE_BACKREF;
5340                         wc->shared_level = level - 1;
5341                 }
5342         } else {
5343                 if (level == 1 &&
5344                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5345                         goto skip;
5346         }
5347
5348         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5349                 btrfs_tree_unlock(next);
5350                 free_extent_buffer(next);
5351                 next = NULL;
5352                 *lookup_info = 1;
5353         }
5354
5355         if (!next) {
5356                 if (reada && level == 1)
5357                         reada_walk_down(trans, root, wc, path);
5358                 next = read_tree_block(fs_info, bytenr, &check);
5359                 if (IS_ERR(next)) {
5360                         return PTR_ERR(next);
5361                 } else if (!extent_buffer_uptodate(next)) {
5362                         free_extent_buffer(next);
5363                         return -EIO;
5364                 }
5365                 btrfs_tree_lock(next);
5366         }
5367
5368         level--;
5369         ASSERT(level == btrfs_header_level(next));
5370         if (level != btrfs_header_level(next)) {
5371                 btrfs_err(root->fs_info, "mismatched level");
5372                 ret = -EIO;
5373                 goto out_unlock;
5374         }
5375         path->nodes[level] = next;
5376         path->slots[level] = 0;
5377         path->locks[level] = BTRFS_WRITE_LOCK;
5378         wc->level = level;
5379         if (wc->level == 1)
5380                 wc->reada_slot = 0;
5381         return 0;
5382 skip:
5383         wc->refs[level - 1] = 0;
5384         wc->flags[level - 1] = 0;
5385         if (wc->stage == DROP_REFERENCE) {
5386                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5387                         parent = path->nodes[level]->start;
5388                 } else {
5389                         ASSERT(root->root_key.objectid ==
5390                                btrfs_header_owner(path->nodes[level]));
5391                         if (root->root_key.objectid !=
5392                             btrfs_header_owner(path->nodes[level])) {
5393                                 btrfs_err(root->fs_info,
5394                                                 "mismatched block owner");
5395                                 ret = -EIO;
5396                                 goto out_unlock;
5397                         }
5398                         parent = 0;
5399                 }
5400
5401                 /*
5402                  * If we had a drop_progress we need to verify the refs are set
5403                  * as expected.  If we find our ref then we know that from here
5404                  * on out everything should be correct, and we can clear the
5405                  * ->restarted flag.
5406                  */
5407                 if (wc->restarted) {
5408                         ret = check_ref_exists(trans, root, bytenr, parent,
5409                                                level - 1);
5410                         if (ret < 0)
5411                                 goto out_unlock;
5412                         if (ret == 0)
5413                                 goto no_delete;
5414                         ret = 0;
5415                         wc->restarted = 0;
5416                 }
5417
5418                 /*
5419                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5420                  * already accounted them at merge time (replace_path),
5421                  * thus we could skip expensive subtree trace here.
5422                  */
5423                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5424                     need_account) {
5425                         ret = btrfs_qgroup_trace_subtree(trans, next,
5426                                                          generation, level - 1);
5427                         if (ret) {
5428                                 btrfs_err_rl(fs_info,
5429                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5430                                              ret);
5431                         }
5432                 }
5433
5434                 /*
5435                  * We need to update the next key in our walk control so we can
5436                  * update the drop_progress key accordingly.  We don't care if
5437                  * find_next_key doesn't find a key because that means we're at
5438                  * the end and are going to clean up now.
5439                  */
5440                 wc->drop_level = level;
5441                 find_next_key(path, level, &wc->drop_progress);
5442
5443                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5444                                        fs_info->nodesize, parent);
5445                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid,
5446                                     0, false);
5447                 ret = btrfs_free_extent(trans, &ref);
5448                 if (ret)
5449                         goto out_unlock;
5450         }
5451 no_delete:
5452         *lookup_info = 1;
5453         ret = 1;
5454
5455 out_unlock:
5456         btrfs_tree_unlock(next);
5457         free_extent_buffer(next);
5458
5459         return ret;
5460 }
5461
5462 /*
5463  * helper to process tree block while walking up the tree.
5464  *
5465  * when wc->stage == DROP_REFERENCE, this function drops
5466  * reference count on the block.
5467  *
5468  * when wc->stage == UPDATE_BACKREF, this function changes
5469  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5470  * to UPDATE_BACKREF previously while processing the block.
5471  *
5472  * NOTE: return value 1 means we should stop walking up.
5473  */
5474 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5475                                  struct btrfs_root *root,
5476                                  struct btrfs_path *path,
5477                                  struct walk_control *wc)
5478 {
5479         struct btrfs_fs_info *fs_info = root->fs_info;
5480         int ret;
5481         int level = wc->level;
5482         struct extent_buffer *eb = path->nodes[level];
5483         u64 parent = 0;
5484
5485         if (wc->stage == UPDATE_BACKREF) {
5486                 BUG_ON(wc->shared_level < level);
5487                 if (level < wc->shared_level)
5488                         goto out;
5489
5490                 ret = find_next_key(path, level + 1, &wc->update_progress);
5491                 if (ret > 0)
5492                         wc->update_ref = 0;
5493
5494                 wc->stage = DROP_REFERENCE;
5495                 wc->shared_level = -1;
5496                 path->slots[level] = 0;
5497
5498                 /*
5499                  * check reference count again if the block isn't locked.
5500                  * we should start walking down the tree again if reference
5501                  * count is one.
5502                  */
5503                 if (!path->locks[level]) {
5504                         BUG_ON(level == 0);
5505                         btrfs_tree_lock(eb);
5506                         path->locks[level] = BTRFS_WRITE_LOCK;
5507
5508                         ret = btrfs_lookup_extent_info(trans, fs_info,
5509                                                        eb->start, level, 1,
5510                                                        &wc->refs[level],
5511                                                        &wc->flags[level]);
5512                         if (ret < 0) {
5513                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5514                                 path->locks[level] = 0;
5515                                 return ret;
5516                         }
5517                         BUG_ON(wc->refs[level] == 0);
5518                         if (wc->refs[level] == 1) {
5519                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5520                                 path->locks[level] = 0;
5521                                 return 1;
5522                         }
5523                 }
5524         }
5525
5526         /* wc->stage == DROP_REFERENCE */
5527         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5528
5529         if (wc->refs[level] == 1) {
5530                 if (level == 0) {
5531                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5532                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5533                         else
5534                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5535                         BUG_ON(ret); /* -ENOMEM */
5536                         if (is_fstree(root->root_key.objectid)) {
5537                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5538                                 if (ret) {
5539                                         btrfs_err_rl(fs_info,
5540         "error %d accounting leaf items, quota is out of sync, rescan required",
5541                                              ret);
5542                                 }
5543                         }
5544                 }
5545                 /* make block locked assertion in btrfs_clean_tree_block happy */
5546                 if (!path->locks[level] &&
5547                     btrfs_header_generation(eb) == trans->transid) {
5548                         btrfs_tree_lock(eb);
5549                         path->locks[level] = BTRFS_WRITE_LOCK;
5550                 }
5551                 btrfs_clean_tree_block(eb);
5552         }
5553
5554         if (eb == root->node) {
5555                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5556                         parent = eb->start;
5557                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5558                         goto owner_mismatch;
5559         } else {
5560                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5561                         parent = path->nodes[level + 1]->start;
5562                 else if (root->root_key.objectid !=
5563                          btrfs_header_owner(path->nodes[level + 1]))
5564                         goto owner_mismatch;
5565         }
5566
5567         btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5568                               wc->refs[level] == 1);
5569 out:
5570         wc->refs[level] = 0;
5571         wc->flags[level] = 0;
5572         return 0;
5573
5574 owner_mismatch:
5575         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5576                      btrfs_header_owner(eb), root->root_key.objectid);
5577         return -EUCLEAN;
5578 }
5579
5580 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5581                                    struct btrfs_root *root,
5582                                    struct btrfs_path *path,
5583                                    struct walk_control *wc)
5584 {
5585         int level = wc->level;
5586         int lookup_info = 1;
5587         int ret;
5588
5589         while (level >= 0) {
5590                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5591                 if (ret > 0)
5592                         break;
5593
5594                 if (level == 0)
5595                         break;
5596
5597                 if (path->slots[level] >=
5598                     btrfs_header_nritems(path->nodes[level]))
5599                         break;
5600
5601                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5602                 if (ret > 0) {
5603                         path->slots[level]++;
5604                         continue;
5605                 } else if (ret < 0)
5606                         return ret;
5607                 level = wc->level;
5608         }
5609         return 0;
5610 }
5611
5612 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5613                                  struct btrfs_root *root,
5614                                  struct btrfs_path *path,
5615                                  struct walk_control *wc, int max_level)
5616 {
5617         int level = wc->level;
5618         int ret;
5619
5620         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5621         while (level < max_level && path->nodes[level]) {
5622                 wc->level = level;
5623                 if (path->slots[level] + 1 <
5624                     btrfs_header_nritems(path->nodes[level])) {
5625                         path->slots[level]++;
5626                         return 0;
5627                 } else {
5628                         ret = walk_up_proc(trans, root, path, wc);
5629                         if (ret > 0)
5630                                 return 0;
5631                         if (ret < 0)
5632                                 return ret;
5633
5634                         if (path->locks[level]) {
5635                                 btrfs_tree_unlock_rw(path->nodes[level],
5636                                                      path->locks[level]);
5637                                 path->locks[level] = 0;
5638                         }
5639                         free_extent_buffer(path->nodes[level]);
5640                         path->nodes[level] = NULL;
5641                         level++;
5642                 }
5643         }
5644         return 1;
5645 }
5646
5647 /*
5648  * drop a subvolume tree.
5649  *
5650  * this function traverses the tree freeing any blocks that only
5651  * referenced by the tree.
5652  *
5653  * when a shared tree block is found. this function decreases its
5654  * reference count by one. if update_ref is true, this function
5655  * also make sure backrefs for the shared block and all lower level
5656  * blocks are properly updated.
5657  *
5658  * If called with for_reloc == 0, may exit early with -EAGAIN
5659  */
5660 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5661 {
5662         const bool is_reloc_root = (root->root_key.objectid ==
5663                                     BTRFS_TREE_RELOC_OBJECTID);
5664         struct btrfs_fs_info *fs_info = root->fs_info;
5665         struct btrfs_path *path;
5666         struct btrfs_trans_handle *trans;
5667         struct btrfs_root *tree_root = fs_info->tree_root;
5668         struct btrfs_root_item *root_item = &root->root_item;
5669         struct walk_control *wc;
5670         struct btrfs_key key;
5671         int err = 0;
5672         int ret;
5673         int level;
5674         bool root_dropped = false;
5675         bool unfinished_drop = false;
5676
5677         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5678
5679         path = btrfs_alloc_path();
5680         if (!path) {
5681                 err = -ENOMEM;
5682                 goto out;
5683         }
5684
5685         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5686         if (!wc) {
5687                 btrfs_free_path(path);
5688                 err = -ENOMEM;
5689                 goto out;
5690         }
5691
5692         /*
5693          * Use join to avoid potential EINTR from transaction start. See
5694          * wait_reserve_ticket and the whole reservation callchain.
5695          */
5696         if (for_reloc)
5697                 trans = btrfs_join_transaction(tree_root);
5698         else
5699                 trans = btrfs_start_transaction(tree_root, 0);
5700         if (IS_ERR(trans)) {
5701                 err = PTR_ERR(trans);
5702                 goto out_free;
5703         }
5704
5705         err = btrfs_run_delayed_items(trans);
5706         if (err)
5707                 goto out_end_trans;
5708
5709         /*
5710          * This will help us catch people modifying the fs tree while we're
5711          * dropping it.  It is unsafe to mess with the fs tree while it's being
5712          * dropped as we unlock the root node and parent nodes as we walk down
5713          * the tree, assuming nothing will change.  If something does change
5714          * then we'll have stale information and drop references to blocks we've
5715          * already dropped.
5716          */
5717         set_bit(BTRFS_ROOT_DELETING, &root->state);
5718         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
5719
5720         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5721                 level = btrfs_header_level(root->node);
5722                 path->nodes[level] = btrfs_lock_root_node(root);
5723                 path->slots[level] = 0;
5724                 path->locks[level] = BTRFS_WRITE_LOCK;
5725                 memset(&wc->update_progress, 0,
5726                        sizeof(wc->update_progress));
5727         } else {
5728                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5729                 memcpy(&wc->update_progress, &key,
5730                        sizeof(wc->update_progress));
5731
5732                 level = btrfs_root_drop_level(root_item);
5733                 BUG_ON(level == 0);
5734                 path->lowest_level = level;
5735                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5736                 path->lowest_level = 0;
5737                 if (ret < 0) {
5738                         err = ret;
5739                         goto out_end_trans;
5740                 }
5741                 WARN_ON(ret > 0);
5742
5743                 /*
5744                  * unlock our path, this is safe because only this
5745                  * function is allowed to delete this snapshot
5746                  */
5747                 btrfs_unlock_up_safe(path, 0);
5748
5749                 level = btrfs_header_level(root->node);
5750                 while (1) {
5751                         btrfs_tree_lock(path->nodes[level]);
5752                         path->locks[level] = BTRFS_WRITE_LOCK;
5753
5754                         ret = btrfs_lookup_extent_info(trans, fs_info,
5755                                                 path->nodes[level]->start,
5756                                                 level, 1, &wc->refs[level],
5757                                                 &wc->flags[level]);
5758                         if (ret < 0) {
5759                                 err = ret;
5760                                 goto out_end_trans;
5761                         }
5762                         BUG_ON(wc->refs[level] == 0);
5763
5764                         if (level == btrfs_root_drop_level(root_item))
5765                                 break;
5766
5767                         btrfs_tree_unlock(path->nodes[level]);
5768                         path->locks[level] = 0;
5769                         WARN_ON(wc->refs[level] != 1);
5770                         level--;
5771                 }
5772         }
5773
5774         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5775         wc->level = level;
5776         wc->shared_level = -1;
5777         wc->stage = DROP_REFERENCE;
5778         wc->update_ref = update_ref;
5779         wc->keep_locks = 0;
5780         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5781
5782         while (1) {
5783
5784                 ret = walk_down_tree(trans, root, path, wc);
5785                 if (ret < 0) {
5786                         err = ret;
5787                         break;
5788                 }
5789
5790                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5791                 if (ret < 0) {
5792                         err = ret;
5793                         break;
5794                 }
5795
5796                 if (ret > 0) {
5797                         BUG_ON(wc->stage != DROP_REFERENCE);
5798                         break;
5799                 }
5800
5801                 if (wc->stage == DROP_REFERENCE) {
5802                         wc->drop_level = wc->level;
5803                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5804                                               &wc->drop_progress,
5805                                               path->slots[wc->drop_level]);
5806                 }
5807                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5808                                       &wc->drop_progress);
5809                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5810
5811                 BUG_ON(wc->level == 0);
5812                 if (btrfs_should_end_transaction(trans) ||
5813                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5814                         ret = btrfs_update_root(trans, tree_root,
5815                                                 &root->root_key,
5816                                                 root_item);
5817                         if (ret) {
5818                                 btrfs_abort_transaction(trans, ret);
5819                                 err = ret;
5820                                 goto out_end_trans;
5821                         }
5822
5823                         if (!is_reloc_root)
5824                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5825
5826                         btrfs_end_transaction_throttle(trans);
5827                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5828                                 btrfs_debug(fs_info,
5829                                             "drop snapshot early exit");
5830                                 err = -EAGAIN;
5831                                 goto out_free;
5832                         }
5833
5834                        /*
5835                         * Use join to avoid potential EINTR from transaction
5836                         * start. See wait_reserve_ticket and the whole
5837                         * reservation callchain.
5838                         */
5839                         if (for_reloc)
5840                                 trans = btrfs_join_transaction(tree_root);
5841                         else
5842                                 trans = btrfs_start_transaction(tree_root, 0);
5843                         if (IS_ERR(trans)) {
5844                                 err = PTR_ERR(trans);
5845                                 goto out_free;
5846                         }
5847                 }
5848         }
5849         btrfs_release_path(path);
5850         if (err)
5851                 goto out_end_trans;
5852
5853         ret = btrfs_del_root(trans, &root->root_key);
5854         if (ret) {
5855                 btrfs_abort_transaction(trans, ret);
5856                 err = ret;
5857                 goto out_end_trans;
5858         }
5859
5860         if (!is_reloc_root) {
5861                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5862                                       NULL, NULL);
5863                 if (ret < 0) {
5864                         btrfs_abort_transaction(trans, ret);
5865                         err = ret;
5866                         goto out_end_trans;
5867                 } else if (ret > 0) {
5868                         /* if we fail to delete the orphan item this time
5869                          * around, it'll get picked up the next time.
5870                          *
5871                          * The most common failure here is just -ENOENT.
5872                          */
5873                         btrfs_del_orphan_item(trans, tree_root,
5874                                               root->root_key.objectid);
5875                 }
5876         }
5877
5878         /*
5879          * This subvolume is going to be completely dropped, and won't be
5880          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5881          * commit transaction time.  So free it here manually.
5882          */
5883         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5884         btrfs_qgroup_free_meta_all_pertrans(root);
5885
5886         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5887                 btrfs_add_dropped_root(trans, root);
5888         else
5889                 btrfs_put_root(root);
5890         root_dropped = true;
5891 out_end_trans:
5892         if (!is_reloc_root)
5893                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5894
5895         btrfs_end_transaction_throttle(trans);
5896 out_free:
5897         kfree(wc);
5898         btrfs_free_path(path);
5899 out:
5900         /*
5901          * We were an unfinished drop root, check to see if there are any
5902          * pending, and if not clear and wake up any waiters.
5903          */
5904         if (!err && unfinished_drop)
5905                 btrfs_maybe_wake_unfinished_drop(fs_info);
5906
5907         /*
5908          * So if we need to stop dropping the snapshot for whatever reason we
5909          * need to make sure to add it back to the dead root list so that we
5910          * keep trying to do the work later.  This also cleans up roots if we
5911          * don't have it in the radix (like when we recover after a power fail
5912          * or unmount) so we don't leak memory.
5913          */
5914         if (!for_reloc && !root_dropped)
5915                 btrfs_add_dead_root(root);
5916         return err;
5917 }
5918
5919 /*
5920  * drop subtree rooted at tree block 'node'.
5921  *
5922  * NOTE: this function will unlock and release tree block 'node'
5923  * only used by relocation code
5924  */
5925 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5926                         struct btrfs_root *root,
5927                         struct extent_buffer *node,
5928                         struct extent_buffer *parent)
5929 {
5930         struct btrfs_fs_info *fs_info = root->fs_info;
5931         struct btrfs_path *path;
5932         struct walk_control *wc;
5933         int level;
5934         int parent_level;
5935         int ret = 0;
5936         int wret;
5937
5938         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5939
5940         path = btrfs_alloc_path();
5941         if (!path)
5942                 return -ENOMEM;
5943
5944         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5945         if (!wc) {
5946                 btrfs_free_path(path);
5947                 return -ENOMEM;
5948         }
5949
5950         btrfs_assert_tree_write_locked(parent);
5951         parent_level = btrfs_header_level(parent);
5952         atomic_inc(&parent->refs);
5953         path->nodes[parent_level] = parent;
5954         path->slots[parent_level] = btrfs_header_nritems(parent);
5955
5956         btrfs_assert_tree_write_locked(node);
5957         level = btrfs_header_level(node);
5958         path->nodes[level] = node;
5959         path->slots[level] = 0;
5960         path->locks[level] = BTRFS_WRITE_LOCK;
5961
5962         wc->refs[parent_level] = 1;
5963         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5964         wc->level = level;
5965         wc->shared_level = -1;
5966         wc->stage = DROP_REFERENCE;
5967         wc->update_ref = 0;
5968         wc->keep_locks = 1;
5969         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5970
5971         while (1) {
5972                 wret = walk_down_tree(trans, root, path, wc);
5973                 if (wret < 0) {
5974                         ret = wret;
5975                         break;
5976                 }
5977
5978                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5979                 if (wret < 0)
5980                         ret = wret;
5981                 if (wret != 0)
5982                         break;
5983         }
5984
5985         kfree(wc);
5986         btrfs_free_path(path);
5987         return ret;
5988 }
5989
5990 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5991                                    u64 start, u64 end)
5992 {
5993         return unpin_extent_range(fs_info, start, end, false);
5994 }
5995
5996 /*
5997  * It used to be that old block groups would be left around forever.
5998  * Iterating over them would be enough to trim unused space.  Since we
5999  * now automatically remove them, we also need to iterate over unallocated
6000  * space.
6001  *
6002  * We don't want a transaction for this since the discard may take a
6003  * substantial amount of time.  We don't require that a transaction be
6004  * running, but we do need to take a running transaction into account
6005  * to ensure that we're not discarding chunks that were released or
6006  * allocated in the current transaction.
6007  *
6008  * Holding the chunks lock will prevent other threads from allocating
6009  * or releasing chunks, but it won't prevent a running transaction
6010  * from committing and releasing the memory that the pending chunks
6011  * list head uses.  For that, we need to take a reference to the
6012  * transaction and hold the commit root sem.  We only need to hold
6013  * it while performing the free space search since we have already
6014  * held back allocations.
6015  */
6016 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
6017 {
6018         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
6019         int ret;
6020
6021         *trimmed = 0;
6022
6023         /* Discard not supported = nothing to do. */
6024         if (!bdev_max_discard_sectors(device->bdev))
6025                 return 0;
6026
6027         /* Not writable = nothing to do. */
6028         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
6029                 return 0;
6030
6031         /* No free space = nothing to do. */
6032         if (device->total_bytes <= device->bytes_used)
6033                 return 0;
6034
6035         ret = 0;
6036
6037         while (1) {
6038                 struct btrfs_fs_info *fs_info = device->fs_info;
6039                 u64 bytes;
6040
6041                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
6042                 if (ret)
6043                         break;
6044
6045                 find_first_clear_extent_bit(&device->alloc_state, start,
6046                                             &start, &end,
6047                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
6048
6049                 /* Check if there are any CHUNK_* bits left */
6050                 if (start > device->total_bytes) {
6051                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6052                         btrfs_warn_in_rcu(fs_info,
6053 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
6054                                           start, end - start + 1,
6055                                           btrfs_dev_name(device),
6056                                           device->total_bytes);
6057                         mutex_unlock(&fs_info->chunk_mutex);
6058                         ret = 0;
6059                         break;
6060                 }
6061
6062                 /* Ensure we skip the reserved space on each device. */
6063                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
6064
6065                 /*
6066                  * If find_first_clear_extent_bit find a range that spans the
6067                  * end of the device it will set end to -1, in this case it's up
6068                  * to the caller to trim the value to the size of the device.
6069                  */
6070                 end = min(end, device->total_bytes - 1);
6071
6072                 len = end - start + 1;
6073
6074                 /* We didn't find any extents */
6075                 if (!len) {
6076                         mutex_unlock(&fs_info->chunk_mutex);
6077                         ret = 0;
6078                         break;
6079                 }
6080
6081                 ret = btrfs_issue_discard(device->bdev, start, len,
6082                                           &bytes);
6083                 if (!ret)
6084                         set_extent_bits(&device->alloc_state, start,
6085                                         start + bytes - 1,
6086                                         CHUNK_TRIMMED);
6087                 mutex_unlock(&fs_info->chunk_mutex);
6088
6089                 if (ret)
6090                         break;
6091
6092                 start += len;
6093                 *trimmed += bytes;
6094
6095                 if (fatal_signal_pending(current)) {
6096                         ret = -ERESTARTSYS;
6097                         break;
6098                 }
6099
6100                 cond_resched();
6101         }
6102
6103         return ret;
6104 }
6105
6106 /*
6107  * Trim the whole filesystem by:
6108  * 1) trimming the free space in each block group
6109  * 2) trimming the unallocated space on each device
6110  *
6111  * This will also continue trimming even if a block group or device encounters
6112  * an error.  The return value will be the last error, or 0 if nothing bad
6113  * happens.
6114  */
6115 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6116 {
6117         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6118         struct btrfs_block_group *cache = NULL;
6119         struct btrfs_device *device;
6120         u64 group_trimmed;
6121         u64 range_end = U64_MAX;
6122         u64 start;
6123         u64 end;
6124         u64 trimmed = 0;
6125         u64 bg_failed = 0;
6126         u64 dev_failed = 0;
6127         int bg_ret = 0;
6128         int dev_ret = 0;
6129         int ret = 0;
6130
6131         if (range->start == U64_MAX)
6132                 return -EINVAL;
6133
6134         /*
6135          * Check range overflow if range->len is set.
6136          * The default range->len is U64_MAX.
6137          */
6138         if (range->len != U64_MAX &&
6139             check_add_overflow(range->start, range->len, &range_end))
6140                 return -EINVAL;
6141
6142         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6143         for (; cache; cache = btrfs_next_block_group(cache)) {
6144                 if (cache->start >= range_end) {
6145                         btrfs_put_block_group(cache);
6146                         break;
6147                 }
6148
6149                 start = max(range->start, cache->start);
6150                 end = min(range_end, cache->start + cache->length);
6151
6152                 if (end - start >= range->minlen) {
6153                         if (!btrfs_block_group_done(cache)) {
6154                                 ret = btrfs_cache_block_group(cache, true);
6155                                 if (ret) {
6156                                         bg_failed++;
6157                                         bg_ret = ret;
6158                                         continue;
6159                                 }
6160                         }
6161                         ret = btrfs_trim_block_group(cache,
6162                                                      &group_trimmed,
6163                                                      start,
6164                                                      end,
6165                                                      range->minlen);
6166
6167                         trimmed += group_trimmed;
6168                         if (ret) {
6169                                 bg_failed++;
6170                                 bg_ret = ret;
6171                                 continue;
6172                         }
6173                 }
6174         }
6175
6176         if (bg_failed)
6177                 btrfs_warn(fs_info,
6178                         "failed to trim %llu block group(s), last error %d",
6179                         bg_failed, bg_ret);
6180
6181         mutex_lock(&fs_devices->device_list_mutex);
6182         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6183                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6184                         continue;
6185
6186                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6187                 if (ret) {
6188                         dev_failed++;
6189                         dev_ret = ret;
6190                         break;
6191                 }
6192
6193                 trimmed += group_trimmed;
6194         }
6195         mutex_unlock(&fs_devices->device_list_mutex);
6196
6197         if (dev_failed)
6198                 btrfs_warn(fs_info,
6199                         "failed to trim %llu device(s), last error %d",
6200                         dev_failed, dev_ret);
6201         range->len = trimmed;
6202         if (bg_ret)
6203                 return bg_ret;
6204         return dev_ret;
6205 }