fs/btrfs/extent-tree.c

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