fs/btrfs/backref.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/mm.h>
   7 #include <linux/rbtree.h>
   8 #include <trace/events/btrfs.h>
   9 #include "ctree.h"
  10 #include "disk-io.h"
  11 #include "backref.h"
  12 #include "ulist.h"
  13 #include "transaction.h"
  14 #include "delayed-ref.h"
  15 #include "locking.h"
  16 #include "misc.h"
  17 #include "tree-mod-log.h"
  18
  19 /* Just an arbitrary number so we can be sure this happened */
  20 #define BACKREF_FOUND_SHARED 6
  21
  22 struct extent_inode_elem {
  23         u64 inum;
  24         u64 offset;
  25         struct extent_inode_elem *next;
  26 };
  27
  28 static int check_extent_in_eb(const struct btrfs_key *key,
  29                               const struct extent_buffer *eb,
  30                               const struct btrfs_file_extent_item *fi,
  31                               u64 extent_item_pos,
  32                               struct extent_inode_elem **eie,
  33                               bool ignore_offset)
  34 {
  35         u64 offset = 0;
  36         struct extent_inode_elem *e;
  37
  38         if (!ignore_offset &&
  39             !btrfs_file_extent_compression(eb, fi) &&
  40             !btrfs_file_extent_encryption(eb, fi) &&
  41             !btrfs_file_extent_other_encoding(eb, fi)) {
  42                 u64 data_offset;
  43                 u64 data_len;
  44
  45                 data_offset = btrfs_file_extent_offset(eb, fi);
  46                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  47
  48                 if (extent_item_pos < data_offset ||
  49                     extent_item_pos >= data_offset + data_len)
  50                         return 1;
  51                 offset = extent_item_pos - data_offset;
  52         }
  53
  54         e = kmalloc(sizeof(*e), GFP_NOFS);
  55         if (!e)
  56                 return -ENOMEM;
  57
  58         e->next = *eie;
  59         e->inum = key->objectid;
  60         e->offset = key->offset + offset;
  61         *eie = e;
  62
  63         return 0;
  64 }
  65
  66 static void free_inode_elem_list(struct extent_inode_elem *eie)
  67 {
  68         struct extent_inode_elem *eie_next;
  69
  70         for (; eie; eie = eie_next) {
  71                 eie_next = eie->next;
  72                 kfree(eie);
  73         }
  74 }
  75
  76 static int find_extent_in_eb(const struct extent_buffer *eb,
  77                              u64 wanted_disk_byte, u64 extent_item_pos,
  78                              struct extent_inode_elem **eie,
  79                              bool ignore_offset)
  80 {
  81         u64 disk_byte;
  82         struct btrfs_key key;
  83         struct btrfs_file_extent_item *fi;
  84         int slot;
  85         int nritems;
  86         int extent_type;
  87         int ret;
  88
  89         /*
  90          * from the shared data ref, we only have the leaf but we need
  91          * the key. thus, we must look into all items and see that we
  92          * find one (some) with a reference to our extent item.
  93          */
  94         nritems = btrfs_header_nritems(eb);
  95         for (slot = 0; slot < nritems; ++slot) {
  96                 btrfs_item_key_to_cpu(eb, &key, slot);
  97                 if (key.type != BTRFS_EXTENT_DATA_KEY)
  98                         continue;
  99                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 100                 extent_type = btrfs_file_extent_type(eb, fi);
 101                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 102                         continue;
 103                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 104                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 105                 if (disk_byte != wanted_disk_byte)
 106                         continue;
 107
 108                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
 109                 if (ret < 0)
 110                         return ret;
 111         }
 112
 113         return 0;
 114 }
 115
 116 struct preftree {
 117         struct rb_root_cached root;
 118         unsigned int count;
 119 };
 120
 121 #define PREFTREE_INIT   { .root = RB_ROOT_CACHED, .count = 0 }
 122
 123 struct preftrees {
 124         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 125         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 126         struct preftree indirect_missing_keys;
 127 };
 128
 129 /*
 130  * Checks for a shared extent during backref search.
 131  *
 132  * The share_count tracks prelim_refs (direct and indirect) having a
 133  * ref->count >0:
 134  *  - incremented when a ref->count transitions to >0
 135  *  - decremented when a ref->count transitions to <1
 136  */
 137 struct share_check {
 138         u64 root_objectid;
 139         u64 inum;
 140         int share_count;
 141         bool have_delayed_delete_refs;
 142 };
 143
 144 static inline int extent_is_shared(struct share_check *sc)
 145 {
 146         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 147 }
 148
 149 static struct kmem_cache *btrfs_prelim_ref_cache;
 150
 151 int __init btrfs_prelim_ref_init(void)
 152 {
 153         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 154                                         sizeof(struct prelim_ref),
 155                                         0,
 156                                         SLAB_MEM_SPREAD,
 157                                         NULL);
 158         if (!btrfs_prelim_ref_cache)
 159                 return -ENOMEM;
 160         return 0;
 161 }
 162
 163 void __cold btrfs_prelim_ref_exit(void)
 164 {
 165         kmem_cache_destroy(btrfs_prelim_ref_cache);
 166 }
 167
 168 static void free_pref(struct prelim_ref *ref)
 169 {
 170         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 171 }
 172
 173 /*
 174  * Return 0 when both refs are for the same block (and can be merged).
 175  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 176  * indicates a 'higher' block.
 177  */
 178 static int prelim_ref_compare(struct prelim_ref *ref1,
 179                               struct prelim_ref *ref2)
 180 {
 181         if (ref1->level < ref2->level)
 182                 return -1;
 183         if (ref1->level > ref2->level)
 184                 return 1;
 185         if (ref1->root_id < ref2->root_id)
 186                 return -1;
 187         if (ref1->root_id > ref2->root_id)
 188                 return 1;
 189         if (ref1->key_for_search.type < ref2->key_for_search.type)
 190                 return -1;
 191         if (ref1->key_for_search.type > ref2->key_for_search.type)
 192                 return 1;
 193         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 194                 return -1;
 195         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 196                 return 1;
 197         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 198                 return -1;
 199         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 200                 return 1;
 201         if (ref1->parent < ref2->parent)
 202                 return -1;
 203         if (ref1->parent > ref2->parent)
 204                 return 1;
 205
 206         return 0;
 207 }
 208
 209 static void update_share_count(struct share_check *sc, int oldcount,
 210                                int newcount)
 211 {
 212         if ((!sc) || (oldcount == 0 && newcount < 1))
 213                 return;
 214
 215         if (oldcount > 0 && newcount < 1)
 216                 sc->share_count--;
 217         else if (oldcount < 1 && newcount > 0)
 218                 sc->share_count++;
 219 }
 220
 221 /*
 222  * Add @newref to the @root rbtree, merging identical refs.
 223  *
 224  * Callers should assume that newref has been freed after calling.
 225  */
 226 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 227                               struct preftree *preftree,
 228                               struct prelim_ref *newref,
 229                               struct share_check *sc)
 230 {
 231         struct rb_root_cached *root;
 232         struct rb_node **p;
 233         struct rb_node *parent = NULL;
 234         struct prelim_ref *ref;
 235         int result;
 236         bool leftmost = true;
 237
 238         root = &preftree->root;
 239         p = &root->rb_root.rb_node;
 240
 241         while (*p) {
 242                 parent = *p;
 243                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 244                 result = prelim_ref_compare(ref, newref);
 245                 if (result < 0) {
 246                         p = &(*p)->rb_left;
 247                 } else if (result > 0) {
 248                         p = &(*p)->rb_right;
 249                         leftmost = false;
 250                 } else {
 251                         /* Identical refs, merge them and free @newref */
 252                         struct extent_inode_elem *eie = ref->inode_list;
 253
 254                         while (eie && eie->next)
 255                                 eie = eie->next;
 256
 257                         if (!eie)
 258                                 ref->inode_list = newref->inode_list;
 259                         else
 260                                 eie->next = newref->inode_list;
 261                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 262                                                      preftree->count);
 263                         /*
 264                          * A delayed ref can have newref->count < 0.
 265                          * The ref->count is updated to follow any
 266                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 267                          */
 268                         update_share_count(sc, ref->count,
 269                                            ref->count + newref->count);
 270                         ref->count += newref->count;
 271                         free_pref(newref);
 272                         return;
 273                 }
 274         }
 275
 276         update_share_count(sc, 0, newref->count);
 277         preftree->count++;
 278         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 279         rb_link_node(&newref->rbnode, parent, p);
 280         rb_insert_color_cached(&newref->rbnode, root, leftmost);
 281 }
 282
 283 /*
 284  * Release the entire tree.  We don't care about internal consistency so
 285  * just free everything and then reset the tree root.
 286  */
 287 static void prelim_release(struct preftree *preftree)
 288 {
 289         struct prelim_ref *ref, *next_ref;
 290
 291         rbtree_postorder_for_each_entry_safe(ref, next_ref,
 292                                              &preftree->root.rb_root, rbnode) {
 293                 free_inode_elem_list(ref->inode_list);
 294                 free_pref(ref);
 295         }
 296
 297         preftree->root = RB_ROOT_CACHED;
 298         preftree->count = 0;
 299 }
 300
 301 /*
 302  * the rules for all callers of this function are:
 303  * - obtaining the parent is the goal
 304  * - if you add a key, you must know that it is a correct key
 305  * - if you cannot add the parent or a correct key, then we will look into the
 306  *   block later to set a correct key
 307  *
 308  * delayed refs
 309  * ============
 310  *        backref type | shared | indirect | shared | indirect
 311  * information         |   tree |     tree |   data |     data
 312  * --------------------+--------+----------+--------+----------
 313  *      parent logical |    y   |     -    |    -   |     -
 314  *      key to resolve |    -   |     y    |    y   |     y
 315  *  tree block logical |    -   |     -    |    -   |     -
 316  *  root for resolving |    y   |     y    |    y   |     y
 317  *
 318  * - column 1:       we've the parent -> done
 319  * - column 2, 3, 4: we use the key to find the parent
 320  *
 321  * on disk refs (inline or keyed)
 322  * ==============================
 323  *        backref type | shared | indirect | shared | indirect
 324  * information         |   tree |     tree |   data |     data
 325  * --------------------+--------+----------+--------+----------
 326  *      parent logical |    y   |     -    |    y   |     -
 327  *      key to resolve |    -   |     -    |    -   |     y
 328  *  tree block logical |    y   |     y    |    y   |     y
 329  *  root for resolving |    -   |     y    |    y   |     y
 330  *
 331  * - column 1, 3: we've the parent -> done
 332  * - column 2:    we take the first key from the block to find the parent
 333  *                (see add_missing_keys)
 334  * - column 4:    we use the key to find the parent
 335  *
 336  * additional information that's available but not required to find the parent
 337  * block might help in merging entries to gain some speed.
 338  */
 339 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 340                           struct preftree *preftree, u64 root_id,
 341                           const struct btrfs_key *key, int level, u64 parent,
 342                           u64 wanted_disk_byte, int count,
 343                           struct share_check *sc, gfp_t gfp_mask)
 344 {
 345         struct prelim_ref *ref;
 346
 347         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 348                 return 0;
 349
 350         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 351         if (!ref)
 352                 return -ENOMEM;
 353
 354         ref->root_id = root_id;
 355         if (key)
 356                 ref->key_for_search = *key;
 357         else
 358                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 359
 360         ref->inode_list = NULL;
 361         ref->level = level;
 362         ref->count = count;
 363         ref->parent = parent;
 364         ref->wanted_disk_byte = wanted_disk_byte;
 365         prelim_ref_insert(fs_info, preftree, ref, sc);
 366         return extent_is_shared(sc);
 367 }
 368
 369 /* direct refs use root == 0, key == NULL */
 370 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 371                           struct preftrees *preftrees, int level, u64 parent,
 372                           u64 wanted_disk_byte, int count,
 373                           struct share_check *sc, gfp_t gfp_mask)
 374 {
 375         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 376                               parent, wanted_disk_byte, count, sc, gfp_mask);
 377 }
 378
 379 /* indirect refs use parent == 0 */
 380 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 381                             struct preftrees *preftrees, u64 root_id,
 382                             const struct btrfs_key *key, int level,
 383                             u64 wanted_disk_byte, int count,
 384                             struct share_check *sc, gfp_t gfp_mask)
 385 {
 386         struct preftree *tree = &preftrees->indirect;
 387
 388         if (!key)
 389                 tree = &preftrees->indirect_missing_keys;
 390         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 391                               wanted_disk_byte, count, sc, gfp_mask);
 392 }
 393
 394 static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
 395 {
 396         struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
 397         struct rb_node *parent = NULL;
 398         struct prelim_ref *ref = NULL;
 399         struct prelim_ref target = {};
 400         int result;
 401
 402         target.parent = bytenr;
 403
 404         while (*p) {
 405                 parent = *p;
 406                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 407                 result = prelim_ref_compare(ref, &target);
 408
 409                 if (result < 0)
 410                         p = &(*p)->rb_left;
 411                 else if (result > 0)
 412                         p = &(*p)->rb_right;
 413                 else
 414                         return 1;
 415         }
 416         return 0;
 417 }
 418
 419 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 420                            struct ulist *parents,
 421                            struct preftrees *preftrees, struct prelim_ref *ref,
 422                            int level, u64 time_seq, const u64 *extent_item_pos,
 423                            bool ignore_offset)
 424 {
 425         int ret = 0;
 426         int slot;
 427         struct extent_buffer *eb;
 428         struct btrfs_key key;
 429         struct btrfs_key *key_for_search = &ref->key_for_search;
 430         struct btrfs_file_extent_item *fi;
 431         struct extent_inode_elem *eie = NULL, *old = NULL;
 432         u64 disk_byte;
 433         u64 wanted_disk_byte = ref->wanted_disk_byte;
 434         u64 count = 0;
 435         u64 data_offset;
 436
 437         if (level != 0) {
 438                 eb = path->nodes[level];
 439                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 440                 if (ret < 0)
 441                         return ret;
 442                 return 0;
 443         }
 444
 445         /*
 446          * 1. We normally enter this function with the path already pointing to
 447          *    the first item to check. But sometimes, we may enter it with
 448          *    slot == nritems.
 449          * 2. We are searching for normal backref but bytenr of this leaf
 450          *    matches shared data backref
 451          * 3. The leaf owner is not equal to the root we are searching
 452          *
 453          * For these cases, go to the next leaf before we continue.
 454          */
 455         eb = path->nodes[0];
 456         if (path->slots[0] >= btrfs_header_nritems(eb) ||
 457             is_shared_data_backref(preftrees, eb->start) ||
 458             ref->root_id != btrfs_header_owner(eb)) {
 459                 if (time_seq == BTRFS_SEQ_LAST)
 460                         ret = btrfs_next_leaf(root, path);
 461                 else
 462                         ret = btrfs_next_old_leaf(root, path, time_seq);
 463         }
 464
 465         while (!ret && count < ref->count) {
 466                 eb = path->nodes[0];
 467                 slot = path->slots[0];
 468
 469                 btrfs_item_key_to_cpu(eb, &key, slot);
 470
 471                 if (key.objectid != key_for_search->objectid ||
 472                     key.type != BTRFS_EXTENT_DATA_KEY)
 473                         break;
 474
 475                 /*
 476                  * We are searching for normal backref but bytenr of this leaf
 477                  * matches shared data backref, OR
 478                  * the leaf owner is not equal to the root we are searching for
 479                  */
 480                 if (slot == 0 &&
 481                     (is_shared_data_backref(preftrees, eb->start) ||
 482                      ref->root_id != btrfs_header_owner(eb))) {
 483                         if (time_seq == BTRFS_SEQ_LAST)
 484                                 ret = btrfs_next_leaf(root, path);
 485                         else
 486                                 ret = btrfs_next_old_leaf(root, path, time_seq);
 487                         continue;
 488                 }
 489                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 490                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 491                 data_offset = btrfs_file_extent_offset(eb, fi);
 492
 493                 if (disk_byte == wanted_disk_byte) {
 494                         eie = NULL;
 495                         old = NULL;
 496                         if (ref->key_for_search.offset == key.offset - data_offset)
 497                                 count++;
 498                         else
 499                                 goto next;
 500                         if (extent_item_pos) {
 501                                 ret = check_extent_in_eb(&key, eb, fi,
 502                                                 *extent_item_pos,
 503                                                 &eie, ignore_offset);
 504                                 if (ret < 0)
 505                                         break;
 506                         }
 507                         if (ret > 0)
 508                                 goto next;
 509                         ret = ulist_add_merge_ptr(parents, eb->start,
 510                                                   eie, (void **)&old, GFP_NOFS);
 511                         if (ret < 0)
 512                                 break;
 513                         if (!ret && extent_item_pos) {
 514                                 while (old->next)
 515                                         old = old->next;
 516                                 old->next = eie;
 517                         }
 518                         eie = NULL;
 519                 }
 520 next:
 521                 if (time_seq == BTRFS_SEQ_LAST)
 522                         ret = btrfs_next_item(root, path);
 523                 else
 524                         ret = btrfs_next_old_item(root, path, time_seq);
 525         }
 526
 527         if (ret > 0)
 528                 ret = 0;
 529         else if (ret < 0)
 530                 free_inode_elem_list(eie);
 531         return ret;
 532 }
 533
 534 /*
 535  * resolve an indirect backref in the form (root_id, key, level)
 536  * to a logical address
 537  */
 538 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 539                                 struct btrfs_path *path, u64 time_seq,
 540                                 struct preftrees *preftrees,
 541                                 struct prelim_ref *ref, struct ulist *parents,
 542                                 const u64 *extent_item_pos, bool ignore_offset)
 543 {
 544         struct btrfs_root *root;
 545         struct extent_buffer *eb;
 546         int ret = 0;
 547         int root_level;
 548         int level = ref->level;
 549         struct btrfs_key search_key = ref->key_for_search;
 550
 551         /*
 552          * If we're search_commit_root we could possibly be holding locks on
 553          * other tree nodes.  This happens when qgroups does backref walks when
 554          * adding new delayed refs.  To deal with this we need to look in cache
 555          * for the root, and if we don't find it then we need to search the
 556          * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
 557          * here.
 558          */
 559         if (path->search_commit_root)
 560                 root = btrfs_get_fs_root_commit_root(fs_info, path, ref->root_id);
 561         else
 562                 root = btrfs_get_fs_root(fs_info, ref->root_id, false);
 563         if (IS_ERR(root)) {
 564                 ret = PTR_ERR(root);
 565                 goto out_free;
 566         }
 567
 568         if (!path->search_commit_root &&
 569             test_bit(BTRFS_ROOT_DELETING, &root->state)) {
 570                 ret = -ENOENT;
 571                 goto out;
 572         }
 573
 574         if (btrfs_is_testing(fs_info)) {
 575                 ret = -ENOENT;
 576                 goto out;
 577         }
 578
 579         if (path->search_commit_root)
 580                 root_level = btrfs_header_level(root->commit_root);
 581         else if (time_seq == BTRFS_SEQ_LAST)
 582                 root_level = btrfs_header_level(root->node);
 583         else
 584                 root_level = btrfs_old_root_level(root, time_seq);
 585
 586         if (root_level + 1 == level)
 587                 goto out;
 588
 589         /*
 590          * We can often find data backrefs with an offset that is too large
 591          * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
 592          * subtracting a file's offset with the data offset of its
 593          * corresponding extent data item. This can happen for example in the
 594          * clone ioctl.
 595          *
 596          * So if we detect such case we set the search key's offset to zero to
 597          * make sure we will find the matching file extent item at
 598          * add_all_parents(), otherwise we will miss it because the offset
 599          * taken form the backref is much larger then the offset of the file
 600          * extent item. This can make us scan a very large number of file
 601          * extent items, but at least it will not make us miss any.
 602          *
 603          * This is an ugly workaround for a behaviour that should have never
 604          * existed, but it does and a fix for the clone ioctl would touch a lot
 605          * of places, cause backwards incompatibility and would not fix the
 606          * problem for extents cloned with older kernels.
 607          */
 608         if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
 609             search_key.offset >= LLONG_MAX)
 610                 search_key.offset = 0;
 611         path->lowest_level = level;
 612         if (time_seq == BTRFS_SEQ_LAST)
 613                 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 614         else
 615                 ret = btrfs_search_old_slot(root, &search_key, path, time_seq);
 616
 617         btrfs_debug(fs_info,
 618                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 619                  ref->root_id, level, ref->count, ret,
 620                  ref->key_for_search.objectid, ref->key_for_search.type,
 621                  ref->key_for_search.offset);
 622         if (ret < 0)
 623                 goto out;
 624
 625         eb = path->nodes[level];
 626         while (!eb) {
 627                 if (WARN_ON(!level)) {
 628                         ret = 1;
 629                         goto out;
 630                 }
 631                 level--;
 632                 eb = path->nodes[level];
 633         }
 634
 635         ret = add_all_parents(root, path, parents, preftrees, ref, level,
 636                               time_seq, extent_item_pos, ignore_offset);
 637 out:
 638         btrfs_put_root(root);
 639 out_free:
 640         path->lowest_level = 0;
 641         btrfs_release_path(path);
 642         return ret;
 643 }
 644
 645 static struct extent_inode_elem *
 646 unode_aux_to_inode_list(struct ulist_node *node)
 647 {
 648         if (!node)
 649                 return NULL;
 650         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 651 }
 652
 653 static void free_leaf_list(struct ulist *ulist)
 654 {
 655         struct ulist_node *node;
 656         struct ulist_iterator uiter;
 657
 658         ULIST_ITER_INIT(&uiter);
 659         while ((node = ulist_next(ulist, &uiter)))
 660                 free_inode_elem_list(unode_aux_to_inode_list(node));
 661
 662         ulist_free(ulist);
 663 }
 664
 665 /*
 666  * We maintain three separate rbtrees: one for direct refs, one for
 667  * indirect refs which have a key, and one for indirect refs which do not
 668  * have a key. Each tree does merge on insertion.
 669  *
 670  * Once all of the references are located, we iterate over the tree of
 671  * indirect refs with missing keys. An appropriate key is located and
 672  * the ref is moved onto the tree for indirect refs. After all missing
 673  * keys are thus located, we iterate over the indirect ref tree, resolve
 674  * each reference, and then insert the resolved reference onto the
 675  * direct tree (merging there too).
 676  *
 677  * New backrefs (i.e., for parent nodes) are added to the appropriate
 678  * rbtree as they are encountered. The new backrefs are subsequently
 679  * resolved as above.
 680  */
 681 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 682                                  struct btrfs_path *path, u64 time_seq,
 683                                  struct preftrees *preftrees,
 684                                  const u64 *extent_item_pos,
 685                                  struct share_check *sc, bool ignore_offset)
 686 {
 687         int err;
 688         int ret = 0;
 689         struct ulist *parents;
 690         struct ulist_node *node;
 691         struct ulist_iterator uiter;
 692         struct rb_node *rnode;
 693
 694         parents = ulist_alloc(GFP_NOFS);
 695         if (!parents)
 696                 return -ENOMEM;
 697
 698         /*
 699          * We could trade memory usage for performance here by iterating
 700          * the tree, allocating new refs for each insertion, and then
 701          * freeing the entire indirect tree when we're done.  In some test
 702          * cases, the tree can grow quite large (~200k objects).
 703          */
 704         while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
 705                 struct prelim_ref *ref;
 706
 707                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 708                 if (WARN(ref->parent,
 709                          "BUG: direct ref found in indirect tree")) {
 710                         ret = -EINVAL;
 711                         goto out;
 712                 }
 713
 714                 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
 715                 preftrees->indirect.count--;
 716
 717                 if (ref->count == 0) {
 718                         free_pref(ref);
 719                         continue;
 720                 }
 721
 722                 if (sc && sc->root_objectid &&
 723                     ref->root_id != sc->root_objectid) {
 724                         free_pref(ref);
 725                         ret = BACKREF_FOUND_SHARED;
 726                         goto out;
 727                 }
 728                 err = resolve_indirect_ref(fs_info, path, time_seq, preftrees,
 729                                            ref, parents, extent_item_pos,
 730                                            ignore_offset);
 731                 /*
 732                  * we can only tolerate ENOENT,otherwise,we should catch error
 733                  * and return directly.
 734                  */
 735                 if (err == -ENOENT) {
 736                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 737                                           NULL);
 738                         continue;
 739                 } else if (err) {
 740                         free_pref(ref);
 741                         ret = err;
 742                         goto out;
 743                 }
 744
 745                 /* we put the first parent into the ref at hand */
 746                 ULIST_ITER_INIT(&uiter);
 747                 node = ulist_next(parents, &uiter);
 748                 ref->parent = node ? node->val : 0;
 749                 ref->inode_list = unode_aux_to_inode_list(node);
 750
 751                 /* Add a prelim_ref(s) for any other parent(s). */
 752                 while ((node = ulist_next(parents, &uiter))) {
 753                         struct prelim_ref *new_ref;
 754
 755                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 756                                                    GFP_NOFS);
 757                         if (!new_ref) {
 758                                 free_pref(ref);
 759                                 ret = -ENOMEM;
 760                                 goto out;
 761                         }
 762                         memcpy(new_ref, ref, sizeof(*ref));
 763                         new_ref->parent = node->val;
 764                         new_ref->inode_list = unode_aux_to_inode_list(node);
 765                         prelim_ref_insert(fs_info, &preftrees->direct,
 766                                           new_ref, NULL);
 767                 }
 768
 769                 /*
 770                  * Now it's a direct ref, put it in the direct tree. We must
 771                  * do this last because the ref could be merged/freed here.
 772                  */
 773                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 774
 775                 ulist_reinit(parents);
 776                 cond_resched();
 777         }
 778 out:
 779         /*
 780          * We may have inode lists attached to refs in the parents ulist, so we
 781          * must free them before freeing the ulist and its refs.
 782          */
 783         free_leaf_list(parents);
 784         return ret;
 785 }
 786
 787 /*
 788  * read tree blocks and add keys where required.
 789  */
 790 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 791                             struct preftrees *preftrees, bool lock)
 792 {
 793         struct prelim_ref *ref;
 794         struct extent_buffer *eb;
 795         struct preftree *tree = &preftrees->indirect_missing_keys;
 796         struct rb_node *node;
 797
 798         while ((node = rb_first_cached(&tree->root))) {
 799                 ref = rb_entry(node, struct prelim_ref, rbnode);
 800                 rb_erase_cached(node, &tree->root);
 801
 802                 BUG_ON(ref->parent);    /* should not be a direct ref */
 803                 BUG_ON(ref->key_for_search.type);
 804                 BUG_ON(!ref->wanted_disk_byte);
 805
 806                 eb = read_tree_block(fs_info, ref->wanted_disk_byte,
 807                                      ref->root_id, 0, ref->level - 1, NULL);
 808                 if (IS_ERR(eb)) {
 809                         free_pref(ref);
 810                         return PTR_ERR(eb);
 811                 } else if (!extent_buffer_uptodate(eb)) {
 812                         free_pref(ref);
 813                         free_extent_buffer(eb);
 814                         return -EIO;
 815                 }
 816                 if (lock)
 817                         btrfs_tree_read_lock(eb);
 818                 if (btrfs_header_level(eb) == 0)
 819                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 820                 else
 821                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 822                 if (lock)
 823                         btrfs_tree_read_unlock(eb);
 824                 free_extent_buffer(eb);
 825                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 826                 cond_resched();
 827         }
 828         return 0;
 829 }
 830
 831 /*
 832  * add all currently queued delayed refs from this head whose seq nr is
 833  * smaller or equal that seq to the list
 834  */
 835 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 836                             struct btrfs_delayed_ref_head *head, u64 seq,
 837                             struct preftrees *preftrees, struct share_check *sc)
 838 {
 839         struct btrfs_delayed_ref_node *node;
 840         struct btrfs_key key;
 841         struct rb_node *n;
 842         int count;
 843         int ret = 0;
 844
 845         spin_lock(&head->lock);
 846         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 847                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 848                                 ref_node);
 849                 if (node->seq > seq)
 850                         continue;
 851
 852                 switch (node->action) {
 853                 case BTRFS_ADD_DELAYED_EXTENT:
 854                 case BTRFS_UPDATE_DELAYED_HEAD:
 855                         WARN_ON(1);
 856                         continue;
 857                 case BTRFS_ADD_DELAYED_REF:
 858                         count = node->ref_mod;
 859                         break;
 860                 case BTRFS_DROP_DELAYED_REF:
 861                         count = node->ref_mod * -1;
 862                         break;
 863                 default:
 864                         BUG();
 865                 }
 866                 switch (node->type) {
 867                 case BTRFS_TREE_BLOCK_REF_KEY: {
 868                         /* NORMAL INDIRECT METADATA backref */
 869                         struct btrfs_delayed_tree_ref *ref;
 870                         struct btrfs_key *key_ptr = NULL;
 871
 872                         if (head->extent_op && head->extent_op->update_key) {
 873                                 btrfs_disk_key_to_cpu(&key, &head->extent_op->key);
 874                                 key_ptr = &key;
 875                         }
 876
 877                         ref = btrfs_delayed_node_to_tree_ref(node);
 878                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 879                                                key_ptr, ref->level + 1,
 880                                                node->bytenr, count, sc,
 881                                                GFP_ATOMIC);
 882                         break;
 883                 }
 884                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 885                         /* SHARED DIRECT METADATA backref */
 886                         struct btrfs_delayed_tree_ref *ref;
 887
 888                         ref = btrfs_delayed_node_to_tree_ref(node);
 889
 890                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 891                                              ref->parent, node->bytenr, count,
 892                                              sc, GFP_ATOMIC);
 893                         break;
 894                 }
 895                 case BTRFS_EXTENT_DATA_REF_KEY: {
 896                         /* NORMAL INDIRECT DATA backref */
 897                         struct btrfs_delayed_data_ref *ref;
 898                         ref = btrfs_delayed_node_to_data_ref(node);
 899
 900                         key.objectid = ref->objectid;
 901                         key.type = BTRFS_EXTENT_DATA_KEY;
 902                         key.offset = ref->offset;
 903
 904                         /*
 905                          * If we have a share check context and a reference for
 906                          * another inode, we can't exit immediately. This is
 907                          * because even if this is a BTRFS_ADD_DELAYED_REF
 908                          * reference we may find next a BTRFS_DROP_DELAYED_REF
 909                          * which cancels out this ADD reference.
 910                          *
 911                          * If this is a DROP reference and there was no previous
 912                          * ADD reference, then we need to signal that when we
 913                          * process references from the extent tree (through
 914                          * add_inline_refs() and add_keyed_refs()), we should
 915                          * not exit early if we find a reference for another
 916                          * inode, because one of the delayed DROP references
 917                          * may cancel that reference in the extent tree.
 918                          */
 919                         if (sc && count < 0)
 920                                 sc->have_delayed_delete_refs = true;
 921
 922                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 923                                                &key, 0, node->bytenr, count, sc,
 924                                                GFP_ATOMIC);
 925                         break;
 926                 }
 927                 case BTRFS_SHARED_DATA_REF_KEY: {
 928                         /* SHARED DIRECT FULL backref */
 929                         struct btrfs_delayed_data_ref *ref;
 930
 931                         ref = btrfs_delayed_node_to_data_ref(node);
 932
 933                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 934                                              node->bytenr, count, sc,
 935                                              GFP_ATOMIC);
 936                         break;
 937                 }
 938                 default:
 939                         WARN_ON(1);
 940                 }
 941                 /*
 942                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 943                  * refs have been checked.
 944                  */
 945                 if (ret && (ret != BACKREF_FOUND_SHARED))
 946                         break;
 947         }
 948         if (!ret)
 949                 ret = extent_is_shared(sc);
 950
 951         spin_unlock(&head->lock);
 952         return ret;
 953 }
 954
 955 /*
 956  * add all inline backrefs for bytenr to the list
 957  *
 958  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 959  */
 960 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 961                            struct btrfs_path *path, u64 bytenr,
 962                            int *info_level, struct preftrees *preftrees,
 963                            struct share_check *sc)
 964 {
 965         int ret = 0;
 966         int slot;
 967         struct extent_buffer *leaf;
 968         struct btrfs_key key;
 969         struct btrfs_key found_key;
 970         unsigned long ptr;
 971         unsigned long end;
 972         struct btrfs_extent_item *ei;
 973         u64 flags;
 974         u64 item_size;
 975
 976         /*
 977          * enumerate all inline refs
 978          */
 979         leaf = path->nodes[0];
 980         slot = path->slots[0];
 981
 982         item_size = btrfs_item_size_nr(leaf, slot);
 983         BUG_ON(item_size < sizeof(*ei));
 984
 985         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 986         flags = btrfs_extent_flags(leaf, ei);
 987         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 988
 989         ptr = (unsigned long)(ei + 1);
 990         end = (unsigned long)ei + item_size;
 991
 992         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 993             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 994                 struct btrfs_tree_block_info *info;
 995
 996                 info = (struct btrfs_tree_block_info *)ptr;
 997                 *info_level = btrfs_tree_block_level(leaf, info);
 998                 ptr += sizeof(struct btrfs_tree_block_info);
 999                 BUG_ON(ptr > end);
1000         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1001                 *info_level = found_key.offset;
1002         } else {
1003                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1004         }
1005
1006         while (ptr < end) {
1007                 struct btrfs_extent_inline_ref *iref;
1008                 u64 offset;
1009                 int type;
1010
1011                 iref = (struct btrfs_extent_inline_ref *)ptr;
1012                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
1013                                                         BTRFS_REF_TYPE_ANY);
1014                 if (type == BTRFS_REF_TYPE_INVALID)
1015                         return -EUCLEAN;
1016
1017                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1018
1019                 switch (type) {
1020                 case BTRFS_SHARED_BLOCK_REF_KEY:
1021                         ret = add_direct_ref(fs_info, preftrees,
1022                                              *info_level + 1, offset,
1023                                              bytenr, 1, NULL, GFP_NOFS);
1024                         break;
1025                 case BTRFS_SHARED_DATA_REF_KEY: {
1026                         struct btrfs_shared_data_ref *sdref;
1027                         int count;
1028
1029                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1030                         count = btrfs_shared_data_ref_count(leaf, sdref);
1031
1032                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
1033                                              bytenr, count, sc, GFP_NOFS);
1034                         break;
1035                 }
1036                 case BTRFS_TREE_BLOCK_REF_KEY:
1037                         ret = add_indirect_ref(fs_info, preftrees, offset,
1038                                                NULL, *info_level + 1,
1039                                                bytenr, 1, NULL, GFP_NOFS);
1040                         break;
1041                 case BTRFS_EXTENT_DATA_REF_KEY: {
1042                         struct btrfs_extent_data_ref *dref;
1043                         int count;
1044                         u64 root;
1045
1046                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1047                         count = btrfs_extent_data_ref_count(leaf, dref);
1048                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1049                                                                       dref);
1050                         key.type = BTRFS_EXTENT_DATA_KEY;
1051                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1052
1053                         if (sc && sc->inum && key.objectid != sc->inum &&
1054                             !sc->have_delayed_delete_refs) {
1055                                 ret = BACKREF_FOUND_SHARED;
1056                                 break;
1057                         }
1058
1059                         root = btrfs_extent_data_ref_root(leaf, dref);
1060
1061                         ret = add_indirect_ref(fs_info, preftrees, root,
1062                                                &key, 0, bytenr, count,
1063                                                sc, GFP_NOFS);
1064
1065                         break;
1066                 }
1067                 default:
1068                         WARN_ON(1);
1069                 }
1070                 if (ret)
1071                         return ret;
1072                 ptr += btrfs_extent_inline_ref_size(type);
1073         }
1074
1075         return 0;
1076 }
1077
1078 /*
1079  * add all non-inline backrefs for bytenr to the list
1080  *
1081  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1082  */
1083 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1084                           struct btrfs_path *path, u64 bytenr,
1085                           int info_level, struct preftrees *preftrees,
1086                           struct share_check *sc)
1087 {
1088         struct btrfs_root *extent_root = fs_info->extent_root;
1089         int ret;
1090         int slot;
1091         struct extent_buffer *leaf;
1092         struct btrfs_key key;
1093
1094         while (1) {
1095                 ret = btrfs_next_item(extent_root, path);
1096                 if (ret < 0)
1097                         break;
1098                 if (ret) {
1099                         ret = 0;
1100                         break;
1101                 }
1102
1103                 slot = path->slots[0];
1104                 leaf = path->nodes[0];
1105                 btrfs_item_key_to_cpu(leaf, &key, slot);
1106
1107                 if (key.objectid != bytenr)
1108                         break;
1109                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1110                         continue;
1111                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1112                         break;
1113
1114                 switch (key.type) {
1115                 case BTRFS_SHARED_BLOCK_REF_KEY:
1116                         /* SHARED DIRECT METADATA backref */
1117                         ret = add_direct_ref(fs_info, preftrees,
1118                                              info_level + 1, key.offset,
1119                                              bytenr, 1, NULL, GFP_NOFS);
1120                         break;
1121                 case BTRFS_SHARED_DATA_REF_KEY: {
1122                         /* SHARED DIRECT FULL backref */
1123                         struct btrfs_shared_data_ref *sdref;
1124                         int count;
1125
1126                         sdref = btrfs_item_ptr(leaf, slot,
1127                                               struct btrfs_shared_data_ref);
1128                         count = btrfs_shared_data_ref_count(leaf, sdref);
1129                         ret = add_direct_ref(fs_info, preftrees, 0,
1130                                              key.offset, bytenr, count,
1131                                              sc, GFP_NOFS);
1132                         break;
1133                 }
1134                 case BTRFS_TREE_BLOCK_REF_KEY:
1135                         /* NORMAL INDIRECT METADATA backref */
1136                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1137                                                NULL, info_level + 1, bytenr,
1138                                                1, NULL, GFP_NOFS);
1139                         break;
1140                 case BTRFS_EXTENT_DATA_REF_KEY: {
1141                         /* NORMAL INDIRECT DATA backref */
1142                         struct btrfs_extent_data_ref *dref;
1143                         int count;
1144                         u64 root;
1145
1146                         dref = btrfs_item_ptr(leaf, slot,
1147                                               struct btrfs_extent_data_ref);
1148                         count = btrfs_extent_data_ref_count(leaf, dref);
1149                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1150                                                                       dref);
1151                         key.type = BTRFS_EXTENT_DATA_KEY;
1152                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1153
1154                         if (sc && sc->inum && key.objectid != sc->inum &&
1155                             !sc->have_delayed_delete_refs) {
1156                                 ret = BACKREF_FOUND_SHARED;
1157                                 break;
1158                         }
1159
1160                         root = btrfs_extent_data_ref_root(leaf, dref);
1161                         ret = add_indirect_ref(fs_info, preftrees, root,
1162                                                &key, 0, bytenr, count,
1163                                                sc, GFP_NOFS);
1164                         break;
1165                 }
1166                 default:
1167                         WARN_ON(1);
1168                 }
1169                 if (ret)
1170                         return ret;
1171
1172         }
1173
1174         return ret;
1175 }
1176
1177 /*
1178  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1179  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1180  * indirect refs to their parent bytenr.
1181  * When roots are found, they're added to the roots list
1182  *
1183  * If time_seq is set to BTRFS_SEQ_LAST, it will not search delayed_refs, and
1184  * behave much like trans == NULL case, the difference only lies in it will not
1185  * commit root.
1186  * The special case is for qgroup to search roots in commit_transaction().
1187  *
1188  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1189  * shared extent is detected.
1190  *
1191  * Otherwise this returns 0 for success and <0 for an error.
1192  *
1193  * If ignore_offset is set to false, only extent refs whose offsets match
1194  * extent_item_pos are returned.  If true, every extent ref is returned
1195  * and extent_item_pos is ignored.
1196  *
1197  * FIXME some caching might speed things up
1198  */
1199 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1200                              struct btrfs_fs_info *fs_info, u64 bytenr,
1201                              u64 time_seq, struct ulist *refs,
1202                              struct ulist *roots, const u64 *extent_item_pos,
1203                              struct share_check *sc, bool ignore_offset)
1204 {
1205         struct btrfs_key key;
1206         struct btrfs_path *path;
1207         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1208         struct btrfs_delayed_ref_head *head;
1209         int info_level = 0;
1210         int ret;
1211         struct prelim_ref *ref;
1212         struct rb_node *node;
1213         struct extent_inode_elem *eie = NULL;
1214         struct preftrees preftrees = {
1215                 .direct = PREFTREE_INIT,
1216                 .indirect = PREFTREE_INIT,
1217                 .indirect_missing_keys = PREFTREE_INIT
1218         };
1219
1220         key.objectid = bytenr;
1221         key.offset = (u64)-1;
1222         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1223                 key.type = BTRFS_METADATA_ITEM_KEY;
1224         else
1225                 key.type = BTRFS_EXTENT_ITEM_KEY;
1226
1227         path = btrfs_alloc_path();
1228         if (!path)
1229                 return -ENOMEM;
1230         if (!trans) {
1231                 path->search_commit_root = 1;
1232                 path->skip_locking = 1;
1233         }
1234
1235         if (time_seq == BTRFS_SEQ_LAST)
1236                 path->skip_locking = 1;
1237
1238         /*
1239          * grab both a lock on the path and a lock on the delayed ref head.
1240          * We need both to get a consistent picture of how the refs look
1241          * at a specified point in time
1242          */
1243 again:
1244         head = NULL;
1245
1246         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1247         if (ret < 0)
1248                 goto out;
1249         if (ret == 0) {
1250                 /* This shouldn't happen, indicates a bug or fs corruption. */
1251                 ASSERT(ret != 0);
1252                 ret = -EUCLEAN;
1253                 goto out;
1254         }
1255
1256 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1257         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1258             time_seq != BTRFS_SEQ_LAST) {
1259 #else
1260         if (trans && time_seq != BTRFS_SEQ_LAST) {
1261 #endif
1262                 /*
1263                  * look if there are updates for this ref queued and lock the
1264                  * head
1265                  */
1266                 delayed_refs = &trans->transaction->delayed_refs;
1267                 spin_lock(&delayed_refs->lock);
1268                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1269                 if (head) {
1270                         if (!mutex_trylock(&head->mutex)) {
1271                                 refcount_inc(&head->refs);
1272                                 spin_unlock(&delayed_refs->lock);
1273
1274                                 btrfs_release_path(path);
1275
1276                                 /*
1277                                  * Mutex was contended, block until it's
1278                                  * released and try again
1279                                  */
1280                                 mutex_lock(&head->mutex);
1281                                 mutex_unlock(&head->mutex);
1282                                 btrfs_put_delayed_ref_head(head);
1283                                 goto again;
1284                         }
1285                         spin_unlock(&delayed_refs->lock);
1286                         ret = add_delayed_refs(fs_info, head, time_seq,
1287                                                &preftrees, sc);
1288                         mutex_unlock(&head->mutex);
1289                         if (ret)
1290                                 goto out;
1291                 } else {
1292                         spin_unlock(&delayed_refs->lock);
1293                 }
1294         }
1295
1296         if (path->slots[0]) {
1297                 struct extent_buffer *leaf;
1298                 int slot;
1299
1300                 path->slots[0]--;
1301                 leaf = path->nodes[0];
1302                 slot = path->slots[0];
1303                 btrfs_item_key_to_cpu(leaf, &key, slot);
1304                 if (key.objectid == bytenr &&
1305                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1306                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1307                         ret = add_inline_refs(fs_info, path, bytenr,
1308                                               &info_level, &preftrees, sc);
1309                         if (ret)
1310                                 goto out;
1311                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1312                                              &preftrees, sc);
1313                         if (ret)
1314                                 goto out;
1315                 }
1316         }
1317
1318         btrfs_release_path(path);
1319
1320         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1321         if (ret)
1322                 goto out;
1323
1324         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1325
1326         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1327                                     extent_item_pos, sc, ignore_offset);
1328         if (ret)
1329                 goto out;
1330
1331         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1332
1333         /*
1334          * This walks the tree of merged and resolved refs. Tree blocks are
1335          * read in as needed. Unique entries are added to the ulist, and
1336          * the list of found roots is updated.
1337          *
1338          * We release the entire tree in one go before returning.
1339          */
1340         node = rb_first_cached(&preftrees.direct.root);
1341         while (node) {
1342                 ref = rb_entry(node, struct prelim_ref, rbnode);
1343                 node = rb_next(&ref->rbnode);
1344                 /*
1345                  * ref->count < 0 can happen here if there are delayed
1346                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1347                  * prelim_ref_insert() relies on this when merging
1348                  * identical refs to keep the overall count correct.
1349                  * prelim_ref_insert() will merge only those refs
1350                  * which compare identically.  Any refs having
1351                  * e.g. different offsets would not be merged,
1352                  * and would retain their original ref->count < 0.
1353                  */
1354                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1355                         if (sc && sc->root_objectid &&
1356                             ref->root_id != sc->root_objectid) {
1357                                 ret = BACKREF_FOUND_SHARED;
1358                                 goto out;
1359                         }
1360
1361                         /* no parent == root of tree */
1362                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1363                         if (ret < 0)
1364                                 goto out;
1365                 }
1366                 if (ref->count && ref->parent) {
1367                         if (extent_item_pos && !ref->inode_list &&
1368                             ref->level == 0) {
1369                                 struct extent_buffer *eb;
1370
1371                                 eb = read_tree_block(fs_info, ref->parent, 0,
1372                                                      0, ref->level, NULL);
1373                                 if (IS_ERR(eb)) {
1374                                         ret = PTR_ERR(eb);
1375                                         goto out;
1376                                 } else if (!extent_buffer_uptodate(eb)) {
1377                                         free_extent_buffer(eb);
1378                                         ret = -EIO;
1379                                         goto out;
1380                                 }
1381
1382                                 if (!path->skip_locking)
1383                                         btrfs_tree_read_lock(eb);
1384                                 ret = find_extent_in_eb(eb, bytenr,
1385                                                         *extent_item_pos, &eie, ignore_offset);
1386                                 if (!path->skip_locking)
1387                                         btrfs_tree_read_unlock(eb);
1388                                 free_extent_buffer(eb);
1389                                 if (ret < 0)
1390                                         goto out;
1391                                 ref->inode_list = eie;
1392                                 /*
1393                                  * We transferred the list ownership to the ref,
1394                                  * so set to NULL to avoid a double free in case
1395                                  * an error happens after this.
1396                                  */
1397                                 eie = NULL;
1398                         }
1399                         ret = ulist_add_merge_ptr(refs, ref->parent,
1400                                                   ref->inode_list,
1401                                                   (void **)&eie, GFP_NOFS);
1402                         if (ret < 0)
1403                                 goto out;
1404                         if (!ret && extent_item_pos) {
1405                                 /*
1406                                  * We've recorded that parent, so we must extend
1407                                  * its inode list here.
1408                                  *
1409                                  * However if there was corruption we may not
1410                                  * have found an eie, return an error in this
1411                                  * case.
1412                                  */
1413                                 ASSERT(eie);
1414                                 if (!eie) {
1415                                         ret = -EUCLEAN;
1416                                         goto out;
1417                                 }
1418                                 while (eie->next)
1419                                         eie = eie->next;
1420                                 eie->next = ref->inode_list;
1421                         }
1422                         eie = NULL;
1423                         /*
1424                          * We have transferred the inode list ownership from
1425                          * this ref to the ref we added to the 'refs' ulist.
1426                          * So set this ref's inode list to NULL to avoid
1427                          * use-after-free when our caller uses it or double
1428                          * frees in case an error happens before we return.
1429                          */
1430                         ref->inode_list = NULL;
1431                 }
1432                 cond_resched();
1433         }
1434
1435 out:
1436         btrfs_free_path(path);
1437
1438         prelim_release(&preftrees.direct);
1439         prelim_release(&preftrees.indirect);
1440         prelim_release(&preftrees.indirect_missing_keys);
1441
1442         if (ret < 0)
1443                 free_inode_elem_list(eie);
1444         return ret;
1445 }
1446
1447 /*
1448  * Finds all leafs with a reference to the specified combination of bytenr and
1449  * offset. key_list_head will point to a list of corresponding keys (caller must
1450  * free each list element). The leafs will be stored in the leafs ulist, which
1451  * must be freed with ulist_free.
1452  *
1453  * returns 0 on success, <0 on error
1454  */
1455 int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1456                          struct btrfs_fs_info *fs_info, u64 bytenr,
1457                          u64 time_seq, struct ulist **leafs,
1458                          const u64 *extent_item_pos, bool ignore_offset)
1459 {
1460         int ret;
1461
1462         *leafs = ulist_alloc(GFP_NOFS);
1463         if (!*leafs)
1464                 return -ENOMEM;
1465
1466         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1467                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1468         if (ret < 0 && ret != -ENOENT) {
1469                 free_leaf_list(*leafs);
1470                 return ret;
1471         }
1472
1473         return 0;
1474 }
1475
1476 /*
1477  * walk all backrefs for a given extent to find all roots that reference this
1478  * extent. Walking a backref means finding all extents that reference this
1479  * extent and in turn walk the backrefs of those, too. Naturally this is a
1480  * recursive process, but here it is implemented in an iterative fashion: We
1481  * find all referencing extents for the extent in question and put them on a
1482  * list. In turn, we find all referencing extents for those, further appending
1483  * to the list. The way we iterate the list allows adding more elements after
1484  * the current while iterating. The process stops when we reach the end of the
1485  * list. Found roots are added to the roots list.
1486  *
1487  * returns 0 on success, < 0 on error.
1488  */
1489 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1490                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1491                                      u64 time_seq, struct ulist **roots,
1492                                      bool ignore_offset)
1493 {
1494         struct ulist *tmp;
1495         struct ulist_node *node = NULL;
1496         struct ulist_iterator uiter;
1497         int ret;
1498
1499         tmp = ulist_alloc(GFP_NOFS);
1500         if (!tmp)
1501                 return -ENOMEM;
1502         *roots = ulist_alloc(GFP_NOFS);
1503         if (!*roots) {
1504                 ulist_free(tmp);
1505                 return -ENOMEM;
1506         }
1507
1508         ULIST_ITER_INIT(&uiter);
1509         while (1) {
1510                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1511                                         tmp, *roots, NULL, NULL, ignore_offset);
1512                 if (ret < 0 && ret != -ENOENT) {
1513                         ulist_free(tmp);
1514                         ulist_free(*roots);
1515                         *roots = NULL;
1516                         return ret;
1517                 }
1518                 node = ulist_next(tmp, &uiter);
1519                 if (!node)
1520                         break;
1521                 bytenr = node->val;
1522                 cond_resched();
1523         }
1524
1525         ulist_free(tmp);
1526         return 0;
1527 }
1528
1529 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1530                          struct btrfs_fs_info *fs_info, u64 bytenr,
1531                          u64 time_seq, struct ulist **roots,
1532                          bool skip_commit_root_sem)
1533 {
1534         int ret;
1535
1536         if (!trans && !skip_commit_root_sem)
1537                 down_read(&fs_info->commit_root_sem);
1538         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1539                                         time_seq, roots, false);
1540         if (!trans && !skip_commit_root_sem)
1541                 up_read(&fs_info->commit_root_sem);
1542         return ret;
1543 }
1544
1545 /**
1546  * Check if an extent is shared or not
1547  *
1548  * @root:   root inode belongs to
1549  * @inum:   inode number of the inode whose extent we are checking
1550  * @bytenr: logical bytenr of the extent we are checking
1551  * @roots:  list of roots this extent is shared among
1552  * @tmp:    temporary list used for iteration
1553  *
1554  * btrfs_check_shared uses the backref walking code but will short
1555  * circuit as soon as it finds a root or inode that doesn't match the
1556  * one passed in. This provides a significant performance benefit for
1557  * callers (such as fiemap) which want to know whether the extent is
1558  * shared but do not need a ref count.
1559  *
1560  * This attempts to attach to the running transaction in order to account for
1561  * delayed refs, but continues on even when no running transaction exists.
1562  *
1563  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1564  */
1565 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr,
1566                 struct ulist *roots, struct ulist *tmp)
1567 {
1568         struct btrfs_fs_info *fs_info = root->fs_info;
1569         struct btrfs_trans_handle *trans;
1570         struct ulist_iterator uiter;
1571         struct ulist_node *node;
1572         struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1573         int ret = 0;
1574         struct share_check shared = {
1575                 .root_objectid = root->root_key.objectid,
1576                 .inum = inum,
1577                 .share_count = 0,
1578                 .have_delayed_delete_refs = false,
1579         };
1580
1581         ulist_init(roots);
1582         ulist_init(tmp);
1583
1584         trans = btrfs_join_transaction_nostart(root);
1585         if (IS_ERR(trans)) {
1586                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1587                         ret = PTR_ERR(trans);
1588                         goto out;
1589                 }
1590                 trans = NULL;
1591                 down_read(&fs_info->commit_root_sem);
1592         } else {
1593                 btrfs_get_tree_mod_seq(fs_info, &elem);
1594         }
1595
1596         ULIST_ITER_INIT(&uiter);
1597         while (1) {
1598                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1599                                         roots, NULL, &shared, false);
1600                 if (ret == BACKREF_FOUND_SHARED) {
1601                         /* this is the only condition under which we return 1 */
1602                         ret = 1;
1603                         break;
1604                 }
1605                 if (ret < 0 && ret != -ENOENT)
1606                         break;
1607                 ret = 0;
1608                 node = ulist_next(tmp, &uiter);
1609                 if (!node)
1610                         break;
1611                 bytenr = node->val;
1612                 shared.share_count = 0;
1613                 shared.have_delayed_delete_refs = false;
1614                 cond_resched();
1615         }
1616
1617         if (trans) {
1618                 btrfs_put_tree_mod_seq(fs_info, &elem);
1619                 btrfs_end_transaction(trans);
1620         } else {
1621                 up_read(&fs_info->commit_root_sem);
1622         }
1623 out:
1624         ulist_release(roots);
1625         ulist_release(tmp);
1626         return ret;
1627 }
1628
1629 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1630                           u64 start_off, struct btrfs_path *path,
1631                           struct btrfs_inode_extref **ret_extref,
1632                           u64 *found_off)
1633 {
1634         int ret, slot;
1635         struct btrfs_key key;
1636         struct btrfs_key found_key;
1637         struct btrfs_inode_extref *extref;
1638         const struct extent_buffer *leaf;
1639         unsigned long ptr;
1640
1641         key.objectid = inode_objectid;
1642         key.type = BTRFS_INODE_EXTREF_KEY;
1643         key.offset = start_off;
1644
1645         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1646         if (ret < 0)
1647                 return ret;
1648
1649         while (1) {
1650                 leaf = path->nodes[0];
1651                 slot = path->slots[0];
1652                 if (slot >= btrfs_header_nritems(leaf)) {
1653                         /*
1654                          * If the item at offset is not found,
1655                          * btrfs_search_slot will point us to the slot
1656                          * where it should be inserted. In our case
1657                          * that will be the slot directly before the
1658                          * next INODE_REF_KEY_V2 item. In the case
1659                          * that we're pointing to the last slot in a
1660                          * leaf, we must move one leaf over.
1661                          */
1662                         ret = btrfs_next_leaf(root, path);
1663                         if (ret) {
1664                                 if (ret >= 1)
1665                                         ret = -ENOENT;
1666                                 break;
1667                         }
1668                         continue;
1669                 }
1670
1671                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1672
1673                 /*
1674                  * Check that we're still looking at an extended ref key for
1675                  * this particular objectid. If we have different
1676                  * objectid or type then there are no more to be found
1677                  * in the tree and we can exit.
1678                  */
1679                 ret = -ENOENT;
1680                 if (found_key.objectid != inode_objectid)
1681                         break;
1682                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1683                         break;
1684
1685                 ret = 0;
1686                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1687                 extref = (struct btrfs_inode_extref *)ptr;
1688                 *ret_extref = extref;
1689                 if (found_off)
1690                         *found_off = found_key.offset;
1691                 break;
1692         }
1693
1694         return ret;
1695 }
1696
1697 /*
1698  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1699  * Elements of the path are separated by '/' and the path is guaranteed to be
1700  * 0-terminated. the path is only given within the current file system.
1701  * Therefore, it never starts with a '/'. the caller is responsible to provide
1702  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1703  * the start point of the resulting string is returned. this pointer is within
1704  * dest, normally.
1705  * in case the path buffer would overflow, the pointer is decremented further
1706  * as if output was written to the buffer, though no more output is actually
1707  * generated. that way, the caller can determine how much space would be
1708  * required for the path to fit into the buffer. in that case, the returned
1709  * value will be smaller than dest. callers must check this!
1710  */
1711 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1712                         u32 name_len, unsigned long name_off,
1713                         struct extent_buffer *eb_in, u64 parent,
1714                         char *dest, u32 size)
1715 {
1716         int slot;
1717         u64 next_inum;
1718         int ret;
1719         s64 bytes_left = ((s64)size) - 1;
1720         struct extent_buffer *eb = eb_in;
1721         struct btrfs_key found_key;
1722         struct btrfs_inode_ref *iref;
1723
1724         if (bytes_left >= 0)
1725                 dest[bytes_left] = '\0';
1726
1727         while (1) {
1728                 bytes_left -= name_len;
1729                 if (bytes_left >= 0)
1730                         read_extent_buffer(eb, dest + bytes_left,
1731                                            name_off, name_len);
1732                 if (eb != eb_in) {
1733                         if (!path->skip_locking)
1734                                 btrfs_tree_read_unlock(eb);
1735                         free_extent_buffer(eb);
1736                 }
1737                 ret = btrfs_find_item(fs_root, path, parent, 0,
1738                                 BTRFS_INODE_REF_KEY, &found_key);
1739                 if (ret > 0)
1740                         ret = -ENOENT;
1741                 if (ret)
1742                         break;
1743
1744                 next_inum = found_key.offset;
1745
1746                 /* regular exit ahead */
1747                 if (parent == next_inum)
1748                         break;
1749
1750                 slot = path->slots[0];
1751                 eb = path->nodes[0];
1752                 /* make sure we can use eb after releasing the path */
1753                 if (eb != eb_in) {
1754                         path->nodes[0] = NULL;
1755                         path->locks[0] = 0;
1756                 }
1757                 btrfs_release_path(path);
1758                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1759
1760                 name_len = btrfs_inode_ref_name_len(eb, iref);
1761                 name_off = (unsigned long)(iref + 1);
1762
1763                 parent = next_inum;
1764                 --bytes_left;
1765                 if (bytes_left >= 0)
1766                         dest[bytes_left] = '/';
1767         }
1768
1769         btrfs_release_path(path);
1770
1771         if (ret)
1772                 return ERR_PTR(ret);
1773
1774         return dest + bytes_left;
1775 }
1776
1777 /*
1778  * this makes the path point to (logical EXTENT_ITEM *)
1779  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1780  * tree blocks and <0 on error.
1781  */
1782 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1783                         struct btrfs_path *path, struct btrfs_key *found_key,
1784                         u64 *flags_ret)
1785 {
1786         int ret;
1787         u64 flags;
1788         u64 size = 0;
1789         u32 item_size;
1790         const struct extent_buffer *eb;
1791         struct btrfs_extent_item *ei;
1792         struct btrfs_key key;
1793
1794         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1795                 key.type = BTRFS_METADATA_ITEM_KEY;
1796         else
1797                 key.type = BTRFS_EXTENT_ITEM_KEY;
1798         key.objectid = logical;
1799         key.offset = (u64)-1;
1800
1801         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1802         if (ret < 0)
1803                 return ret;
1804
1805         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1806         if (ret) {
1807                 if (ret > 0)
1808                         ret = -ENOENT;
1809                 return ret;
1810         }
1811         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1812         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1813                 size = fs_info->nodesize;
1814         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1815                 size = found_key->offset;
1816
1817         if (found_key->objectid > logical ||
1818             found_key->objectid + size <= logical) {
1819                 btrfs_debug(fs_info,
1820                         "logical %llu is not within any extent", logical);
1821                 return -ENOENT;
1822         }
1823
1824         eb = path->nodes[0];
1825         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1826         BUG_ON(item_size < sizeof(*ei));
1827
1828         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1829         flags = btrfs_extent_flags(eb, ei);
1830
1831         btrfs_debug(fs_info,
1832                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1833                  logical, logical - found_key->objectid, found_key->objectid,
1834                  found_key->offset, flags, item_size);
1835
1836         WARN_ON(!flags_ret);
1837         if (flags_ret) {
1838                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1839                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1840                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1841                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1842                 else
1843                         BUG();
1844                 return 0;
1845         }
1846
1847         return -EIO;
1848 }
1849
1850 /*
1851  * helper function to iterate extent inline refs. ptr must point to a 0 value
1852  * for the first call and may be modified. it is used to track state.
1853  * if more refs exist, 0 is returned and the next call to
1854  * get_extent_inline_ref must pass the modified ptr parameter to get the
1855  * next ref. after the last ref was processed, 1 is returned.
1856  * returns <0 on error
1857  */
1858 static int get_extent_inline_ref(unsigned long *ptr,
1859                                  const struct extent_buffer *eb,
1860                                  const struct btrfs_key *key,
1861                                  const struct btrfs_extent_item *ei,
1862                                  u32 item_size,
1863                                  struct btrfs_extent_inline_ref **out_eiref,
1864                                  int *out_type)
1865 {
1866         unsigned long end;
1867         u64 flags;
1868         struct btrfs_tree_block_info *info;
1869
1870         if (!*ptr) {
1871                 /* first call */
1872                 flags = btrfs_extent_flags(eb, ei);
1873                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1874                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1875                                 /* a skinny metadata extent */
1876                                 *out_eiref =
1877                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1878                         } else {
1879                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1880                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1881                                 *out_eiref =
1882                                    (struct btrfs_extent_inline_ref *)(info + 1);
1883                         }
1884                 } else {
1885                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1886                 }
1887                 *ptr = (unsigned long)*out_eiref;
1888                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1889                         return -ENOENT;
1890         }
1891
1892         end = (unsigned long)ei + item_size;
1893         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1894         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1895                                                      BTRFS_REF_TYPE_ANY);
1896         if (*out_type == BTRFS_REF_TYPE_INVALID)
1897                 return -EUCLEAN;
1898
1899         *ptr += btrfs_extent_inline_ref_size(*out_type);
1900         WARN_ON(*ptr > end);
1901         if (*ptr == end)
1902                 return 1; /* last */
1903
1904         return 0;
1905 }
1906
1907 /*
1908  * reads the tree block backref for an extent. tree level and root are returned
1909  * through out_level and out_root. ptr must point to a 0 value for the first
1910  * call and may be modified (see get_extent_inline_ref comment).
1911  * returns 0 if data was provided, 1 if there was no more data to provide or
1912  * <0 on error.
1913  */
1914 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1915                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1916                             u32 item_size, u64 *out_root, u8 *out_level)
1917 {
1918         int ret;
1919         int type;
1920         struct btrfs_extent_inline_ref *eiref;
1921
1922         if (*ptr == (unsigned long)-1)
1923                 return 1;
1924
1925         while (1) {
1926                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1927                                               &eiref, &type);
1928                 if (ret < 0)
1929                         return ret;
1930
1931                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1932                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1933                         break;
1934
1935                 if (ret == 1)
1936                         return 1;
1937         }
1938
1939         /* we can treat both ref types equally here */
1940         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1941
1942         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1943                 struct btrfs_tree_block_info *info;
1944
1945                 info = (struct btrfs_tree_block_info *)(ei + 1);
1946                 *out_level = btrfs_tree_block_level(eb, info);
1947         } else {
1948                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1949                 *out_level = (u8)key->offset;
1950         }
1951
1952         if (ret == 1)
1953                 *ptr = (unsigned long)-1;
1954
1955         return 0;
1956 }
1957
1958 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1959                              struct extent_inode_elem *inode_list,
1960                              u64 root, u64 extent_item_objectid,
1961                              iterate_extent_inodes_t *iterate, void *ctx)
1962 {
1963         struct extent_inode_elem *eie;
1964         int ret = 0;
1965
1966         for (eie = inode_list; eie; eie = eie->next) {
1967                 btrfs_debug(fs_info,
1968                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1969                             extent_item_objectid, eie->inum,
1970                             eie->offset, root);
1971                 ret = iterate(eie->inum, eie->offset, root, ctx);
1972                 if (ret) {
1973                         btrfs_debug(fs_info,
1974                                     "stopping iteration for %llu due to ret=%d",
1975                                     extent_item_objectid, ret);
1976                         break;
1977                 }
1978         }
1979
1980         return ret;
1981 }
1982
1983 /*
1984  * calls iterate() for every inode that references the extent identified by
1985  * the given parameters.
1986  * when the iterator function returns a non-zero value, iteration stops.
1987  */
1988 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1989                                 u64 extent_item_objectid, u64 extent_item_pos,
1990                                 int search_commit_root,
1991                                 iterate_extent_inodes_t *iterate, void *ctx,
1992                                 bool ignore_offset)
1993 {
1994         int ret;
1995         struct btrfs_trans_handle *trans = NULL;
1996         struct ulist *refs = NULL;
1997         struct ulist *roots = NULL;
1998         struct ulist_node *ref_node = NULL;
1999         struct ulist_node *root_node = NULL;
2000         struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
2001         struct ulist_iterator ref_uiter;
2002         struct ulist_iterator root_uiter;
2003
2004         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
2005                         extent_item_objectid);
2006
2007         if (!search_commit_root) {
2008                 trans = btrfs_attach_transaction(fs_info->extent_root);
2009                 if (IS_ERR(trans)) {
2010                         if (PTR_ERR(trans) != -ENOENT &&
2011                             PTR_ERR(trans) != -EROFS)
2012                                 return PTR_ERR(trans);
2013                         trans = NULL;
2014                 }
2015         }
2016
2017         if (trans)
2018                 btrfs_get_tree_mod_seq(fs_info, &seq_elem);
2019         else
2020                 down_read(&fs_info->commit_root_sem);
2021
2022         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
2023                                    seq_elem.seq, &refs,
2024                                    &extent_item_pos, ignore_offset);
2025         if (ret)
2026                 goto out;
2027
2028         ULIST_ITER_INIT(&ref_uiter);
2029         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2030                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
2031                                                 seq_elem.seq, &roots,
2032                                                 ignore_offset);
2033                 if (ret)
2034                         break;
2035                 ULIST_ITER_INIT(&root_uiter);
2036                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
2037                         btrfs_debug(fs_info,
2038                                     "root %llu references leaf %llu, data list %#llx",
2039                                     root_node->val, ref_node->val,
2040                                     ref_node->aux);
2041                         ret = iterate_leaf_refs(fs_info,
2042                                                 (struct extent_inode_elem *)
2043                                                 (uintptr_t)ref_node->aux,
2044                                                 root_node->val,
2045                                                 extent_item_objectid,
2046                                                 iterate, ctx);
2047                 }
2048                 ulist_free(roots);
2049         }
2050
2051         free_leaf_list(refs);
2052 out:
2053         if (trans) {
2054                 btrfs_put_tree_mod_seq(fs_info, &seq_elem);
2055                 btrfs_end_transaction(trans);
2056         } else {
2057                 up_read(&fs_info->commit_root_sem);
2058         }
2059
2060         return ret;
2061 }
2062
2063 static int build_ino_list(u64 inum, u64 offset, u64 root, void *ctx)
2064 {
2065         struct btrfs_data_container *inodes = ctx;
2066         const size_t c = 3 * sizeof(u64);
2067
2068         if (inodes->bytes_left >= c) {
2069                 inodes->bytes_left -= c;
2070                 inodes->val[inodes->elem_cnt] = inum;
2071                 inodes->val[inodes->elem_cnt + 1] = offset;
2072                 inodes->val[inodes->elem_cnt + 2] = root;
2073                 inodes->elem_cnt += 3;
2074         } else {
2075                 inodes->bytes_missing += c - inodes->bytes_left;
2076                 inodes->bytes_left = 0;
2077                 inodes->elem_missed += 3;
2078         }
2079
2080         return 0;
2081 }
2082
2083 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2084                                 struct btrfs_path *path,
2085                                 void *ctx, bool ignore_offset)
2086 {
2087         int ret;
2088         u64 extent_item_pos;
2089         u64 flags = 0;
2090         struct btrfs_key found_key;
2091         int search_commit_root = path->search_commit_root;
2092
2093         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2094         btrfs_release_path(path);
2095         if (ret < 0)
2096                 return ret;
2097         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2098                 return -EINVAL;
2099
2100         extent_item_pos = logical - found_key.objectid;
2101         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2102                                         extent_item_pos, search_commit_root,
2103                                         build_ino_list, ctx, ignore_offset);
2104
2105         return ret;
2106 }
2107
2108 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2109                               struct extent_buffer *eb, void *ctx);
2110
2111 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2112                               struct btrfs_path *path,
2113                               iterate_irefs_t *iterate, void *ctx)
2114 {
2115         int ret = 0;
2116         int slot;
2117         u32 cur;
2118         u32 len;
2119         u32 name_len;
2120         u64 parent = 0;
2121         int found = 0;
2122         struct extent_buffer *eb;
2123         struct btrfs_item *item;
2124         struct btrfs_inode_ref *iref;
2125         struct btrfs_key found_key;
2126
2127         while (!ret) {
2128                 ret = btrfs_find_item(fs_root, path, inum,
2129                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2130                                 &found_key);
2131
2132                 if (ret < 0)
2133                         break;
2134                 if (ret) {
2135                         ret = found ? 0 : -ENOENT;
2136                         break;
2137                 }
2138                 ++found;
2139
2140                 parent = found_key.offset;
2141                 slot = path->slots[0];
2142                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2143                 if (!eb) {
2144                         ret = -ENOMEM;
2145                         break;
2146                 }
2147                 btrfs_release_path(path);
2148
2149                 item = btrfs_item_nr(slot);
2150                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2151
2152                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2153                         name_len = btrfs_inode_ref_name_len(eb, iref);
2154                         /* path must be released before calling iterate()! */
2155                         btrfs_debug(fs_root->fs_info,
2156                                 "following ref at offset %u for inode %llu in tree %llu",
2157                                 cur, found_key.objectid,
2158                                 fs_root->root_key.objectid);
2159                         ret = iterate(parent, name_len,
2160                                       (unsigned long)(iref + 1), eb, ctx);
2161                         if (ret)
2162                                 break;
2163                         len = sizeof(*iref) + name_len;
2164                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2165                 }
2166                 free_extent_buffer(eb);
2167         }
2168
2169         btrfs_release_path(path);
2170
2171         return ret;
2172 }
2173
2174 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2175                                  struct btrfs_path *path,
2176                                  iterate_irefs_t *iterate, void *ctx)
2177 {
2178         int ret;
2179         int slot;
2180         u64 offset = 0;
2181         u64 parent;
2182         int found = 0;
2183         struct extent_buffer *eb;
2184         struct btrfs_inode_extref *extref;
2185         u32 item_size;
2186         u32 cur_offset;
2187         unsigned long ptr;
2188
2189         while (1) {
2190                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2191                                             &offset);
2192                 if (ret < 0)
2193                         break;
2194                 if (ret) {
2195                         ret = found ? 0 : -ENOENT;
2196                         break;
2197                 }
2198                 ++found;
2199
2200                 slot = path->slots[0];
2201                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2202                 if (!eb) {
2203                         ret = -ENOMEM;
2204                         break;
2205                 }
2206                 btrfs_release_path(path);
2207
2208                 item_size = btrfs_item_size_nr(eb, slot);
2209                 ptr = btrfs_item_ptr_offset(eb, slot);
2210                 cur_offset = 0;
2211
2212                 while (cur_offset < item_size) {
2213                         u32 name_len;
2214
2215                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2216                         parent = btrfs_inode_extref_parent(eb, extref);
2217                         name_len = btrfs_inode_extref_name_len(eb, extref);
2218                         ret = iterate(parent, name_len,
2219                                       (unsigned long)&extref->name, eb, ctx);
2220                         if (ret)
2221                                 break;
2222
2223                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2224                         cur_offset += sizeof(*extref);
2225                 }
2226                 free_extent_buffer(eb);
2227
2228                 offset++;
2229         }
2230
2231         btrfs_release_path(path);
2232
2233         return ret;
2234 }
2235
2236 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2237                          struct btrfs_path *path, iterate_irefs_t *iterate,
2238                          void *ctx)
2239 {
2240         int ret;
2241         int found_refs = 0;
2242
2243         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2244         if (!ret)
2245                 ++found_refs;
2246         else if (ret != -ENOENT)
2247                 return ret;
2248
2249         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2250         if (ret == -ENOENT && found_refs)
2251                 return 0;
2252
2253         return ret;
2254 }
2255
2256 /*
2257  * returns 0 if the path could be dumped (probably truncated)
2258  * returns <0 in case of an error
2259  */
2260 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2261                          struct extent_buffer *eb, void *ctx)
2262 {
2263         struct inode_fs_paths *ipath = ctx;
2264         char *fspath;
2265         char *fspath_min;
2266         int i = ipath->fspath->elem_cnt;
2267         const int s_ptr = sizeof(char *);
2268         u32 bytes_left;
2269
2270         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2271                                         ipath->fspath->bytes_left - s_ptr : 0;
2272
2273         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2274         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2275                                    name_off, eb, inum, fspath_min, bytes_left);
2276         if (IS_ERR(fspath))
2277                 return PTR_ERR(fspath);
2278
2279         if (fspath > fspath_min) {
2280                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2281                 ++ipath->fspath->elem_cnt;
2282                 ipath->fspath->bytes_left = fspath - fspath_min;
2283         } else {
2284                 ++ipath->fspath->elem_missed;
2285                 ipath->fspath->bytes_missing += fspath_min - fspath;
2286                 ipath->fspath->bytes_left = 0;
2287         }
2288
2289         return 0;
2290 }
2291
2292 /*
2293  * this dumps all file system paths to the inode into the ipath struct, provided
2294  * is has been created large enough. each path is zero-terminated and accessed
2295  * from ipath->fspath->val[i].
2296  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2297  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2298  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2299  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2300  * have been needed to return all paths.
2301  */
2302 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2303 {
2304         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2305                              inode_to_path, ipath);
2306 }
2307
2308 struct btrfs_data_container *init_data_container(u32 total_bytes)
2309 {
2310         struct btrfs_data_container *data;
2311         size_t alloc_bytes;
2312
2313         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2314         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2315         if (!data)
2316                 return ERR_PTR(-ENOMEM);
2317
2318         if (total_bytes >= sizeof(*data)) {
2319                 data->bytes_left = total_bytes - sizeof(*data);
2320                 data->bytes_missing = 0;
2321         } else {
2322                 data->bytes_missing = sizeof(*data) - total_bytes;
2323                 data->bytes_left = 0;
2324         }
2325
2326         data->elem_cnt = 0;
2327         data->elem_missed = 0;
2328
2329         return data;
2330 }
2331
2332 /*
2333  * allocates space to return multiple file system paths for an inode.
2334  * total_bytes to allocate are passed, note that space usable for actual path
2335  * information will be total_bytes - sizeof(struct inode_fs_paths).
2336  * the returned pointer must be freed with free_ipath() in the end.
2337  */
2338 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2339                                         struct btrfs_path *path)
2340 {
2341         struct inode_fs_paths *ifp;
2342         struct btrfs_data_container *fspath;
2343
2344         fspath = init_data_container(total_bytes);
2345         if (IS_ERR(fspath))
2346                 return ERR_CAST(fspath);
2347
2348         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2349         if (!ifp) {
2350                 kvfree(fspath);
2351                 return ERR_PTR(-ENOMEM);
2352         }
2353
2354         ifp->btrfs_path = path;
2355         ifp->fspath = fspath;
2356         ifp->fs_root = fs_root;
2357
2358         return ifp;
2359 }
2360
2361 void free_ipath(struct inode_fs_paths *ipath)
2362 {
2363         if (!ipath)
2364                 return;
2365         kvfree(ipath->fspath);
2366         kfree(ipath);
2367 }
2368
2369 struct btrfs_backref_iter *btrfs_backref_iter_alloc(
2370                 struct btrfs_fs_info *fs_info, gfp_t gfp_flag)
2371 {
2372         struct btrfs_backref_iter *ret;
2373
2374         ret = kzalloc(sizeof(*ret), gfp_flag);
2375         if (!ret)
2376                 return NULL;
2377
2378         ret->path = btrfs_alloc_path();
2379         if (!ret->path) {
2380                 kfree(ret);
2381                 return NULL;
2382         }
2383
2384         /* Current backref iterator only supports iteration in commit root */
2385         ret->path->search_commit_root = 1;
2386         ret->path->skip_locking = 1;
2387         ret->fs_info = fs_info;
2388
2389         return ret;
2390 }
2391
2392 int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2393 {
2394         struct btrfs_fs_info *fs_info = iter->fs_info;
2395         struct btrfs_path *path = iter->path;
2396         struct btrfs_extent_item *ei;
2397         struct btrfs_key key;
2398         int ret;
2399
2400         key.objectid = bytenr;
2401         key.type = BTRFS_METADATA_ITEM_KEY;
2402         key.offset = (u64)-1;
2403         iter->bytenr = bytenr;
2404
2405         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
2406         if (ret < 0)
2407                 return ret;
2408         if (ret == 0) {
2409                 ret = -EUCLEAN;
2410                 goto release;
2411         }
2412         if (path->slots[0] == 0) {
2413                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2414                 ret = -EUCLEAN;
2415                 goto release;
2416         }
2417         path->slots[0]--;
2418
2419         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2420         if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2421              key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2422                 ret = -ENOENT;
2423                 goto release;
2424         }
2425         memcpy(&iter->cur_key, &key, sizeof(key));
2426         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2427                                                     path->slots[0]);
2428         iter->end_ptr = (u32)(iter->item_ptr +
2429                         btrfs_item_size_nr(path->nodes[0], path->slots[0]));
2430         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2431                             struct btrfs_extent_item);
2432
2433         /*
2434          * Only support iteration on tree backref yet.
2435          *
2436          * This is an extra precaution for non skinny-metadata, where
2437          * EXTENT_ITEM is also used for tree blocks, that we can only use
2438          * extent flags to determine if it's a tree block.
2439          */
2440         if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2441                 ret = -ENOTSUPP;
2442                 goto release;
2443         }
2444         iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2445
2446         /* If there is no inline backref, go search for keyed backref */
2447         if (iter->cur_ptr >= iter->end_ptr) {
2448                 ret = btrfs_next_item(fs_info->extent_root, path);
2449
2450                 /* No inline nor keyed ref */
2451                 if (ret > 0) {
2452                         ret = -ENOENT;
2453                         goto release;
2454                 }
2455                 if (ret < 0)
2456                         goto release;
2457
2458                 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2459                                 path->slots[0]);
2460                 if (iter->cur_key.objectid != bytenr ||
2461                     (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2462                      iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2463                         ret = -ENOENT;
2464                         goto release;
2465                 }
2466                 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2467                                                            path->slots[0]);
2468                 iter->item_ptr = iter->cur_ptr;
2469                 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size_nr(
2470                                       path->nodes[0], path->slots[0]));
2471         }
2472
2473         return 0;
2474 release:
2475         btrfs_backref_iter_release(iter);
2476         return ret;
2477 }
2478
2479 /*
2480  * Go to the next backref item of current bytenr, can be either inlined or
2481  * keyed.
2482  *
2483  * Caller needs to check whether it's inline ref or not by iter->cur_key.
2484  *
2485  * Return 0 if we get next backref without problem.
2486  * Return >0 if there is no extra backref for this bytenr.
2487  * Return <0 if there is something wrong happened.
2488  */
2489 int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2490 {
2491         struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2492         struct btrfs_path *path = iter->path;
2493         struct btrfs_extent_inline_ref *iref;
2494         int ret;
2495         u32 size;
2496
2497         if (btrfs_backref_iter_is_inline_ref(iter)) {
2498                 /* We're still inside the inline refs */
2499                 ASSERT(iter->cur_ptr < iter->end_ptr);
2500
2501                 if (btrfs_backref_has_tree_block_info(iter)) {
2502                         /* First tree block info */
2503                         size = sizeof(struct btrfs_tree_block_info);
2504                 } else {
2505                         /* Use inline ref type to determine the size */
2506                         int type;
2507
2508                         iref = (struct btrfs_extent_inline_ref *)
2509                                 ((unsigned long)iter->cur_ptr);
2510                         type = btrfs_extent_inline_ref_type(eb, iref);
2511
2512                         size = btrfs_extent_inline_ref_size(type);
2513                 }
2514                 iter->cur_ptr += size;
2515                 if (iter->cur_ptr < iter->end_ptr)
2516                         return 0;
2517
2518                 /* All inline items iterated, fall through */
2519         }
2520
2521         /* We're at keyed items, there is no inline item, go to the next one */
2522         ret = btrfs_next_item(iter->fs_info->extent_root, iter->path);
2523         if (ret)
2524                 return ret;
2525
2526         btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2527         if (iter->cur_key.objectid != iter->bytenr ||
2528             (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2529              iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2530                 return 1;
2531         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2532                                         path->slots[0]);
2533         iter->cur_ptr = iter->item_ptr;
2534         iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size_nr(path->nodes[0],
2535                                                 path->slots[0]);
2536         return 0;
2537 }
2538
2539 void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
2540                               struct btrfs_backref_cache *cache, int is_reloc)
2541 {
2542         int i;
2543
2544         cache->rb_root = RB_ROOT;
2545         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2546                 INIT_LIST_HEAD(&cache->pending[i]);
2547         INIT_LIST_HEAD(&cache->changed);
2548         INIT_LIST_HEAD(&cache->detached);
2549         INIT_LIST_HEAD(&cache->leaves);
2550         INIT_LIST_HEAD(&cache->pending_edge);
2551         INIT_LIST_HEAD(&cache->useless_node);
2552         cache->fs_info = fs_info;
2553         cache->is_reloc = is_reloc;
2554 }
2555
2556 struct btrfs_backref_node *btrfs_backref_alloc_node(
2557                 struct btrfs_backref_cache *cache, u64 bytenr, int level)
2558 {
2559         struct btrfs_backref_node *node;
2560
2561         ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
2562         node = kzalloc(sizeof(*node), GFP_NOFS);
2563         if (!node)
2564                 return node;
2565
2566         INIT_LIST_HEAD(&node->list);
2567         INIT_LIST_HEAD(&node->upper);
2568         INIT_LIST_HEAD(&node->lower);
2569         RB_CLEAR_NODE(&node->rb_node);
2570         cache->nr_nodes++;
2571         node->level = level;
2572         node->bytenr = bytenr;
2573
2574         return node;
2575 }
2576
2577 struct btrfs_backref_edge *btrfs_backref_alloc_edge(
2578                 struct btrfs_backref_cache *cache)
2579 {
2580         struct btrfs_backref_edge *edge;
2581
2582         edge = kzalloc(sizeof(*edge), GFP_NOFS);
2583         if (edge)
2584                 cache->nr_edges++;
2585         return edge;
2586 }
2587
2588 /*
2589  * Drop the backref node from cache, also cleaning up all its
2590  * upper edges and any uncached nodes in the path.
2591  *
2592  * This cleanup happens bottom up, thus the node should either
2593  * be the lowest node in the cache or a detached node.
2594  */
2595 void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
2596                                 struct btrfs_backref_node *node)
2597 {
2598         struct btrfs_backref_node *upper;
2599         struct btrfs_backref_edge *edge;
2600
2601         if (!node)
2602                 return;
2603
2604         BUG_ON(!node->lowest && !node->detached);
2605         while (!list_empty(&node->upper)) {
2606                 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
2607                                   list[LOWER]);
2608                 upper = edge->node[UPPER];
2609                 list_del(&edge->list[LOWER]);
2610                 list_del(&edge->list[UPPER]);
2611                 btrfs_backref_free_edge(cache, edge);
2612
2613                 /*
2614                  * Add the node to leaf node list if no other child block
2615                  * cached.
2616                  */
2617                 if (list_empty(&upper->lower)) {
2618                         list_add_tail(&upper->lower, &cache->leaves);
2619                         upper->lowest = 1;
2620                 }
2621         }
2622
2623         btrfs_backref_drop_node(cache, node);
2624 }
2625
2626 /*
2627  * Release all nodes/edges from current cache
2628  */
2629 void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
2630 {
2631         struct btrfs_backref_node *node;
2632         int i;
2633
2634         while (!list_empty(&cache->detached)) {
2635                 node = list_entry(cache->detached.next,
2636                                   struct btrfs_backref_node, list);
2637                 btrfs_backref_cleanup_node(cache, node);
2638         }
2639
2640         while (!list_empty(&cache->leaves)) {
2641                 node = list_entry(cache->leaves.next,
2642                                   struct btrfs_backref_node, lower);
2643                 btrfs_backref_cleanup_node(cache, node);
2644         }
2645
2646         cache->last_trans = 0;
2647
2648         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2649                 ASSERT(list_empty(&cache->pending[i]));
2650         ASSERT(list_empty(&cache->pending_edge));
2651         ASSERT(list_empty(&cache->useless_node));
2652         ASSERT(list_empty(&cache->changed));
2653         ASSERT(list_empty(&cache->detached));
2654         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
2655         ASSERT(!cache->nr_nodes);
2656         ASSERT(!cache->nr_edges);
2657 }
2658
2659 /*
2660  * Handle direct tree backref
2661  *
2662  * Direct tree backref means, the backref item shows its parent bytenr
2663  * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
2664  *
2665  * @ref_key:    The converted backref key.
2666  *              For keyed backref, it's the item key.
2667  *              For inlined backref, objectid is the bytenr,
2668  *              type is btrfs_inline_ref_type, offset is
2669  *              btrfs_inline_ref_offset.
2670  */
2671 static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
2672                                       struct btrfs_key *ref_key,
2673                                       struct btrfs_backref_node *cur)
2674 {
2675         struct btrfs_backref_edge *edge;
2676         struct btrfs_backref_node *upper;
2677         struct rb_node *rb_node;
2678
2679         ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
2680
2681         /* Only reloc root uses backref pointing to itself */
2682         if (ref_key->objectid == ref_key->offset) {
2683                 struct btrfs_root *root;
2684
2685                 cur->is_reloc_root = 1;
2686                 /* Only reloc backref cache cares about a specific root */
2687                 if (cache->is_reloc) {
2688                         root = find_reloc_root(cache->fs_info, cur->bytenr);
2689                         if (!root)
2690                                 return -ENOENT;
2691                         cur->root = root;
2692                 } else {
2693                         /*
2694                          * For generic purpose backref cache, reloc root node
2695                          * is useless.
2696                          */
2697                         list_add(&cur->list, &cache->useless_node);
2698                 }
2699                 return 0;
2700         }
2701
2702         edge = btrfs_backref_alloc_edge(cache);
2703         if (!edge)
2704                 return -ENOMEM;
2705
2706         rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
2707         if (!rb_node) {
2708                 /* Parent node not yet cached */
2709                 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
2710                                            cur->level + 1);
2711                 if (!upper) {
2712                         btrfs_backref_free_edge(cache, edge);
2713                         return -ENOMEM;
2714                 }
2715
2716                 /*
2717                  *  Backrefs for the upper level block isn't cached, add the
2718                  *  block to pending list
2719                  */
2720                 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2721         } else {
2722                 /* Parent node already cached */
2723                 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
2724                 ASSERT(upper->checked);
2725                 INIT_LIST_HEAD(&edge->list[UPPER]);
2726         }
2727         btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
2728         return 0;
2729 }
2730
2731 /*
2732  * Handle indirect tree backref
2733  *
2734  * Indirect tree backref means, we only know which tree the node belongs to.
2735  * We still need to do a tree search to find out the parents. This is for
2736  * TREE_BLOCK_REF backref (keyed or inlined).
2737  *
2738  * @ref_key:    The same as @ref_key in  handle_direct_tree_backref()
2739  * @tree_key:   The first key of this tree block.
2740  * @path:       A clean (released) path, to avoid allocating path every time
2741  *              the function get called.
2742  */
2743 static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache,
2744                                         struct btrfs_path *path,
2745                                         struct btrfs_key *ref_key,
2746                                         struct btrfs_key *tree_key,
2747                                         struct btrfs_backref_node *cur)
2748 {
2749         struct btrfs_fs_info *fs_info = cache->fs_info;
2750         struct btrfs_backref_node *upper;
2751         struct btrfs_backref_node *lower;
2752         struct btrfs_backref_edge *edge;
2753         struct extent_buffer *eb;
2754         struct btrfs_root *root;
2755         struct rb_node *rb_node;
2756         int level;
2757         bool need_check = true;
2758         int ret;
2759
2760         root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
2761         if (IS_ERR(root))
2762                 return PTR_ERR(root);
2763         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2764                 cur->cowonly = 1;
2765
2766         if (btrfs_root_level(&root->root_item) == cur->level) {
2767                 /* Tree root */
2768                 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
2769                 /*
2770                  * For reloc backref cache, we may ignore reloc root.  But for
2771                  * general purpose backref cache, we can't rely on
2772                  * btrfs_should_ignore_reloc_root() as it may conflict with
2773                  * current running relocation and lead to missing root.
2774                  *
2775                  * For general purpose backref cache, reloc root detection is
2776                  * completely relying on direct backref (key->offset is parent
2777                  * bytenr), thus only do such check for reloc cache.
2778                  */
2779                 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
2780                         btrfs_put_root(root);
2781                         list_add(&cur->list, &cache->useless_node);
2782                 } else {
2783                         cur->root = root;
2784                 }
2785                 return 0;
2786         }
2787
2788         level = cur->level + 1;
2789
2790         /* Search the tree to find parent blocks referring to the block */
2791         path->search_commit_root = 1;
2792         path->skip_locking = 1;
2793         path->lowest_level = level;
2794         ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
2795         path->lowest_level = 0;
2796         if (ret < 0) {
2797                 btrfs_put_root(root);
2798                 return ret;
2799         }
2800         if (ret > 0 && path->slots[level] > 0)
2801                 path->slots[level]--;
2802
2803         eb = path->nodes[level];
2804         if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
2805                 btrfs_err(fs_info,
2806 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
2807                           cur->bytenr, level - 1, root->root_key.objectid,
2808                           tree_key->objectid, tree_key->type, tree_key->offset);
2809                 btrfs_put_root(root);
2810                 ret = -ENOENT;
2811                 goto out;
2812         }
2813         lower = cur;
2814
2815         /* Add all nodes and edges in the path */
2816         for (; level < BTRFS_MAX_LEVEL; level++) {
2817                 if (!path->nodes[level]) {
2818                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
2819                                lower->bytenr);
2820                         /* Same as previous should_ignore_reloc_root() call */
2821                         if (btrfs_should_ignore_reloc_root(root) &&
2822                             cache->is_reloc) {
2823                                 btrfs_put_root(root);
2824                                 list_add(&lower->list, &cache->useless_node);
2825                         } else {
2826                                 lower->root = root;
2827                         }
2828                         break;
2829                 }
2830
2831                 edge = btrfs_backref_alloc_edge(cache);
2832                 if (!edge) {
2833                         btrfs_put_root(root);
2834                         ret = -ENOMEM;
2835                         goto out;
2836                 }
2837
2838                 eb = path->nodes[level];
2839                 rb_node = rb_simple_search(&cache->rb_root, eb->start);
2840                 if (!rb_node) {
2841                         upper = btrfs_backref_alloc_node(cache, eb->start,
2842                                                          lower->level + 1);
2843                         if (!upper) {
2844                                 btrfs_put_root(root);
2845                                 btrfs_backref_free_edge(cache, edge);
2846                                 ret = -ENOMEM;
2847                                 goto out;
2848                         }
2849                         upper->owner = btrfs_header_owner(eb);
2850                         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2851                                 upper->cowonly = 1;
2852
2853                         /*
2854                          * If we know the block isn't shared we can avoid
2855                          * checking its backrefs.
2856                          */
2857                         if (btrfs_block_can_be_shared(root, eb))
2858                                 upper->checked = 0;
2859                         else
2860                                 upper->checked = 1;
2861
2862                         /*
2863                          * Add the block to pending list if we need to check its
2864                          * backrefs, we only do this once while walking up a
2865                          * tree as we will catch anything else later on.
2866                          */
2867                         if (!upper->checked && need_check) {
2868                                 need_check = false;
2869                                 list_add_tail(&edge->list[UPPER],
2870                                               &cache->pending_edge);
2871                         } else {
2872                                 if (upper->checked)
2873                                         need_check = true;
2874                                 INIT_LIST_HEAD(&edge->list[UPPER]);
2875                         }
2876                 } else {
2877                         upper = rb_entry(rb_node, struct btrfs_backref_node,
2878                                          rb_node);
2879                         ASSERT(upper->checked);
2880                         INIT_LIST_HEAD(&edge->list[UPPER]);
2881                         if (!upper->owner)
2882                                 upper->owner = btrfs_header_owner(eb);
2883                 }
2884                 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
2885
2886                 if (rb_node) {
2887                         btrfs_put_root(root);
2888                         break;
2889                 }
2890                 lower = upper;
2891                 upper = NULL;
2892         }
2893 out:
2894         btrfs_release_path(path);
2895         return ret;
2896 }
2897
2898 /*
2899  * Add backref node @cur into @cache.
2900  *
2901  * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
2902  *       links aren't yet bi-directional. Needs to finish such links.
2903  *       Use btrfs_backref_finish_upper_links() to finish such linkage.
2904  *
2905  * @path:       Released path for indirect tree backref lookup
2906  * @iter:       Released backref iter for extent tree search
2907  * @node_key:   The first key of the tree block
2908  */
2909 int btrfs_backref_add_tree_node(struct btrfs_backref_cache *cache,
2910                                 struct btrfs_path *path,
2911                                 struct btrfs_backref_iter *iter,
2912                                 struct btrfs_key *node_key,
2913                                 struct btrfs_backref_node *cur)
2914 {
2915         struct btrfs_fs_info *fs_info = cache->fs_info;
2916         struct btrfs_backref_edge *edge;
2917         struct btrfs_backref_node *exist;
2918         int ret;
2919
2920         ret = btrfs_backref_iter_start(iter, cur->bytenr);
2921         if (ret < 0)
2922                 return ret;
2923         /*
2924          * We skip the first btrfs_tree_block_info, as we don't use the key
2925          * stored in it, but fetch it from the tree block
2926          */
2927         if (btrfs_backref_has_tree_block_info(iter)) {
2928                 ret = btrfs_backref_iter_next(iter);
2929                 if (ret < 0)
2930                         goto out;
2931                 /* No extra backref? This means the tree block is corrupted */
2932                 if (ret > 0) {
2933                         ret = -EUCLEAN;
2934                         goto out;
2935                 }
2936         }
2937         WARN_ON(cur->checked);
2938         if (!list_empty(&cur->upper)) {
2939                 /*
2940                  * The backref was added previously when processing backref of
2941                  * type BTRFS_TREE_BLOCK_REF_KEY
2942                  */
2943                 ASSERT(list_is_singular(&cur->upper));
2944                 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
2945                                   list[LOWER]);
2946                 ASSERT(list_empty(&edge->list[UPPER]));
2947                 exist = edge->node[UPPER];
2948                 /*
2949                  * Add the upper level block to pending list if we need check
2950                  * its backrefs
2951                  */
2952                 if (!exist->checked)
2953                         list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2954         } else {
2955                 exist = NULL;
2956         }
2957
2958         for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
2959                 struct extent_buffer *eb;
2960                 struct btrfs_key key;
2961                 int type;
2962
2963                 cond_resched();
2964                 eb = btrfs_backref_get_eb(iter);
2965
2966                 key.objectid = iter->bytenr;
2967                 if (btrfs_backref_iter_is_inline_ref(iter)) {
2968                         struct btrfs_extent_inline_ref *iref;
2969
2970                         /* Update key for inline backref */
2971                         iref = (struct btrfs_extent_inline_ref *)
2972                                 ((unsigned long)iter->cur_ptr);
2973                         type = btrfs_get_extent_inline_ref_type(eb, iref,
2974                                                         BTRFS_REF_TYPE_BLOCK);
2975                         if (type == BTRFS_REF_TYPE_INVALID) {
2976                                 ret = -EUCLEAN;
2977                                 goto out;
2978                         }
2979                         key.type = type;
2980                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
2981                 } else {
2982                         key.type = iter->cur_key.type;
2983                         key.offset = iter->cur_key.offset;
2984                 }
2985
2986                 /*
2987                  * Parent node found and matches current inline ref, no need to
2988                  * rebuild this node for this inline ref
2989                  */
2990                 if (exist &&
2991                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
2992                       exist->owner == key.offset) ||
2993                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
2994                       exist->bytenr == key.offset))) {
2995                         exist = NULL;
2996                         continue;
2997                 }
2998
2999                 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
3000                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
3001                         ret = handle_direct_tree_backref(cache, &key, cur);
3002                         if (ret < 0)
3003                                 goto out;
3004                         continue;
3005                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
3006                         ret = -EINVAL;
3007                         btrfs_print_v0_err(fs_info);
3008                         btrfs_handle_fs_error(fs_info, ret, NULL);
3009                         goto out;
3010                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
3011                         continue;
3012                 }
3013
3014                 /*
3015                  * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
3016                  * means the root objectid. We need to search the tree to get
3017                  * its parent bytenr.
3018                  */
3019                 ret = handle_indirect_tree_backref(cache, path, &key, node_key,
3020                                                    cur);
3021                 if (ret < 0)
3022                         goto out;
3023         }
3024         ret = 0;
3025         cur->checked = 1;
3026         WARN_ON(exist);
3027 out:
3028         btrfs_backref_iter_release(iter);
3029         return ret;
3030 }
3031
3032 /*
3033  * Finish the upwards linkage created by btrfs_backref_add_tree_node()
3034  */
3035 int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
3036                                      struct btrfs_backref_node *start)
3037 {
3038         struct list_head *useless_node = &cache->useless_node;
3039         struct btrfs_backref_edge *edge;
3040         struct rb_node *rb_node;
3041         LIST_HEAD(pending_edge);
3042
3043         ASSERT(start->checked);
3044
3045         /* Insert this node to cache if it's not COW-only */
3046         if (!start->cowonly) {
3047                 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
3048                                            &start->rb_node);
3049                 if (rb_node)
3050                         btrfs_backref_panic(cache->fs_info, start->bytenr,
3051                                             -EEXIST);
3052                 list_add_tail(&start->lower, &cache->leaves);
3053         }
3054
3055         /*
3056          * Use breadth first search to iterate all related edges.
3057          *
3058          * The starting points are all the edges of this node
3059          */
3060         list_for_each_entry(edge, &start->upper, list[LOWER])
3061                 list_add_tail(&edge->list[UPPER], &pending_edge);
3062
3063         while (!list_empty(&pending_edge)) {
3064                 struct btrfs_backref_node *upper;
3065                 struct btrfs_backref_node *lower;
3066
3067                 edge = list_first_entry(&pending_edge,
3068                                 struct btrfs_backref_edge, list[UPPER]);
3069                 list_del_init(&edge->list[UPPER]);
3070                 upper = edge->node[UPPER];
3071                 lower = edge->node[LOWER];
3072
3073                 /* Parent is detached, no need to keep any edges */
3074                 if (upper->detached) {
3075                         list_del(&edge->list[LOWER]);
3076                         btrfs_backref_free_edge(cache, edge);
3077
3078                         /* Lower node is orphan, queue for cleanup */
3079                         if (list_empty(&lower->upper))
3080                                 list_add(&lower->list, useless_node);
3081                         continue;
3082                 }
3083
3084                 /*
3085                  * All new nodes added in current build_backref_tree() haven't
3086                  * been linked to the cache rb tree.
3087                  * So if we have upper->rb_node populated, this means a cache
3088                  * hit. We only need to link the edge, as @upper and all its
3089                  * parents have already been linked.
3090                  */
3091                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3092                         if (upper->lowest) {
3093                                 list_del_init(&upper->lower);
3094                                 upper->lowest = 0;
3095                         }
3096
3097                         list_add_tail(&edge->list[UPPER], &upper->lower);
3098                         continue;
3099                 }
3100
3101                 /* Sanity check, we shouldn't have any unchecked nodes */
3102                 if (!upper->checked) {
3103                         ASSERT(0);
3104                         return -EUCLEAN;
3105                 }
3106
3107                 /* Sanity check, COW-only node has non-COW-only parent */
3108                 if (start->cowonly != upper->cowonly) {
3109                         ASSERT(0);
3110                         return -EUCLEAN;
3111                 }
3112
3113                 /* Only cache non-COW-only (subvolume trees) tree blocks */
3114                 if (!upper->cowonly) {
3115                         rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3116                                                    &upper->rb_node);
3117                         if (rb_node) {
3118                                 btrfs_backref_panic(cache->fs_info,
3119                                                 upper->bytenr, -EEXIST);
3120                                 return -EUCLEAN;
3121                         }
3122                 }
3123
3124                 list_add_tail(&edge->list[UPPER], &upper->lower);
3125
3126                 /*
3127                  * Also queue all the parent edges of this uncached node
3128                  * to finish the upper linkage
3129                  */
3130                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3131                         list_add_tail(&edge->list[UPPER], &pending_edge);
3132         }
3133         return 0;
3134 }
3135
3136 void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3137                                  struct btrfs_backref_node *node)
3138 {
3139         struct btrfs_backref_node *lower;
3140         struct btrfs_backref_node *upper;
3141         struct btrfs_backref_edge *edge;
3142
3143         while (!list_empty(&cache->useless_node)) {
3144                 lower = list_first_entry(&cache->useless_node,
3145                                    struct btrfs_backref_node, list);
3146                 list_del_init(&lower->list);
3147         }
3148         while (!list_empty(&cache->pending_edge)) {
3149                 edge = list_first_entry(&cache->pending_edge,
3150                                 struct btrfs_backref_edge, list[UPPER]);
3151                 list_del(&edge->list[UPPER]);
3152                 list_del(&edge->list[LOWER]);
3153                 lower = edge->node[LOWER];
3154                 upper = edge->node[UPPER];
3155                 btrfs_backref_free_edge(cache, edge);
3156
3157                 /*
3158                  * Lower is no longer linked to any upper backref nodes and
3159                  * isn't in the cache, we can free it ourselves.
3160                  */
3161                 if (list_empty(&lower->upper) &&
3162                     RB_EMPTY_NODE(&lower->rb_node))
3163                         list_add(&lower->list, &cache->useless_node);
3164
3165                 if (!RB_EMPTY_NODE(&upper->rb_node))
3166                         continue;
3167
3168                 /* Add this guy's upper edges to the list to process */
3169                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3170                         list_add_tail(&edge->list[UPPER],
3171                                       &cache->pending_edge);
3172                 if (list_empty(&upper->upper))
3173                         list_add(&upper->list, &cache->useless_node);
3174         }
3175
3176         while (!list_empty(&cache->useless_node)) {
3177                 lower = list_first_entry(&cache->useless_node,
3178                                    struct btrfs_backref_node, list);
3179                 list_del_init(&lower->list);
3180                 if (lower == node)
3181                         node = NULL;
3182                 btrfs_backref_drop_node(cache, lower);
3183         }
3184
3185         btrfs_backref_cleanup_node(cache, node);
3186         ASSERT(list_empty(&cache->useless_node) &&
3187                list_empty(&cache->pending_edge));
3188 }