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