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
  17 /* Just an arbitrary number so we can be sure this happened */
  18 #define BACKREF_FOUND_SHARED 6
  19
  20 struct extent_inode_elem {
  21         u64 inum;
  22         u64 offset;
  23         struct extent_inode_elem *next;
  24 };
  25
  26 static int check_extent_in_eb(const struct btrfs_key *key,
  27                               const struct extent_buffer *eb,
  28                               const struct btrfs_file_extent_item *fi,
  29                               u64 extent_item_pos,
  30                               struct extent_inode_elem **eie,
  31                               bool ignore_offset)
  32 {
  33         u64 offset = 0;
  34         struct extent_inode_elem *e;
  35
  36         if (!ignore_offset &&
  37             !btrfs_file_extent_compression(eb, fi) &&
  38             !btrfs_file_extent_encryption(eb, fi) &&
  39             !btrfs_file_extent_other_encoding(eb, fi)) {
  40                 u64 data_offset;
  41                 u64 data_len;
  42
  43                 data_offset = btrfs_file_extent_offset(eb, fi);
  44                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  45
  46                 if (extent_item_pos < data_offset ||
  47                     extent_item_pos >= data_offset + data_len)
  48                         return 1;
  49                 offset = extent_item_pos - data_offset;
  50         }
  51
  52         e = kmalloc(sizeof(*e), GFP_NOFS);
  53         if (!e)
  54                 return -ENOMEM;
  55
  56         e->next = *eie;
  57         e->inum = key->objectid;
  58         e->offset = key->offset + offset;
  59         *eie = e;
  60
  61         return 0;
  62 }
  63
  64 static void free_inode_elem_list(struct extent_inode_elem *eie)
  65 {
  66         struct extent_inode_elem *eie_next;
  67
  68         for (; eie; eie = eie_next) {
  69                 eie_next = eie->next;
  70                 kfree(eie);
  71         }
  72 }
  73
  74 static int find_extent_in_eb(const struct extent_buffer *eb,
  75                              u64 wanted_disk_byte, u64 extent_item_pos,
  76                              struct extent_inode_elem **eie,
  77                              bool ignore_offset)
  78 {
  79         u64 disk_byte;
  80         struct btrfs_key key;
  81         struct btrfs_file_extent_item *fi;
  82         int slot;
  83         int nritems;
  84         int extent_type;
  85         int ret;
  86
  87         /*
  88          * from the shared data ref, we only have the leaf but we need
  89          * the key. thus, we must look into all items and see that we
  90          * find one (some) with a reference to our extent item.
  91          */
  92         nritems = btrfs_header_nritems(eb);
  93         for (slot = 0; slot < nritems; ++slot) {
  94                 btrfs_item_key_to_cpu(eb, &key, slot);
  95                 if (key.type != BTRFS_EXTENT_DATA_KEY)
  96                         continue;
  97                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
  98                 extent_type = btrfs_file_extent_type(eb, fi);
  99                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 100                         continue;
 101                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 102                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 103                 if (disk_byte != wanted_disk_byte)
 104                         continue;
 105
 106                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
 107                 if (ret < 0)
 108                         return ret;
 109         }
 110
 111         return 0;
 112 }
 113
 114 struct preftree {
 115         struct rb_root_cached root;
 116         unsigned int count;
 117 };
 118
 119 #define PREFTREE_INIT   { .root = RB_ROOT_CACHED, .count = 0 }
 120
 121 struct preftrees {
 122         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 123         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 124         struct preftree indirect_missing_keys;
 125 };
 126
 127 /*
 128  * Checks for a shared extent during backref search.
 129  *
 130  * The share_count tracks prelim_refs (direct and indirect) having a
 131  * ref->count >0:
 132  *  - incremented when a ref->count transitions to >0
 133  *  - decremented when a ref->count transitions to <1
 134  */
 135 struct share_check {
 136         u64 root_objectid;
 137         u64 inum;
 138         int share_count;
 139         bool have_delayed_delete_refs;
 140 };
 141
 142 static inline int extent_is_shared(struct share_check *sc)
 143 {
 144         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 145 }
 146
 147 static struct kmem_cache *btrfs_prelim_ref_cache;
 148
 149 int __init btrfs_prelim_ref_init(void)
 150 {
 151         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 152                                         sizeof(struct prelim_ref),
 153                                         0,
 154                                         SLAB_MEM_SPREAD,
 155                                         NULL);
 156         if (!btrfs_prelim_ref_cache)
 157                 return -ENOMEM;
 158         return 0;
 159 }
 160
 161 void __cold btrfs_prelim_ref_exit(void)
 162 {
 163         kmem_cache_destroy(btrfs_prelim_ref_cache);
 164 }
 165
 166 static void free_pref(struct prelim_ref *ref)
 167 {
 168         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 169 }
 170
 171 /*
 172  * Return 0 when both refs are for the same block (and can be merged).
 173  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 174  * indicates a 'higher' block.
 175  */
 176 static int prelim_ref_compare(struct prelim_ref *ref1,
 177                               struct prelim_ref *ref2)
 178 {
 179         if (ref1->level < ref2->level)
 180                 return -1;
 181         if (ref1->level > ref2->level)
 182                 return 1;
 183         if (ref1->root_id < ref2->root_id)
 184                 return -1;
 185         if (ref1->root_id > ref2->root_id)
 186                 return 1;
 187         if (ref1->key_for_search.type < ref2->key_for_search.type)
 188                 return -1;
 189         if (ref1->key_for_search.type > ref2->key_for_search.type)
 190                 return 1;
 191         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 192                 return -1;
 193         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 194                 return 1;
 195         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 196                 return -1;
 197         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 198                 return 1;
 199         if (ref1->parent < ref2->parent)
 200                 return -1;
 201         if (ref1->parent > ref2->parent)
 202                 return 1;
 203
 204         return 0;
 205 }
 206
 207 static void update_share_count(struct share_check *sc, int oldcount,
 208                                int newcount)
 209 {
 210         if ((!sc) || (oldcount == 0 && newcount < 1))
 211                 return;
 212
 213         if (oldcount > 0 && newcount < 1)
 214                 sc->share_count--;
 215         else if (oldcount < 1 && newcount > 0)
 216                 sc->share_count++;
 217 }
 218
 219 /*
 220  * Add @newref to the @root rbtree, merging identical refs.
 221  *
 222  * Callers should assume that newref has been freed after calling.
 223  */
 224 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 225                               struct preftree *preftree,
 226                               struct prelim_ref *newref,
 227                               struct share_check *sc)
 228 {
 229         struct rb_root_cached *root;
 230         struct rb_node **p;
 231         struct rb_node *parent = NULL;
 232         struct prelim_ref *ref;
 233         int result;
 234         bool leftmost = true;
 235
 236         root = &preftree->root;
 237         p = &root->rb_root.rb_node;
 238
 239         while (*p) {
 240                 parent = *p;
 241                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 242                 result = prelim_ref_compare(ref, newref);
 243                 if (result < 0) {
 244                         p = &(*p)->rb_left;
 245                 } else if (result > 0) {
 246                         p = &(*p)->rb_right;
 247                         leftmost = false;
 248                 } else {
 249                         /* Identical refs, merge them and free @newref */
 250                         struct extent_inode_elem *eie = ref->inode_list;
 251
 252                         while (eie && eie->next)
 253                                 eie = eie->next;
 254
 255                         if (!eie)
 256                                 ref->inode_list = newref->inode_list;
 257                         else
 258                                 eie->next = newref->inode_list;
 259                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 260                                                      preftree->count);
 261                         /*
 262                          * A delayed ref can have newref->count < 0.
 263                          * The ref->count is updated to follow any
 264                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 265                          */
 266                         update_share_count(sc, ref->count,
 267                                            ref->count + newref->count);
 268                         ref->count += newref->count;
 269                         free_pref(newref);
 270                         return;
 271                 }
 272         }
 273
 274         update_share_count(sc, 0, newref->count);
 275         preftree->count++;
 276         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 277         rb_link_node(&newref->rbnode, parent, p);
 278         rb_insert_color_cached(&newref->rbnode, root, leftmost);
 279 }
 280
 281 /*
 282  * Release the entire tree.  We don't care about internal consistency so
 283  * just free everything and then reset the tree root.
 284  */
 285 static void prelim_release(struct preftree *preftree)
 286 {
 287         struct prelim_ref *ref, *next_ref;
 288
 289         rbtree_postorder_for_each_entry_safe(ref, next_ref,
 290                                              &preftree->root.rb_root, rbnode) {
 291                 free_inode_elem_list(ref->inode_list);
 292                 free_pref(ref);
 293         }
 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 = {0};
 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         u8 type;
 435
 436         if (level != 0) {
 437                 eb = path->nodes[level];
 438                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 439                 if (ret < 0)
 440                         return ret;
 441                 return 0;
 442         }
 443
 444         /*
 445          * 1. We normally enter this function with the path already pointing to
 446          *    the first item to check. But sometimes, we may enter it with
 447          *    slot == nritems.
 448          * 2. We are searching for normal backref but bytenr of this leaf
 449          *    matches shared data backref
 450          * 3. The leaf owner is not equal to the root we are searching
 451          *
 452          * For these cases, go to the next leaf before we continue.
 453          */
 454         eb = path->nodes[0];
 455         if (path->slots[0] >= btrfs_header_nritems(eb) ||
 456             is_shared_data_backref(preftrees, eb->start) ||
 457             ref->root_id != btrfs_header_owner(eb)) {
 458                 if (time_seq == SEQ_LAST)
 459                         ret = btrfs_next_leaf(root, path);
 460                 else
 461                         ret = btrfs_next_old_leaf(root, path, time_seq);
 462         }
 463
 464         while (!ret && count < ref->count) {
 465                 eb = path->nodes[0];
 466                 slot = path->slots[0];
 467
 468                 btrfs_item_key_to_cpu(eb, &key, slot);
 469
 470                 if (key.objectid != key_for_search->objectid ||
 471                     key.type != BTRFS_EXTENT_DATA_KEY)
 472                         break;
 473
 474                 /*
 475                  * We are searching for normal backref but bytenr of this leaf
 476                  * matches shared data backref, OR
 477                  * the leaf owner is not equal to the root we are searching for
 478                  */
 479                 if (slot == 0 &&
 480                     (is_shared_data_backref(preftrees, eb->start) ||
 481                      ref->root_id != btrfs_header_owner(eb))) {
 482                         if (time_seq == SEQ_LAST)
 483                                 ret = btrfs_next_leaf(root, path);
 484                         else
 485                                 ret = btrfs_next_old_leaf(root, path, time_seq);
 486                         continue;
 487                 }
 488                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 489                 type = btrfs_file_extent_type(eb, fi);
 490                 if (type == BTRFS_FILE_EXTENT_INLINE)
 491                         goto next;
 492                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 493                 data_offset = btrfs_file_extent_offset(eb, fi);
 494
 495                 if (disk_byte == wanted_disk_byte) {
 496                         eie = NULL;
 497                         old = NULL;
 498                         if (ref->key_for_search.offset == key.offset - data_offset)
 499                                 count++;
 500                         else
 501                                 goto next;
 502                         if (extent_item_pos) {
 503                                 ret = check_extent_in_eb(&key, eb, fi,
 504                                                 *extent_item_pos,
 505                                                 &eie, ignore_offset);
 506                                 if (ret < 0)
 507                                         break;
 508                         }
 509                         if (ret > 0)
 510                                 goto next;
 511                         ret = ulist_add_merge_ptr(parents, eb->start,
 512                                                   eie, (void **)&old, GFP_NOFS);
 513                         if (ret < 0)
 514                                 break;
 515                         if (!ret && extent_item_pos) {
 516                                 while (old->next)
 517                                         old = old->next;
 518                                 old->next = eie;
 519                         }
 520                         eie = NULL;
 521                 }
 522 next:
 523                 if (time_seq == SEQ_LAST)
 524                         ret = btrfs_next_item(root, path);
 525                 else
 526                         ret = btrfs_next_old_item(root, path, time_seq);
 527         }
 528
 529         if (ret > 0)
 530                 ret = 0;
 531         else if (ret < 0)
 532                 free_inode_elem_list(eie);
 533         return ret;
 534 }
 535
 536 /*
 537  * resolve an indirect backref in the form (root_id, key, level)
 538  * to a logical address
 539  */
 540 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 541                                 struct btrfs_path *path, u64 time_seq,
 542                                 struct preftrees *preftrees,
 543                                 struct prelim_ref *ref, struct ulist *parents,
 544                                 const u64 *extent_item_pos, bool ignore_offset)
 545 {
 546         struct btrfs_root *root;
 547         struct btrfs_key root_key;
 548         struct extent_buffer *eb;
 549         int ret = 0;
 550         int root_level;
 551         int level = ref->level;
 552         int index;
 553         struct btrfs_key search_key = ref->key_for_search;
 554
 555         root_key.objectid = ref->root_id;
 556         root_key.type = BTRFS_ROOT_ITEM_KEY;
 557         root_key.offset = (u64)-1;
 558
 559         index = srcu_read_lock(&fs_info->subvol_srcu);
 560
 561         root = btrfs_get_fs_root(fs_info, &root_key, false);
 562         if (IS_ERR(root)) {
 563                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 564                 ret = PTR_ERR(root);
 565                 goto out;
 566         }
 567
 568         if (btrfs_is_testing(fs_info)) {
 569                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 570                 ret = -ENOENT;
 571                 goto out;
 572         }
 573
 574         if (path->search_commit_root)
 575                 root_level = btrfs_header_level(root->commit_root);
 576         else if (time_seq == SEQ_LAST)
 577                 root_level = btrfs_header_level(root->node);
 578         else
 579                 root_level = btrfs_old_root_level(root, time_seq);
 580
 581         if (root_level + 1 == level) {
 582                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 583                 goto out;
 584         }
 585
 586         /*
 587          * We can often find data backrefs with an offset that is too large
 588          * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
 589          * subtracting a file's offset with the data offset of its
 590          * corresponding extent data item. This can happen for example in the
 591          * clone ioctl.
 592          *
 593          * So if we detect such case we set the search key's offset to zero to
 594          * make sure we will find the matching file extent item at
 595          * add_all_parents(), otherwise we will miss it because the offset
 596          * taken form the backref is much larger then the offset of the file
 597          * extent item. This can make us scan a very large number of file
 598          * extent items, but at least it will not make us miss any.
 599          *
 600          * This is an ugly workaround for a behaviour that should have never
 601          * existed, but it does and a fix for the clone ioctl would touch a lot
 602          * of places, cause backwards incompatibility and would not fix the
 603          * problem for extents cloned with older kernels.
 604          */
 605         if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
 606             search_key.offset >= LLONG_MAX)
 607                 search_key.offset = 0;
 608         path->lowest_level = level;
 609         if (time_seq == SEQ_LAST)
 610                 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 611         else
 612                 ret = btrfs_search_old_slot(root, &search_key, path, time_seq);
 613
 614         /* root node has been locked, we can release @subvol_srcu safely here */
 615         srcu_read_unlock(&fs_info->subvol_srcu, index);
 616
 617         btrfs_debug(fs_info,
 618                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 619                  ref->root_id, level, ref->count, ret,
 620                  ref->key_for_search.objectid, ref->key_for_search.type,
 621                  ref->key_for_search.offset);
 622         if (ret < 0)
 623                 goto out;
 624
 625         eb = path->nodes[level];
 626         while (!eb) {
 627                 if (WARN_ON(!level)) {
 628                         ret = 1;
 629                         goto out;
 630                 }
 631                 level--;
 632                 eb = path->nodes[level];
 633         }
 634
 635         ret = add_all_parents(root, path, parents, preftrees, ref, level,
 636                               time_seq, extent_item_pos, ignore_offset);
 637 out:
 638         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 static void free_leaf_list(struct ulist *ulist)
 652 {
 653         struct ulist_node *node;
 654         struct ulist_iterator uiter;
 655
 656         ULIST_ITER_INIT(&uiter);
 657         while ((node = ulist_next(ulist, &uiter)))
 658                 free_inode_elem_list(unode_aux_to_inode_list(node));
 659
 660         ulist_free(ulist);
 661 }
 662
 663 /*
 664  * We maintain three separate rbtrees: one for direct refs, one for
 665  * indirect refs which have a key, and one for indirect refs which do not
 666  * have a key. Each tree does merge on insertion.
 667  *
 668  * Once all of the references are located, we iterate over the tree of
 669  * indirect refs with missing keys. An appropriate key is located and
 670  * the ref is moved onto the tree for indirect refs. After all missing
 671  * keys are thus located, we iterate over the indirect ref tree, resolve
 672  * each reference, and then insert the resolved reference onto the
 673  * direct tree (merging there too).
 674  *
 675  * New backrefs (i.e., for parent nodes) are added to the appropriate
 676  * rbtree as they are encountered. The new backrefs are subsequently
 677  * resolved as above.
 678  */
 679 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 680                                  struct btrfs_path *path, u64 time_seq,
 681                                  struct preftrees *preftrees,
 682                                  const u64 *extent_item_pos,
 683                                  struct share_check *sc, bool ignore_offset)
 684 {
 685         int err;
 686         int ret = 0;
 687         struct ulist *parents;
 688         struct ulist_node *node;
 689         struct ulist_iterator uiter;
 690         struct rb_node *rnode;
 691
 692         parents = ulist_alloc(GFP_NOFS);
 693         if (!parents)
 694                 return -ENOMEM;
 695
 696         /*
 697          * We could trade memory usage for performance here by iterating
 698          * the tree, allocating new refs for each insertion, and then
 699          * freeing the entire indirect tree when we're done.  In some test
 700          * cases, the tree can grow quite large (~200k objects).
 701          */
 702         while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
 703                 struct prelim_ref *ref;
 704
 705                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 706                 if (WARN(ref->parent,
 707                          "BUG: direct ref found in indirect tree")) {
 708                         ret = -EINVAL;
 709                         goto out;
 710                 }
 711
 712                 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
 713                 preftrees->indirect.count--;
 714
 715                 if (ref->count == 0) {
 716                         free_pref(ref);
 717                         continue;
 718                 }
 719
 720                 if (sc && sc->root_objectid &&
 721                     ref->root_id != sc->root_objectid) {
 722                         free_pref(ref);
 723                         ret = BACKREF_FOUND_SHARED;
 724                         goto out;
 725                 }
 726                 err = resolve_indirect_ref(fs_info, path, time_seq, preftrees,
 727                                            ref, parents, extent_item_pos,
 728                                            ignore_offset);
 729                 /*
 730                  * we can only tolerate ENOENT,otherwise,we should catch error
 731                  * and return directly.
 732                  */
 733                 if (err == -ENOENT) {
 734                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 735                                           NULL);
 736                         continue;
 737                 } else if (err) {
 738                         free_pref(ref);
 739                         ret = err;
 740                         goto out;
 741                 }
 742
 743                 /* we put the first parent into the ref at hand */
 744                 ULIST_ITER_INIT(&uiter);
 745                 node = ulist_next(parents, &uiter);
 746                 ref->parent = node ? node->val : 0;
 747                 ref->inode_list = unode_aux_to_inode_list(node);
 748
 749                 /* Add a prelim_ref(s) for any other parent(s). */
 750                 while ((node = ulist_next(parents, &uiter))) {
 751                         struct prelim_ref *new_ref;
 752
 753                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 754                                                    GFP_NOFS);
 755                         if (!new_ref) {
 756                                 free_pref(ref);
 757                                 ret = -ENOMEM;
 758                                 goto out;
 759                         }
 760                         memcpy(new_ref, ref, sizeof(*ref));
 761                         new_ref->parent = node->val;
 762                         new_ref->inode_list = unode_aux_to_inode_list(node);
 763                         prelim_ref_insert(fs_info, &preftrees->direct,
 764                                           new_ref, NULL);
 765                 }
 766
 767                 /*
 768                  * Now it's a direct ref, put it in the direct tree. We must
 769                  * do this last because the ref could be merged/freed here.
 770                  */
 771                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 772
 773                 ulist_reinit(parents);
 774                 cond_resched();
 775         }
 776 out:
 777         /*
 778          * We may have inode lists attached to refs in the parents ulist, so we
 779          * must free them before freeing the ulist and its refs.
 780          */
 781         free_leaf_list(parents);
 782         return ret;
 783 }
 784
 785 /*
 786  * read tree blocks and add keys where required.
 787  */
 788 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 789                             struct preftrees *preftrees, bool lock)
 790 {
 791         struct prelim_ref *ref;
 792         struct extent_buffer *eb;
 793         struct preftree *tree = &preftrees->indirect_missing_keys;
 794         struct rb_node *node;
 795
 796         while ((node = rb_first_cached(&tree->root))) {
 797                 ref = rb_entry(node, struct prelim_ref, rbnode);
 798                 rb_erase_cached(node, &tree->root);
 799
 800                 BUG_ON(ref->parent);    /* should not be a direct ref */
 801                 BUG_ON(ref->key_for_search.type);
 802                 BUG_ON(!ref->wanted_disk_byte);
 803
 804                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0,
 805                                      ref->level - 1, NULL);
 806                 if (IS_ERR(eb)) {
 807                         free_pref(ref);
 808                         return PTR_ERR(eb);
 809                 } else if (!extent_buffer_uptodate(eb)) {
 810                         free_pref(ref);
 811                         free_extent_buffer(eb);
 812                         return -EIO;
 813                 }
 814                 if (lock)
 815                         btrfs_tree_read_lock(eb);
 816                 if (btrfs_header_level(eb) == 0)
 817                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 818                 else
 819                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 820                 if (lock)
 821                         btrfs_tree_read_unlock(eb);
 822                 free_extent_buffer(eb);
 823                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 824                 cond_resched();
 825         }
 826         return 0;
 827 }
 828
 829 /*
 830  * add all currently queued delayed refs from this head whose seq nr is
 831  * smaller or equal that seq to the list
 832  */
 833 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 834                             struct btrfs_delayed_ref_head *head, u64 seq,
 835                             struct preftrees *preftrees, struct share_check *sc)
 836 {
 837         struct btrfs_delayed_ref_node *node;
 838         struct btrfs_key key;
 839         struct rb_node *n;
 840         int count;
 841         int ret = 0;
 842
 843         spin_lock(&head->lock);
 844         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 845                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 846                                 ref_node);
 847                 if (node->seq > seq)
 848                         continue;
 849
 850                 switch (node->action) {
 851                 case BTRFS_ADD_DELAYED_EXTENT:
 852                 case BTRFS_UPDATE_DELAYED_HEAD:
 853                         WARN_ON(1);
 854                         continue;
 855                 case BTRFS_ADD_DELAYED_REF:
 856                         count = node->ref_mod;
 857                         break;
 858                 case BTRFS_DROP_DELAYED_REF:
 859                         count = node->ref_mod * -1;
 860                         break;
 861                 default:
 862                         BUG();
 863                 }
 864                 switch (node->type) {
 865                 case BTRFS_TREE_BLOCK_REF_KEY: {
 866                         /* NORMAL INDIRECT METADATA backref */
 867                         struct btrfs_delayed_tree_ref *ref;
 868                         struct btrfs_key *key_ptr = NULL;
 869
 870                         if (head->extent_op && head->extent_op->update_key) {
 871                                 btrfs_disk_key_to_cpu(&key, &head->extent_op->key);
 872                                 key_ptr = &key;
 873                         }
 874
 875                         ref = btrfs_delayed_node_to_tree_ref(node);
 876                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 877                                                key_ptr, ref->level + 1,
 878                                                node->bytenr, count, sc,
 879                                                GFP_ATOMIC);
 880                         break;
 881                 }
 882                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 883                         /* SHARED DIRECT METADATA backref */
 884                         struct btrfs_delayed_tree_ref *ref;
 885
 886                         ref = btrfs_delayed_node_to_tree_ref(node);
 887
 888                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 889                                              ref->parent, node->bytenr, count,
 890                                              sc, GFP_ATOMIC);
 891                         break;
 892                 }
 893                 case BTRFS_EXTENT_DATA_REF_KEY: {
 894                         /* NORMAL INDIRECT DATA backref */
 895                         struct btrfs_delayed_data_ref *ref;
 896                         ref = btrfs_delayed_node_to_data_ref(node);
 897
 898                         key.objectid = ref->objectid;
 899                         key.type = BTRFS_EXTENT_DATA_KEY;
 900                         key.offset = ref->offset;
 901
 902                         /*
 903                          * If we have a share check context and a reference for
 904                          * another inode, we can't exit immediately. This is
 905                          * because even if this is a BTRFS_ADD_DELAYED_REF
 906                          * reference we may find next a BTRFS_DROP_DELAYED_REF
 907                          * which cancels out this ADD reference.
 908                          *
 909                          * If this is a DROP reference and there was no previous
 910                          * ADD reference, then we need to signal that when we
 911                          * process references from the extent tree (through
 912                          * add_inline_refs() and add_keyed_refs()), we should
 913                          * not exit early if we find a reference for another
 914                          * inode, because one of the delayed DROP references
 915                          * may cancel that reference in the extent tree.
 916                          */
 917                         if (sc && count < 0)
 918                                 sc->have_delayed_delete_refs = true;
 919
 920                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 921                                                &key, 0, node->bytenr, count, sc,
 922                                                GFP_ATOMIC);
 923                         break;
 924                 }
 925                 case BTRFS_SHARED_DATA_REF_KEY: {
 926                         /* SHARED DIRECT FULL backref */
 927                         struct btrfs_delayed_data_ref *ref;
 928
 929                         ref = btrfs_delayed_node_to_data_ref(node);
 930
 931                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 932                                              node->bytenr, count, sc,
 933                                              GFP_ATOMIC);
 934                         break;
 935                 }
 936                 default:
 937                         WARN_ON(1);
 938                 }
 939                 /*
 940                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 941                  * refs have been checked.
 942                  */
 943                 if (ret && (ret != BACKREF_FOUND_SHARED))
 944                         break;
 945         }
 946         if (!ret)
 947                 ret = extent_is_shared(sc);
 948
 949         spin_unlock(&head->lock);
 950         return ret;
 951 }
 952
 953 /*
 954  * add all inline backrefs for bytenr to the list
 955  *
 956  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 957  */
 958 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 959                            struct btrfs_path *path, u64 bytenr,
 960                            int *info_level, struct preftrees *preftrees,
 961                            struct share_check *sc)
 962 {
 963         int ret = 0;
 964         int slot;
 965         struct extent_buffer *leaf;
 966         struct btrfs_key key;
 967         struct btrfs_key found_key;
 968         unsigned long ptr;
 969         unsigned long end;
 970         struct btrfs_extent_item *ei;
 971         u64 flags;
 972         u64 item_size;
 973
 974         /*
 975          * enumerate all inline refs
 976          */
 977         leaf = path->nodes[0];
 978         slot = path->slots[0];
 979
 980         item_size = btrfs_item_size_nr(leaf, slot);
 981         BUG_ON(item_size < sizeof(*ei));
 982
 983         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 984         flags = btrfs_extent_flags(leaf, ei);
 985         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 986
 987         ptr = (unsigned long)(ei + 1);
 988         end = (unsigned long)ei + item_size;
 989
 990         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 991             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 992                 struct btrfs_tree_block_info *info;
 993
 994                 info = (struct btrfs_tree_block_info *)ptr;
 995                 *info_level = btrfs_tree_block_level(leaf, info);
 996                 ptr += sizeof(struct btrfs_tree_block_info);
 997                 BUG_ON(ptr > end);
 998         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 999                 *info_level = found_key.offset;
1000         } else {
1001                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1002         }
1003
1004         while (ptr < end) {
1005                 struct btrfs_extent_inline_ref *iref;
1006                 u64 offset;
1007                 int type;
1008
1009                 iref = (struct btrfs_extent_inline_ref *)ptr;
1010                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
1011                                                         BTRFS_REF_TYPE_ANY);
1012                 if (type == BTRFS_REF_TYPE_INVALID)
1013                         return -EUCLEAN;
1014
1015                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1016
1017                 switch (type) {
1018                 case BTRFS_SHARED_BLOCK_REF_KEY:
1019                         ret = add_direct_ref(fs_info, preftrees,
1020                                              *info_level + 1, offset,
1021                                              bytenr, 1, NULL, GFP_NOFS);
1022                         break;
1023                 case BTRFS_SHARED_DATA_REF_KEY: {
1024                         struct btrfs_shared_data_ref *sdref;
1025                         int count;
1026
1027                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1028                         count = btrfs_shared_data_ref_count(leaf, sdref);
1029
1030                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
1031                                              bytenr, count, sc, GFP_NOFS);
1032                         break;
1033                 }
1034                 case BTRFS_TREE_BLOCK_REF_KEY:
1035                         ret = add_indirect_ref(fs_info, preftrees, offset,
1036                                                NULL, *info_level + 1,
1037                                                bytenr, 1, NULL, GFP_NOFS);
1038                         break;
1039                 case BTRFS_EXTENT_DATA_REF_KEY: {
1040                         struct btrfs_extent_data_ref *dref;
1041                         int count;
1042                         u64 root;
1043
1044                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1045                         count = btrfs_extent_data_ref_count(leaf, dref);
1046                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1047                                                                       dref);
1048                         key.type = BTRFS_EXTENT_DATA_KEY;
1049                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1050
1051                         if (sc && sc->inum && key.objectid != sc->inum &&
1052                             !sc->have_delayed_delete_refs) {
1053                                 ret = BACKREF_FOUND_SHARED;
1054                                 break;
1055                         }
1056
1057                         root = btrfs_extent_data_ref_root(leaf, dref);
1058
1059                         ret = add_indirect_ref(fs_info, preftrees, root,
1060                                                &key, 0, bytenr, count,
1061                                                sc, GFP_NOFS);
1062
1063                         break;
1064                 }
1065                 default:
1066                         WARN_ON(1);
1067                 }
1068                 if (ret)
1069                         return ret;
1070                 ptr += btrfs_extent_inline_ref_size(type);
1071         }
1072
1073         return 0;
1074 }
1075
1076 /*
1077  * add all non-inline backrefs for bytenr to the list
1078  *
1079  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1080  */
1081 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1082                           struct btrfs_path *path, u64 bytenr,
1083                           int info_level, struct preftrees *preftrees,
1084                           struct share_check *sc)
1085 {
1086         struct btrfs_root *extent_root = fs_info->extent_root;
1087         int ret;
1088         int slot;
1089         struct extent_buffer *leaf;
1090         struct btrfs_key key;
1091
1092         while (1) {
1093                 ret = btrfs_next_item(extent_root, path);
1094                 if (ret < 0)
1095                         break;
1096                 if (ret) {
1097                         ret = 0;
1098                         break;
1099                 }
1100
1101                 slot = path->slots[0];
1102                 leaf = path->nodes[0];
1103                 btrfs_item_key_to_cpu(leaf, &key, slot);
1104
1105                 if (key.objectid != bytenr)
1106                         break;
1107                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1108                         continue;
1109                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1110                         break;
1111
1112                 switch (key.type) {
1113                 case BTRFS_SHARED_BLOCK_REF_KEY:
1114                         /* SHARED DIRECT METADATA backref */
1115                         ret = add_direct_ref(fs_info, preftrees,
1116                                              info_level + 1, key.offset,
1117                                              bytenr, 1, NULL, GFP_NOFS);
1118                         break;
1119                 case BTRFS_SHARED_DATA_REF_KEY: {
1120                         /* SHARED DIRECT FULL backref */
1121                         struct btrfs_shared_data_ref *sdref;
1122                         int count;
1123
1124                         sdref = btrfs_item_ptr(leaf, slot,
1125                                               struct btrfs_shared_data_ref);
1126                         count = btrfs_shared_data_ref_count(leaf, sdref);
1127                         ret = add_direct_ref(fs_info, preftrees, 0,
1128                                              key.offset, bytenr, count,
1129                                              sc, GFP_NOFS);
1130                         break;
1131                 }
1132                 case BTRFS_TREE_BLOCK_REF_KEY:
1133                         /* NORMAL INDIRECT METADATA backref */
1134                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1135                                                NULL, info_level + 1, bytenr,
1136                                                1, NULL, GFP_NOFS);
1137                         break;
1138                 case BTRFS_EXTENT_DATA_REF_KEY: {
1139                         /* NORMAL INDIRECT DATA backref */
1140                         struct btrfs_extent_data_ref *dref;
1141                         int count;
1142                         u64 root;
1143
1144                         dref = btrfs_item_ptr(leaf, slot,
1145                                               struct btrfs_extent_data_ref);
1146                         count = btrfs_extent_data_ref_count(leaf, dref);
1147                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1148                                                                       dref);
1149                         key.type = BTRFS_EXTENT_DATA_KEY;
1150                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1151
1152                         if (sc && sc->inum && key.objectid != sc->inum &&
1153                             !sc->have_delayed_delete_refs) {
1154                                 ret = BACKREF_FOUND_SHARED;
1155                                 break;
1156                         }
1157
1158                         root = btrfs_extent_data_ref_root(leaf, dref);
1159                         ret = add_indirect_ref(fs_info, preftrees, root,
1160                                                &key, 0, bytenr, count,
1161                                                sc, GFP_NOFS);
1162                         break;
1163                 }
1164                 default:
1165                         WARN_ON(1);
1166                 }
1167                 if (ret)
1168                         return ret;
1169
1170         }
1171
1172         return ret;
1173 }
1174
1175 /*
1176  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1177  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1178  * indirect refs to their parent bytenr.
1179  * When roots are found, they're added to the roots list
1180  *
1181  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1182  * much like trans == NULL case, the difference only lies in it will not
1183  * commit root.
1184  * The special case is for qgroup to search roots in commit_transaction().
1185  *
1186  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1187  * shared extent is detected.
1188  *
1189  * Otherwise this returns 0 for success and <0 for an error.
1190  *
1191  * If ignore_offset is set to false, only extent refs whose offsets match
1192  * extent_item_pos are returned.  If true, every extent ref is returned
1193  * and extent_item_pos is ignored.
1194  *
1195  * FIXME some caching might speed things up
1196  */
1197 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1198                              struct btrfs_fs_info *fs_info, u64 bytenr,
1199                              u64 time_seq, struct ulist *refs,
1200                              struct ulist *roots, const u64 *extent_item_pos,
1201                              struct share_check *sc, bool ignore_offset)
1202 {
1203         struct btrfs_key key;
1204         struct btrfs_path *path;
1205         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1206         struct btrfs_delayed_ref_head *head;
1207         int info_level = 0;
1208         int ret;
1209         struct prelim_ref *ref;
1210         struct rb_node *node;
1211         struct extent_inode_elem *eie = NULL;
1212         struct preftrees preftrees = {
1213                 .direct = PREFTREE_INIT,
1214                 .indirect = PREFTREE_INIT,
1215                 .indirect_missing_keys = PREFTREE_INIT
1216         };
1217
1218         key.objectid = bytenr;
1219         key.offset = (u64)-1;
1220         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1221                 key.type = BTRFS_METADATA_ITEM_KEY;
1222         else
1223                 key.type = BTRFS_EXTENT_ITEM_KEY;
1224
1225         path = btrfs_alloc_path();
1226         if (!path)
1227                 return -ENOMEM;
1228         if (!trans) {
1229                 path->search_commit_root = 1;
1230                 path->skip_locking = 1;
1231         }
1232
1233         if (time_seq == SEQ_LAST)
1234                 path->skip_locking = 1;
1235
1236         /*
1237          * grab both a lock on the path and a lock on the delayed ref head.
1238          * We need both to get a consistent picture of how the refs look
1239          * at a specified point in time
1240          */
1241 again:
1242         head = NULL;
1243
1244         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1245         if (ret < 0)
1246                 goto out;
1247         if (ret == 0) {
1248                 /* This shouldn't happen, indicates a bug or fs corruption. */
1249                 ASSERT(ret != 0);
1250                 ret = -EUCLEAN;
1251                 goto out;
1252         }
1253
1254 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1255         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1256             time_seq != SEQ_LAST) {
1257 #else
1258         if (trans && time_seq != SEQ_LAST) {
1259 #endif
1260                 /*
1261                  * look if there are updates for this ref queued and lock the
1262                  * head
1263                  */
1264                 delayed_refs = &trans->transaction->delayed_refs;
1265                 spin_lock(&delayed_refs->lock);
1266                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1267                 if (head) {
1268                         if (!mutex_trylock(&head->mutex)) {
1269                                 refcount_inc(&head->refs);
1270                                 spin_unlock(&delayed_refs->lock);
1271
1272                                 btrfs_release_path(path);
1273
1274                                 /*
1275                                  * Mutex was contended, block until it's
1276                                  * released and try again
1277                                  */
1278                                 mutex_lock(&head->mutex);
1279                                 mutex_unlock(&head->mutex);
1280                                 btrfs_put_delayed_ref_head(head);
1281                                 goto again;
1282                         }
1283                         spin_unlock(&delayed_refs->lock);
1284                         ret = add_delayed_refs(fs_info, head, time_seq,
1285                                                &preftrees, sc);
1286                         mutex_unlock(&head->mutex);
1287                         if (ret)
1288                                 goto out;
1289                 } else {
1290                         spin_unlock(&delayed_refs->lock);
1291                 }
1292         }
1293
1294         if (path->slots[0]) {
1295                 struct extent_buffer *leaf;
1296                 int slot;
1297
1298                 path->slots[0]--;
1299                 leaf = path->nodes[0];
1300                 slot = path->slots[0];
1301                 btrfs_item_key_to_cpu(leaf, &key, slot);
1302                 if (key.objectid == bytenr &&
1303                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1304                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1305                         ret = add_inline_refs(fs_info, path, bytenr,
1306                                               &info_level, &preftrees, sc);
1307                         if (ret)
1308                                 goto out;
1309                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1310                                              &preftrees, sc);
1311                         if (ret)
1312                                 goto out;
1313                 }
1314         }
1315
1316         btrfs_release_path(path);
1317
1318         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1319         if (ret)
1320                 goto out;
1321
1322         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1323
1324         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1325                                     extent_item_pos, sc, ignore_offset);
1326         if (ret)
1327                 goto out;
1328
1329         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1330
1331         /*
1332          * This walks the tree of merged and resolved refs. Tree blocks are
1333          * read in as needed. Unique entries are added to the ulist, and
1334          * the list of found roots is updated.
1335          *
1336          * We release the entire tree in one go before returning.
1337          */
1338         node = rb_first_cached(&preftrees.direct.root);
1339         while (node) {
1340                 ref = rb_entry(node, struct prelim_ref, rbnode);
1341                 node = rb_next(&ref->rbnode);
1342                 /*
1343                  * ref->count < 0 can happen here if there are delayed
1344                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1345                  * prelim_ref_insert() relies on this when merging
1346                  * identical refs to keep the overall count correct.
1347                  * prelim_ref_insert() will merge only those refs
1348                  * which compare identically.  Any refs having
1349                  * e.g. different offsets would not be merged,
1350                  * and would retain their original ref->count < 0.
1351                  */
1352                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1353                         if (sc && sc->root_objectid &&
1354                             ref->root_id != sc->root_objectid) {
1355                                 ret = BACKREF_FOUND_SHARED;
1356                                 goto out;
1357                         }
1358
1359                         /* no parent == root of tree */
1360                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1361                         if (ret < 0)
1362                                 goto out;
1363                 }
1364                 if (ref->count && ref->parent) {
1365                         if (extent_item_pos && !ref->inode_list &&
1366                             ref->level == 0) {
1367                                 struct extent_buffer *eb;
1368
1369                                 eb = read_tree_block(fs_info, ref->parent, 0,
1370                                                      ref->level, NULL);
1371                                 if (IS_ERR(eb)) {
1372                                         ret = PTR_ERR(eb);
1373                                         goto out;
1374                                 } else if (!extent_buffer_uptodate(eb)) {
1375                                         free_extent_buffer(eb);
1376                                         ret = -EIO;
1377                                         goto out;
1378                                 }
1379
1380                                 if (!path->skip_locking) {
1381                                         btrfs_tree_read_lock(eb);
1382                                         btrfs_set_lock_blocking_read(eb);
1383                                 }
1384                                 ret = find_extent_in_eb(eb, bytenr,
1385                                                         *extent_item_pos, &eie, ignore_offset);
1386                                 if (!path->skip_locking)
1387                                         btrfs_tree_read_unlock_blocking(eb);
1388                                 free_extent_buffer(eb);
1389                                 if (ret < 0)
1390                                         goto out;
1391                                 ref->inode_list = eie;
1392                                 /*
1393                                  * We transferred the list ownership to the ref,
1394                                  * so set to NULL to avoid a double free in case
1395                                  * an error happens after this.
1396                                  */
1397                                 eie = NULL;
1398                         }
1399                         ret = ulist_add_merge_ptr(refs, ref->parent,
1400                                                   ref->inode_list,
1401                                                   (void **)&eie, GFP_NOFS);
1402                         if (ret < 0)
1403                                 goto out;
1404                         if (!ret && extent_item_pos) {
1405                                 /*
1406                                  * We've recorded that parent, so we must extend
1407                                  * its inode list here.
1408                                  *
1409                                  * However if there was corruption we may not
1410                                  * have found an eie, return an error in this
1411                                  * case.
1412                                  */
1413                                 ASSERT(eie);
1414                                 if (!eie) {
1415                                         ret = -EUCLEAN;
1416                                         goto out;
1417                                 }
1418                                 while (eie->next)
1419                                         eie = eie->next;
1420                                 eie->next = ref->inode_list;
1421                         }
1422                         eie = NULL;
1423                         /*
1424                          * We have transferred the inode list ownership from
1425                          * this ref to the ref we added to the 'refs' ulist.
1426                          * So set this ref's inode list to NULL to avoid
1427                          * use-after-free when our caller uses it or double
1428                          * frees in case an error happens before we return.
1429                          */
1430                         ref->inode_list = NULL;
1431                 }
1432                 cond_resched();
1433         }
1434
1435 out:
1436         btrfs_free_path(path);
1437
1438         prelim_release(&preftrees.direct);
1439         prelim_release(&preftrees.indirect);
1440         prelim_release(&preftrees.indirect_missing_keys);
1441
1442         if (ret < 0)
1443                 free_inode_elem_list(eie);
1444         return ret;
1445 }
1446
1447 /*
1448  * Finds all leafs with a reference to the specified combination of bytenr and
1449  * offset. key_list_head will point to a list of corresponding keys (caller must
1450  * free each list element). The leafs will be stored in the leafs ulist, which
1451  * must be freed with ulist_free.
1452  *
1453  * returns 0 on success, <0 on error
1454  */
1455 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1456                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1457                                 u64 time_seq, struct ulist **leafs,
1458                                 const u64 *extent_item_pos, bool ignore_offset)
1459 {
1460         int ret;
1461
1462         *leafs = ulist_alloc(GFP_NOFS);
1463         if (!*leafs)
1464                 return -ENOMEM;
1465
1466         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1467                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1468         if (ret < 0 && ret != -ENOENT) {
1469                 free_leaf_list(*leafs);
1470                 return ret;
1471         }
1472
1473         return 0;
1474 }
1475
1476 /*
1477  * walk all backrefs for a given extent to find all roots that reference this
1478  * extent. Walking a backref means finding all extents that reference this
1479  * extent and in turn walk the backrefs of those, too. Naturally this is a
1480  * recursive process, but here it is implemented in an iterative fashion: We
1481  * find all referencing extents for the extent in question and put them on a
1482  * list. In turn, we find all referencing extents for those, further appending
1483  * to the list. The way we iterate the list allows adding more elements after
1484  * the current while iterating. The process stops when we reach the end of the
1485  * list. Found roots are added to the roots list.
1486  *
1487  * returns 0 on success, < 0 on error.
1488  */
1489 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1490                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1491                                      u64 time_seq, struct ulist **roots,
1492                                      bool ignore_offset)
1493 {
1494         struct ulist *tmp;
1495         struct ulist_node *node = NULL;
1496         struct ulist_iterator uiter;
1497         int ret;
1498
1499         tmp = ulist_alloc(GFP_NOFS);
1500         if (!tmp)
1501                 return -ENOMEM;
1502         *roots = ulist_alloc(GFP_NOFS);
1503         if (!*roots) {
1504                 ulist_free(tmp);
1505                 return -ENOMEM;
1506         }
1507
1508         ULIST_ITER_INIT(&uiter);
1509         while (1) {
1510                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1511                                         tmp, *roots, NULL, NULL, ignore_offset);
1512                 if (ret < 0 && ret != -ENOENT) {
1513                         ulist_free(tmp);
1514                         ulist_free(*roots);
1515                         *roots = NULL;
1516                         return ret;
1517                 }
1518                 node = ulist_next(tmp, &uiter);
1519                 if (!node)
1520                         break;
1521                 bytenr = node->val;
1522                 cond_resched();
1523         }
1524
1525         ulist_free(tmp);
1526         return 0;
1527 }
1528
1529 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1530                          struct btrfs_fs_info *fs_info, u64 bytenr,
1531                          u64 time_seq, struct ulist **roots,
1532                          bool ignore_offset)
1533 {
1534         int ret;
1535
1536         if (!trans)
1537                 down_read(&fs_info->commit_root_sem);
1538         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1539                                         time_seq, roots, ignore_offset);
1540         if (!trans)
1541                 up_read(&fs_info->commit_root_sem);
1542         return ret;
1543 }
1544
1545 /**
1546  * btrfs_check_shared - tell us whether an extent is shared
1547  *
1548  * btrfs_check_shared uses the backref walking code but will short
1549  * circuit as soon as it finds a root or inode that doesn't match the
1550  * one passed in. This provides a significant performance benefit for
1551  * callers (such as fiemap) which want to know whether the extent is
1552  * shared but do not need a ref count.
1553  *
1554  * This attempts to attach to the running transaction in order to account for
1555  * delayed refs, but continues on even when no running transaction exists.
1556  *
1557  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1558  */
1559 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr,
1560                 struct ulist *roots, struct ulist *tmp)
1561 {
1562         struct btrfs_fs_info *fs_info = root->fs_info;
1563         struct btrfs_trans_handle *trans;
1564         struct ulist_iterator uiter;
1565         struct ulist_node *node;
1566         struct seq_list elem = SEQ_LIST_INIT(elem);
1567         int ret = 0;
1568         struct share_check shared = {
1569                 .root_objectid = root->root_key.objectid,
1570                 .inum = inum,
1571                 .share_count = 0,
1572                 .have_delayed_delete_refs = false,
1573         };
1574
1575         ulist_init(roots);
1576         ulist_init(tmp);
1577
1578         trans = btrfs_join_transaction_nostart(root);
1579         if (IS_ERR(trans)) {
1580                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1581                         ret = PTR_ERR(trans);
1582                         goto out;
1583                 }
1584                 trans = NULL;
1585                 down_read(&fs_info->commit_root_sem);
1586         } else {
1587                 btrfs_get_tree_mod_seq(fs_info, &elem);
1588         }
1589
1590         ULIST_ITER_INIT(&uiter);
1591         while (1) {
1592                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1593                                         roots, NULL, &shared, false);
1594                 if (ret == BACKREF_FOUND_SHARED) {
1595                         /* this is the only condition under which we return 1 */
1596                         ret = 1;
1597                         break;
1598                 }
1599                 if (ret < 0 && ret != -ENOENT)
1600                         break;
1601                 ret = 0;
1602                 node = ulist_next(tmp, &uiter);
1603                 if (!node)
1604                         break;
1605                 bytenr = node->val;
1606                 shared.share_count = 0;
1607                 shared.have_delayed_delete_refs = false;
1608                 cond_resched();
1609         }
1610
1611         if (trans) {
1612                 btrfs_put_tree_mod_seq(fs_info, &elem);
1613                 btrfs_end_transaction(trans);
1614         } else {
1615                 up_read(&fs_info->commit_root_sem);
1616         }
1617 out:
1618         ulist_release(roots);
1619         ulist_release(tmp);
1620         return ret;
1621 }
1622
1623 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1624                           u64 start_off, struct btrfs_path *path,
1625                           struct btrfs_inode_extref **ret_extref,
1626                           u64 *found_off)
1627 {
1628         int ret, slot;
1629         struct btrfs_key key;
1630         struct btrfs_key found_key;
1631         struct btrfs_inode_extref *extref;
1632         const struct extent_buffer *leaf;
1633         unsigned long ptr;
1634
1635         key.objectid = inode_objectid;
1636         key.type = BTRFS_INODE_EXTREF_KEY;
1637         key.offset = start_off;
1638
1639         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1640         if (ret < 0)
1641                 return ret;
1642
1643         while (1) {
1644                 leaf = path->nodes[0];
1645                 slot = path->slots[0];
1646                 if (slot >= btrfs_header_nritems(leaf)) {
1647                         /*
1648                          * If the item at offset is not found,
1649                          * btrfs_search_slot will point us to the slot
1650                          * where it should be inserted. In our case
1651                          * that will be the slot directly before the
1652                          * next INODE_REF_KEY_V2 item. In the case
1653                          * that we're pointing to the last slot in a
1654                          * leaf, we must move one leaf over.
1655                          */
1656                         ret = btrfs_next_leaf(root, path);
1657                         if (ret) {
1658                                 if (ret >= 1)
1659                                         ret = -ENOENT;
1660                                 break;
1661                         }
1662                         continue;
1663                 }
1664
1665                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1666
1667                 /*
1668                  * Check that we're still looking at an extended ref key for
1669                  * this particular objectid. If we have different
1670                  * objectid or type then there are no more to be found
1671                  * in the tree and we can exit.
1672                  */
1673                 ret = -ENOENT;
1674                 if (found_key.objectid != inode_objectid)
1675                         break;
1676                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1677                         break;
1678
1679                 ret = 0;
1680                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1681                 extref = (struct btrfs_inode_extref *)ptr;
1682                 *ret_extref = extref;
1683                 if (found_off)
1684                         *found_off = found_key.offset;
1685                 break;
1686         }
1687
1688         return ret;
1689 }
1690
1691 /*
1692  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1693  * Elements of the path are separated by '/' and the path is guaranteed to be
1694  * 0-terminated. the path is only given within the current file system.
1695  * Therefore, it never starts with a '/'. the caller is responsible to provide
1696  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1697  * the start point of the resulting string is returned. this pointer is within
1698  * dest, normally.
1699  * in case the path buffer would overflow, the pointer is decremented further
1700  * as if output was written to the buffer, though no more output is actually
1701  * generated. that way, the caller can determine how much space would be
1702  * required for the path to fit into the buffer. in that case, the returned
1703  * value will be smaller than dest. callers must check this!
1704  */
1705 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1706                         u32 name_len, unsigned long name_off,
1707                         struct extent_buffer *eb_in, u64 parent,
1708                         char *dest, u32 size)
1709 {
1710         int slot;
1711         u64 next_inum;
1712         int ret;
1713         s64 bytes_left = ((s64)size) - 1;
1714         struct extent_buffer *eb = eb_in;
1715         struct btrfs_key found_key;
1716         int leave_spinning = path->leave_spinning;
1717         struct btrfs_inode_ref *iref;
1718
1719         if (bytes_left >= 0)
1720                 dest[bytes_left] = '\0';
1721
1722         path->leave_spinning = 1;
1723         while (1) {
1724                 bytes_left -= name_len;
1725                 if (bytes_left >= 0)
1726                         read_extent_buffer(eb, dest + bytes_left,
1727                                            name_off, name_len);
1728                 if (eb != eb_in) {
1729                         if (!path->skip_locking)
1730                                 btrfs_tree_read_unlock_blocking(eb);
1731                         free_extent_buffer(eb);
1732                 }
1733                 ret = btrfs_find_item(fs_root, path, parent, 0,
1734                                 BTRFS_INODE_REF_KEY, &found_key);
1735                 if (ret > 0)
1736                         ret = -ENOENT;
1737                 if (ret)
1738                         break;
1739
1740                 next_inum = found_key.offset;
1741
1742                 /* regular exit ahead */
1743                 if (parent == next_inum)
1744                         break;
1745
1746                 slot = path->slots[0];
1747                 eb = path->nodes[0];
1748                 /* make sure we can use eb after releasing the path */
1749                 if (eb != eb_in) {
1750                         if (!path->skip_locking)
1751                                 btrfs_set_lock_blocking_read(eb);
1752                         path->nodes[0] = NULL;
1753                         path->locks[0] = 0;
1754                 }
1755                 btrfs_release_path(path);
1756                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1757
1758                 name_len = btrfs_inode_ref_name_len(eb, iref);
1759                 name_off = (unsigned long)(iref + 1);
1760
1761                 parent = next_inum;
1762                 --bytes_left;
1763                 if (bytes_left >= 0)
1764                         dest[bytes_left] = '/';
1765         }
1766
1767         btrfs_release_path(path);
1768         path->leave_spinning = leave_spinning;
1769
1770         if (ret)
1771                 return ERR_PTR(ret);
1772
1773         return dest + bytes_left;
1774 }
1775
1776 /*
1777  * this makes the path point to (logical EXTENT_ITEM *)
1778  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1779  * tree blocks and <0 on error.
1780  */
1781 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1782                         struct btrfs_path *path, struct btrfs_key *found_key,
1783                         u64 *flags_ret)
1784 {
1785         int ret;
1786         u64 flags;
1787         u64 size = 0;
1788         u32 item_size;
1789         const struct extent_buffer *eb;
1790         struct btrfs_extent_item *ei;
1791         struct btrfs_key key;
1792
1793         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1794                 key.type = BTRFS_METADATA_ITEM_KEY;
1795         else
1796                 key.type = BTRFS_EXTENT_ITEM_KEY;
1797         key.objectid = logical;
1798         key.offset = (u64)-1;
1799
1800         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1801         if (ret < 0)
1802                 return ret;
1803
1804         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1805         if (ret) {
1806                 if (ret > 0)
1807                         ret = -ENOENT;
1808                 return ret;
1809         }
1810         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1811         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1812                 size = fs_info->nodesize;
1813         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1814                 size = found_key->offset;
1815
1816         if (found_key->objectid > logical ||
1817             found_key->objectid + size <= logical) {
1818                 btrfs_debug(fs_info,
1819                         "logical %llu is not within any extent", logical);
1820                 return -ENOENT;
1821         }
1822
1823         eb = path->nodes[0];
1824         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1825         BUG_ON(item_size < sizeof(*ei));
1826
1827         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1828         flags = btrfs_extent_flags(eb, ei);
1829
1830         btrfs_debug(fs_info,
1831                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1832                  logical, logical - found_key->objectid, found_key->objectid,
1833                  found_key->offset, flags, item_size);
1834
1835         WARN_ON(!flags_ret);
1836         if (flags_ret) {
1837                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1838                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1839                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1840                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1841                 else
1842                         BUG();
1843                 return 0;
1844         }
1845
1846         return -EIO;
1847 }
1848
1849 /*
1850  * helper function to iterate extent inline refs. ptr must point to a 0 value
1851  * for the first call and may be modified. it is used to track state.
1852  * if more refs exist, 0 is returned and the next call to
1853  * get_extent_inline_ref must pass the modified ptr parameter to get the
1854  * next ref. after the last ref was processed, 1 is returned.
1855  * returns <0 on error
1856  */
1857 static int get_extent_inline_ref(unsigned long *ptr,
1858                                  const struct extent_buffer *eb,
1859                                  const struct btrfs_key *key,
1860                                  const struct btrfs_extent_item *ei,
1861                                  u32 item_size,
1862                                  struct btrfs_extent_inline_ref **out_eiref,
1863                                  int *out_type)
1864 {
1865         unsigned long end;
1866         u64 flags;
1867         struct btrfs_tree_block_info *info;
1868
1869         if (!*ptr) {
1870                 /* first call */
1871                 flags = btrfs_extent_flags(eb, ei);
1872                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1873                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1874                                 /* a skinny metadata extent */
1875                                 *out_eiref =
1876                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1877                         } else {
1878                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1879                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1880                                 *out_eiref =
1881                                    (struct btrfs_extent_inline_ref *)(info + 1);
1882                         }
1883                 } else {
1884                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1885                 }
1886                 *ptr = (unsigned long)*out_eiref;
1887                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1888                         return -ENOENT;
1889         }
1890
1891         end = (unsigned long)ei + item_size;
1892         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1893         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1894                                                      BTRFS_REF_TYPE_ANY);
1895         if (*out_type == BTRFS_REF_TYPE_INVALID)
1896                 return -EUCLEAN;
1897
1898         *ptr += btrfs_extent_inline_ref_size(*out_type);
1899         WARN_ON(*ptr > end);
1900         if (*ptr == end)
1901                 return 1; /* last */
1902
1903         return 0;
1904 }
1905
1906 /*
1907  * reads the tree block backref for an extent. tree level and root are returned
1908  * through out_level and out_root. ptr must point to a 0 value for the first
1909  * call and may be modified (see get_extent_inline_ref comment).
1910  * returns 0 if data was provided, 1 if there was no more data to provide or
1911  * <0 on error.
1912  */
1913 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1914                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1915                             u32 item_size, u64 *out_root, u8 *out_level)
1916 {
1917         int ret;
1918         int type;
1919         struct btrfs_extent_inline_ref *eiref;
1920
1921         if (*ptr == (unsigned long)-1)
1922                 return 1;
1923
1924         while (1) {
1925                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1926                                               &eiref, &type);
1927                 if (ret < 0)
1928                         return ret;
1929
1930                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1931                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1932                         break;
1933
1934                 if (ret == 1)
1935                         return 1;
1936         }
1937
1938         /* we can treat both ref types equally here */
1939         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1940
1941         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1942                 struct btrfs_tree_block_info *info;
1943
1944                 info = (struct btrfs_tree_block_info *)(ei + 1);
1945                 *out_level = btrfs_tree_block_level(eb, info);
1946         } else {
1947                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1948                 *out_level = (u8)key->offset;
1949         }
1950
1951         if (ret == 1)
1952                 *ptr = (unsigned long)-1;
1953
1954         return 0;
1955 }
1956
1957 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1958                              struct extent_inode_elem *inode_list,
1959                              u64 root, u64 extent_item_objectid,
1960                              iterate_extent_inodes_t *iterate, void *ctx)
1961 {
1962         struct extent_inode_elem *eie;
1963         int ret = 0;
1964
1965         for (eie = inode_list; eie; eie = eie->next) {
1966                 btrfs_debug(fs_info,
1967                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1968                             extent_item_objectid, eie->inum,
1969                             eie->offset, root);
1970                 ret = iterate(eie->inum, eie->offset, root, ctx);
1971                 if (ret) {
1972                         btrfs_debug(fs_info,
1973                                     "stopping iteration for %llu due to ret=%d",
1974                                     extent_item_objectid, ret);
1975                         break;
1976                 }
1977         }
1978
1979         return ret;
1980 }
1981
1982 /*
1983  * calls iterate() for every inode that references the extent identified by
1984  * the given parameters.
1985  * when the iterator function returns a non-zero value, iteration stops.
1986  */
1987 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1988                                 u64 extent_item_objectid, u64 extent_item_pos,
1989                                 int search_commit_root,
1990                                 iterate_extent_inodes_t *iterate, void *ctx,
1991                                 bool ignore_offset)
1992 {
1993         int ret;
1994         struct btrfs_trans_handle *trans = NULL;
1995         struct ulist *refs = NULL;
1996         struct ulist *roots = NULL;
1997         struct ulist_node *ref_node = NULL;
1998         struct ulist_node *root_node = NULL;
1999         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
2000         struct ulist_iterator ref_uiter;
2001         struct ulist_iterator root_uiter;
2002
2003         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
2004                         extent_item_objectid);
2005
2006         if (!search_commit_root) {
2007                 trans = btrfs_attach_transaction(fs_info->extent_root);
2008                 if (IS_ERR(trans)) {
2009                         if (PTR_ERR(trans) != -ENOENT &&
2010                             PTR_ERR(trans) != -EROFS)
2011                                 return PTR_ERR(trans);
2012                         trans = NULL;
2013                 }
2014         }
2015
2016         if (trans)
2017                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2018         else
2019                 down_read(&fs_info->commit_root_sem);
2020
2021         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
2022                                    tree_mod_seq_elem.seq, &refs,
2023                                    &extent_item_pos, ignore_offset);
2024         if (ret)
2025                 goto out;
2026
2027         ULIST_ITER_INIT(&ref_uiter);
2028         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2029                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
2030                                                 tree_mod_seq_elem.seq, &roots,
2031                                                 ignore_offset);
2032                 if (ret)
2033                         break;
2034                 ULIST_ITER_INIT(&root_uiter);
2035                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
2036                         btrfs_debug(fs_info,
2037                                     "root %llu references leaf %llu, data list %#llx",
2038                                     root_node->val, ref_node->val,
2039                                     ref_node->aux);
2040                         ret = iterate_leaf_refs(fs_info,
2041                                                 (struct extent_inode_elem *)
2042                                                 (uintptr_t)ref_node->aux,
2043                                                 root_node->val,
2044                                                 extent_item_objectid,
2045                                                 iterate, ctx);
2046                 }
2047                 ulist_free(roots);
2048         }
2049
2050         free_leaf_list(refs);
2051 out:
2052         if (trans) {
2053                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2054                 btrfs_end_transaction(trans);
2055         } else {
2056                 up_read(&fs_info->commit_root_sem);
2057         }
2058
2059         return ret;
2060 }
2061
2062 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2063                                 struct btrfs_path *path,
2064                                 iterate_extent_inodes_t *iterate, void *ctx,
2065                                 bool ignore_offset)
2066 {
2067         int ret;
2068         u64 extent_item_pos;
2069         u64 flags = 0;
2070         struct btrfs_key found_key;
2071         int search_commit_root = path->search_commit_root;
2072
2073         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2074         btrfs_release_path(path);
2075         if (ret < 0)
2076                 return ret;
2077         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2078                 return -EINVAL;
2079
2080         extent_item_pos = logical - found_key.objectid;
2081         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2082                                         extent_item_pos, search_commit_root,
2083                                         iterate, ctx, ignore_offset);
2084
2085         return ret;
2086 }
2087
2088 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2089                               struct extent_buffer *eb, void *ctx);
2090
2091 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2092                               struct btrfs_path *path,
2093                               iterate_irefs_t *iterate, void *ctx)
2094 {
2095         int ret = 0;
2096         int slot;
2097         u32 cur;
2098         u32 len;
2099         u32 name_len;
2100         u64 parent = 0;
2101         int found = 0;
2102         struct extent_buffer *eb;
2103         struct btrfs_item *item;
2104         struct btrfs_inode_ref *iref;
2105         struct btrfs_key found_key;
2106
2107         while (!ret) {
2108                 ret = btrfs_find_item(fs_root, path, inum,
2109                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2110                                 &found_key);
2111
2112                 if (ret < 0)
2113                         break;
2114                 if (ret) {
2115                         ret = found ? 0 : -ENOENT;
2116                         break;
2117                 }
2118                 ++found;
2119
2120                 parent = found_key.offset;
2121                 slot = path->slots[0];
2122                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2123                 if (!eb) {
2124                         ret = -ENOMEM;
2125                         break;
2126                 }
2127                 btrfs_release_path(path);
2128
2129                 item = btrfs_item_nr(slot);
2130                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2131
2132                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2133                         name_len = btrfs_inode_ref_name_len(eb, iref);
2134                         /* path must be released before calling iterate()! */
2135                         btrfs_debug(fs_root->fs_info,
2136                                 "following ref at offset %u for inode %llu in tree %llu",
2137                                 cur, found_key.objectid,
2138                                 fs_root->root_key.objectid);
2139                         ret = iterate(parent, name_len,
2140                                       (unsigned long)(iref + 1), eb, ctx);
2141                         if (ret)
2142                                 break;
2143                         len = sizeof(*iref) + name_len;
2144                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2145                 }
2146                 free_extent_buffer(eb);
2147         }
2148
2149         btrfs_release_path(path);
2150
2151         return ret;
2152 }
2153
2154 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2155                                  struct btrfs_path *path,
2156                                  iterate_irefs_t *iterate, void *ctx)
2157 {
2158         int ret;
2159         int slot;
2160         u64 offset = 0;
2161         u64 parent;
2162         int found = 0;
2163         struct extent_buffer *eb;
2164         struct btrfs_inode_extref *extref;
2165         u32 item_size;
2166         u32 cur_offset;
2167         unsigned long ptr;
2168
2169         while (1) {
2170                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2171                                             &offset);
2172                 if (ret < 0)
2173                         break;
2174                 if (ret) {
2175                         ret = found ? 0 : -ENOENT;
2176                         break;
2177                 }
2178                 ++found;
2179
2180                 slot = path->slots[0];
2181                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2182                 if (!eb) {
2183                         ret = -ENOMEM;
2184                         break;
2185                 }
2186                 btrfs_release_path(path);
2187
2188                 item_size = btrfs_item_size_nr(eb, slot);
2189                 ptr = btrfs_item_ptr_offset(eb, slot);
2190                 cur_offset = 0;
2191
2192                 while (cur_offset < item_size) {
2193                         u32 name_len;
2194
2195                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2196                         parent = btrfs_inode_extref_parent(eb, extref);
2197                         name_len = btrfs_inode_extref_name_len(eb, extref);
2198                         ret = iterate(parent, name_len,
2199                                       (unsigned long)&extref->name, eb, ctx);
2200                         if (ret)
2201                                 break;
2202
2203                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2204                         cur_offset += sizeof(*extref);
2205                 }
2206                 free_extent_buffer(eb);
2207
2208                 offset++;
2209         }
2210
2211         btrfs_release_path(path);
2212
2213         return ret;
2214 }
2215
2216 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2217                          struct btrfs_path *path, iterate_irefs_t *iterate,
2218                          void *ctx)
2219 {
2220         int ret;
2221         int found_refs = 0;
2222
2223         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2224         if (!ret)
2225                 ++found_refs;
2226         else if (ret != -ENOENT)
2227                 return ret;
2228
2229         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2230         if (ret == -ENOENT && found_refs)
2231                 return 0;
2232
2233         return ret;
2234 }
2235
2236 /*
2237  * returns 0 if the path could be dumped (probably truncated)
2238  * returns <0 in case of an error
2239  */
2240 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2241                          struct extent_buffer *eb, void *ctx)
2242 {
2243         struct inode_fs_paths *ipath = ctx;
2244         char *fspath;
2245         char *fspath_min;
2246         int i = ipath->fspath->elem_cnt;
2247         const int s_ptr = sizeof(char *);
2248         u32 bytes_left;
2249
2250         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2251                                         ipath->fspath->bytes_left - s_ptr : 0;
2252
2253         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2254         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2255                                    name_off, eb, inum, fspath_min, bytes_left);
2256         if (IS_ERR(fspath))
2257                 return PTR_ERR(fspath);
2258
2259         if (fspath > fspath_min) {
2260                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2261                 ++ipath->fspath->elem_cnt;
2262                 ipath->fspath->bytes_left = fspath - fspath_min;
2263         } else {
2264                 ++ipath->fspath->elem_missed;
2265                 ipath->fspath->bytes_missing += fspath_min - fspath;
2266                 ipath->fspath->bytes_left = 0;
2267         }
2268
2269         return 0;
2270 }
2271
2272 /*
2273  * this dumps all file system paths to the inode into the ipath struct, provided
2274  * is has been created large enough. each path is zero-terminated and accessed
2275  * from ipath->fspath->val[i].
2276  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2277  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2278  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2279  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2280  * have been needed to return all paths.
2281  */
2282 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2283 {
2284         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2285                              inode_to_path, ipath);
2286 }
2287
2288 struct btrfs_data_container *init_data_container(u32 total_bytes)
2289 {
2290         struct btrfs_data_container *data;
2291         size_t alloc_bytes;
2292
2293         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2294         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2295         if (!data)
2296                 return ERR_PTR(-ENOMEM);
2297
2298         if (total_bytes >= sizeof(*data)) {
2299                 data->bytes_left = total_bytes - sizeof(*data);
2300                 data->bytes_missing = 0;
2301         } else {
2302                 data->bytes_missing = sizeof(*data) - total_bytes;
2303                 data->bytes_left = 0;
2304         }
2305
2306         data->elem_cnt = 0;
2307         data->elem_missed = 0;
2308
2309         return data;
2310 }
2311
2312 /*
2313  * allocates space to return multiple file system paths for an inode.
2314  * total_bytes to allocate are passed, note that space usable for actual path
2315  * information will be total_bytes - sizeof(struct inode_fs_paths).
2316  * the returned pointer must be freed with free_ipath() in the end.
2317  */
2318 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2319                                         struct btrfs_path *path)
2320 {
2321         struct inode_fs_paths *ifp;
2322         struct btrfs_data_container *fspath;
2323
2324         fspath = init_data_container(total_bytes);
2325         if (IS_ERR(fspath))
2326                 return ERR_CAST(fspath);
2327
2328         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2329         if (!ifp) {
2330                 kvfree(fspath);
2331                 return ERR_PTR(-ENOMEM);
2332         }
2333
2334         ifp->btrfs_path = path;
2335         ifp->fspath = fspath;
2336         ifp->fs_root = fs_root;
2337
2338         return ifp;
2339 }
2340
2341 void free_ipath(struct inode_fs_paths *ipath)
2342 {
2343         if (!ipath)
2344                 return;
2345         kvfree(ipath->fspath);
2346         kfree(ipath);
2347 }