fs/btrfs/backref.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/mm.h>
   7 #include <linux/rbtree.h>
   8 #include <trace/events/btrfs.h>
   9 #include "ctree.h"
  10 #include "disk-io.h"
  11 #include "backref.h"
  12 #include "ulist.h"
  13 #include "transaction.h"
  14 #include "delayed-ref.h"
  15 #include "locking.h"
  16
  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 root;
 116         unsigned int count;
 117 };
 118
 119 #define PREFTREE_INIT   { .root = RB_ROOT, .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 };
 140
 141 static inline int extent_is_shared(struct share_check *sc)
 142 {
 143         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 144 }
 145
 146 static struct kmem_cache *btrfs_prelim_ref_cache;
 147
 148 int __init btrfs_prelim_ref_init(void)
 149 {
 150         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 151                                         sizeof(struct prelim_ref),
 152                                         0,
 153                                         SLAB_MEM_SPREAD,
 154                                         NULL);
 155         if (!btrfs_prelim_ref_cache)
 156                 return -ENOMEM;
 157         return 0;
 158 }
 159
 160 void __cold btrfs_prelim_ref_exit(void)
 161 {
 162         kmem_cache_destroy(btrfs_prelim_ref_cache);
 163 }
 164
 165 static void free_pref(struct prelim_ref *ref)
 166 {
 167         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 168 }
 169
 170 /*
 171  * Return 0 when both refs are for the same block (and can be merged).
 172  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 173  * indicates a 'higher' block.
 174  */
 175 static int prelim_ref_compare(struct prelim_ref *ref1,
 176                               struct prelim_ref *ref2)
 177 {
 178         if (ref1->level < ref2->level)
 179                 return -1;
 180         if (ref1->level > ref2->level)
 181                 return 1;
 182         if (ref1->root_id < ref2->root_id)
 183                 return -1;
 184         if (ref1->root_id > ref2->root_id)
 185                 return 1;
 186         if (ref1->key_for_search.type < ref2->key_for_search.type)
 187                 return -1;
 188         if (ref1->key_for_search.type > ref2->key_for_search.type)
 189                 return 1;
 190         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 191                 return -1;
 192         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 193                 return 1;
 194         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 195                 return -1;
 196         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 197                 return 1;
 198         if (ref1->parent < ref2->parent)
 199                 return -1;
 200         if (ref1->parent > ref2->parent)
 201                 return 1;
 202
 203         return 0;
 204 }
 205
 206 static void update_share_count(struct share_check *sc, int oldcount,
 207                                int newcount)
 208 {
 209         if ((!sc) || (oldcount == 0 && newcount < 1))
 210                 return;
 211
 212         if (oldcount > 0 && newcount < 1)
 213                 sc->share_count--;
 214         else if (oldcount < 1 && newcount > 0)
 215                 sc->share_count++;
 216 }
 217
 218 /*
 219  * Add @newref to the @root rbtree, merging identical refs.
 220  *
 221  * Callers should assume that newref has been freed after calling.
 222  */
 223 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 224                               struct preftree *preftree,
 225                               struct prelim_ref *newref,
 226                               struct share_check *sc)
 227 {
 228         struct rb_root *root;
 229         struct rb_node **p;
 230         struct rb_node *parent = NULL;
 231         struct prelim_ref *ref;
 232         int result;
 233
 234         root = &preftree->root;
 235         p = &root->rb_node;
 236
 237         while (*p) {
 238                 parent = *p;
 239                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 240                 result = prelim_ref_compare(ref, newref);
 241                 if (result < 0) {
 242                         p = &(*p)->rb_left;
 243                 } else if (result > 0) {
 244                         p = &(*p)->rb_right;
 245                 } else {
 246                         /* Identical refs, merge them and free @newref */
 247                         struct extent_inode_elem *eie = ref->inode_list;
 248
 249                         while (eie && eie->next)
 250                                 eie = eie->next;
 251
 252                         if (!eie)
 253                                 ref->inode_list = newref->inode_list;
 254                         else
 255                                 eie->next = newref->inode_list;
 256                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 257                                                      preftree->count);
 258                         /*
 259                          * A delayed ref can have newref->count < 0.
 260                          * The ref->count is updated to follow any
 261                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 262                          */
 263                         update_share_count(sc, ref->count,
 264                                            ref->count + newref->count);
 265                         ref->count += newref->count;
 266                         free_pref(newref);
 267                         return;
 268                 }
 269         }
 270
 271         update_share_count(sc, 0, newref->count);
 272         preftree->count++;
 273         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 274         rb_link_node(&newref->rbnode, parent, p);
 275         rb_insert_color(&newref->rbnode, root);
 276 }
 277
 278 /*
 279  * Release the entire tree.  We don't care about internal consistency so
 280  * just free everything and then reset the tree root.
 281  */
 282 static void prelim_release(struct preftree *preftree)
 283 {
 284         struct prelim_ref *ref, *next_ref;
 285
 286         rbtree_postorder_for_each_entry_safe(ref, next_ref, &preftree->root,
 287                                              rbnode)
 288                 free_pref(ref);
 289
 290         preftree->root = RB_ROOT;
 291         preftree->count = 0;
 292 }
 293
 294 /*
 295  * the rules for all callers of this function are:
 296  * - obtaining the parent is the goal
 297  * - if you add a key, you must know that it is a correct key
 298  * - if you cannot add the parent or a correct key, then we will look into the
 299  *   block later to set a correct key
 300  *
 301  * delayed refs
 302  * ============
 303  *        backref type | shared | indirect | shared | indirect
 304  * information         |   tree |     tree |   data |     data
 305  * --------------------+--------+----------+--------+----------
 306  *      parent logical |    y   |     -    |    -   |     -
 307  *      key to resolve |    -   |     y    |    y   |     y
 308  *  tree block logical |    -   |     -    |    -   |     -
 309  *  root for resolving |    y   |     y    |    y   |     y
 310  *
 311  * - column 1:       we've the parent -> done
 312  * - column 2, 3, 4: we use the key to find the parent
 313  *
 314  * on disk refs (inline or keyed)
 315  * ==============================
 316  *        backref type | shared | indirect | shared | indirect
 317  * information         |   tree |     tree |   data |     data
 318  * --------------------+--------+----------+--------+----------
 319  *      parent logical |    y   |     -    |    y   |     -
 320  *      key to resolve |    -   |     -    |    -   |     y
 321  *  tree block logical |    y   |     y    |    y   |     y
 322  *  root for resolving |    -   |     y    |    y   |     y
 323  *
 324  * - column 1, 3: we've the parent -> done
 325  * - column 2:    we take the first key from the block to find the parent
 326  *                (see add_missing_keys)
 327  * - column 4:    we use the key to find the parent
 328  *
 329  * additional information that's available but not required to find the parent
 330  * block might help in merging entries to gain some speed.
 331  */
 332 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 333                           struct preftree *preftree, u64 root_id,
 334                           const struct btrfs_key *key, int level, u64 parent,
 335                           u64 wanted_disk_byte, int count,
 336                           struct share_check *sc, gfp_t gfp_mask)
 337 {
 338         struct prelim_ref *ref;
 339
 340         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 341                 return 0;
 342
 343         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 344         if (!ref)
 345                 return -ENOMEM;
 346
 347         ref->root_id = root_id;
 348         if (key) {
 349                 ref->key_for_search = *key;
 350                 /*
 351                  * We can often find data backrefs with an offset that is too
 352                  * large (>= LLONG_MAX, maximum allowed file offset) due to
 353                  * underflows when subtracting a file's offset with the data
 354                  * offset of its corresponding extent data item. This can
 355                  * happen for example in the clone ioctl.
 356                  * So if we detect such case we set the search key's offset to
 357                  * zero to make sure we will find the matching file extent item
 358                  * at add_all_parents(), otherwise we will miss it because the
 359                  * offset taken form the backref is much larger then the offset
 360                  * of the file extent item. This can make us scan a very large
 361                  * number of file extent items, but at least it will not make
 362                  * us miss any.
 363                  * This is an ugly workaround for a behaviour that should have
 364                  * never existed, but it does and a fix for the clone ioctl
 365                  * would touch a lot of places, cause backwards incompatibility
 366                  * and would not fix the problem for extents cloned with older
 367                  * kernels.
 368                  */
 369                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
 370                     ref->key_for_search.offset >= LLONG_MAX)
 371                         ref->key_for_search.offset = 0;
 372         } else {
 373                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 374         }
 375
 376         ref->inode_list = NULL;
 377         ref->level = level;
 378         ref->count = count;
 379         ref->parent = parent;
 380         ref->wanted_disk_byte = wanted_disk_byte;
 381         prelim_ref_insert(fs_info, preftree, ref, sc);
 382         return extent_is_shared(sc);
 383 }
 384
 385 /* direct refs use root == 0, key == NULL */
 386 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 387                           struct preftrees *preftrees, int level, u64 parent,
 388                           u64 wanted_disk_byte, int count,
 389                           struct share_check *sc, gfp_t gfp_mask)
 390 {
 391         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 392                               parent, wanted_disk_byte, count, sc, gfp_mask);
 393 }
 394
 395 /* indirect refs use parent == 0 */
 396 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 397                             struct preftrees *preftrees, u64 root_id,
 398                             const struct btrfs_key *key, int level,
 399                             u64 wanted_disk_byte, int count,
 400                             struct share_check *sc, gfp_t gfp_mask)
 401 {
 402         struct preftree *tree = &preftrees->indirect;
 403
 404         if (!key)
 405                 tree = &preftrees->indirect_missing_keys;
 406         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 407                               wanted_disk_byte, count, sc, gfp_mask);
 408 }
 409
 410 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 411                            struct ulist *parents, struct prelim_ref *ref,
 412                            int level, u64 time_seq, const u64 *extent_item_pos,
 413                            u64 total_refs, bool ignore_offset)
 414 {
 415         int ret = 0;
 416         int slot;
 417         struct extent_buffer *eb;
 418         struct btrfs_key key;
 419         struct btrfs_key *key_for_search = &ref->key_for_search;
 420         struct btrfs_file_extent_item *fi;
 421         struct extent_inode_elem *eie = NULL, *old = NULL;
 422         u64 disk_byte;
 423         u64 wanted_disk_byte = ref->wanted_disk_byte;
 424         u64 count = 0;
 425
 426         if (level != 0) {
 427                 eb = path->nodes[level];
 428                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 429                 if (ret < 0)
 430                         return ret;
 431                 return 0;
 432         }
 433
 434         /*
 435          * We normally enter this function with the path already pointing to
 436          * the first item to check. But sometimes, we may enter it with
 437          * slot==nritems. In that case, go to the next leaf before we continue.
 438          */
 439         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 440                 if (time_seq == SEQ_LAST)
 441                         ret = btrfs_next_leaf(root, path);
 442                 else
 443                         ret = btrfs_next_old_leaf(root, path, time_seq);
 444         }
 445
 446         while (!ret && count < total_refs) {
 447                 eb = path->nodes[0];
 448                 slot = path->slots[0];
 449
 450                 btrfs_item_key_to_cpu(eb, &key, slot);
 451
 452                 if (key.objectid != key_for_search->objectid ||
 453                     key.type != BTRFS_EXTENT_DATA_KEY)
 454                         break;
 455
 456                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 457                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 458
 459                 if (disk_byte == wanted_disk_byte) {
 460                         eie = NULL;
 461                         old = NULL;
 462                         count++;
 463                         if (extent_item_pos) {
 464                                 ret = check_extent_in_eb(&key, eb, fi,
 465                                                 *extent_item_pos,
 466                                                 &eie, ignore_offset);
 467                                 if (ret < 0)
 468                                         break;
 469                         }
 470                         if (ret > 0)
 471                                 goto next;
 472                         ret = ulist_add_merge_ptr(parents, eb->start,
 473                                                   eie, (void **)&old, GFP_NOFS);
 474                         if (ret < 0)
 475                                 break;
 476                         if (!ret && extent_item_pos) {
 477                                 while (old->next)
 478                                         old = old->next;
 479                                 old->next = eie;
 480                         }
 481                         eie = NULL;
 482                 }
 483 next:
 484                 if (time_seq == SEQ_LAST)
 485                         ret = btrfs_next_item(root, path);
 486                 else
 487                         ret = btrfs_next_old_item(root, path, time_seq);
 488         }
 489
 490         if (ret > 0)
 491                 ret = 0;
 492         else if (ret < 0)
 493                 free_inode_elem_list(eie);
 494         return ret;
 495 }
 496
 497 /*
 498  * resolve an indirect backref in the form (root_id, key, level)
 499  * to a logical address
 500  */
 501 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 502                                 struct btrfs_path *path, u64 time_seq,
 503                                 struct prelim_ref *ref, struct ulist *parents,
 504                                 const u64 *extent_item_pos, u64 total_refs,
 505                                 bool ignore_offset)
 506 {
 507         struct btrfs_root *root;
 508         struct btrfs_key root_key;
 509         struct extent_buffer *eb;
 510         int ret = 0;
 511         int root_level;
 512         int level = ref->level;
 513         int index;
 514
 515         root_key.objectid = ref->root_id;
 516         root_key.type = BTRFS_ROOT_ITEM_KEY;
 517         root_key.offset = (u64)-1;
 518
 519         index = srcu_read_lock(&fs_info->subvol_srcu);
 520
 521         root = btrfs_get_fs_root(fs_info, &root_key, false);
 522         if (IS_ERR(root)) {
 523                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 524                 ret = PTR_ERR(root);
 525                 goto out;
 526         }
 527
 528         if (btrfs_is_testing(fs_info)) {
 529                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 530                 ret = -ENOENT;
 531                 goto out;
 532         }
 533
 534         if (path->search_commit_root)
 535                 root_level = btrfs_header_level(root->commit_root);
 536         else if (time_seq == SEQ_LAST)
 537                 root_level = btrfs_header_level(root->node);
 538         else
 539                 root_level = btrfs_old_root_level(root, time_seq);
 540
 541         if (root_level + 1 == level) {
 542                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 543                 goto out;
 544         }
 545
 546         path->lowest_level = level;
 547         if (time_seq == SEQ_LAST)
 548                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
 549                                         0, 0);
 550         else
 551                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
 552                                             time_seq);
 553
 554         /* root node has been locked, we can release @subvol_srcu safely here */
 555         srcu_read_unlock(&fs_info->subvol_srcu, index);
 556
 557         btrfs_debug(fs_info,
 558                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 559                  ref->root_id, level, ref->count, ret,
 560                  ref->key_for_search.objectid, ref->key_for_search.type,
 561                  ref->key_for_search.offset);
 562         if (ret < 0)
 563                 goto out;
 564
 565         eb = path->nodes[level];
 566         while (!eb) {
 567                 if (WARN_ON(!level)) {
 568                         ret = 1;
 569                         goto out;
 570                 }
 571                 level--;
 572                 eb = path->nodes[level];
 573         }
 574
 575         ret = add_all_parents(root, path, parents, ref, level, time_seq,
 576                               extent_item_pos, total_refs, ignore_offset);
 577 out:
 578         path->lowest_level = 0;
 579         btrfs_release_path(path);
 580         return ret;
 581 }
 582
 583 static struct extent_inode_elem *
 584 unode_aux_to_inode_list(struct ulist_node *node)
 585 {
 586         if (!node)
 587                 return NULL;
 588         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 589 }
 590
 591 /*
 592  * We maintain three seperate rbtrees: one for direct refs, one for
 593  * indirect refs which have a key, and one for indirect refs which do not
 594  * have a key. Each tree does merge on insertion.
 595  *
 596  * Once all of the references are located, we iterate over the tree of
 597  * indirect refs with missing keys. An appropriate key is located and
 598  * the ref is moved onto the tree for indirect refs. After all missing
 599  * keys are thus located, we iterate over the indirect ref tree, resolve
 600  * each reference, and then insert the resolved reference onto the
 601  * direct tree (merging there too).
 602  *
 603  * New backrefs (i.e., for parent nodes) are added to the appropriate
 604  * rbtree as they are encountered. The new backrefs are subsequently
 605  * resolved as above.
 606  */
 607 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 608                                  struct btrfs_path *path, u64 time_seq,
 609                                  struct preftrees *preftrees,
 610                                  const u64 *extent_item_pos, u64 total_refs,
 611                                  struct share_check *sc, bool ignore_offset)
 612 {
 613         int err;
 614         int ret = 0;
 615         struct ulist *parents;
 616         struct ulist_node *node;
 617         struct ulist_iterator uiter;
 618         struct rb_node *rnode;
 619
 620         parents = ulist_alloc(GFP_NOFS);
 621         if (!parents)
 622                 return -ENOMEM;
 623
 624         /*
 625          * We could trade memory usage for performance here by iterating
 626          * the tree, allocating new refs for each insertion, and then
 627          * freeing the entire indirect tree when we're done.  In some test
 628          * cases, the tree can grow quite large (~200k objects).
 629          */
 630         while ((rnode = rb_first(&preftrees->indirect.root))) {
 631                 struct prelim_ref *ref;
 632
 633                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 634                 if (WARN(ref->parent,
 635                          "BUG: direct ref found in indirect tree")) {
 636                         ret = -EINVAL;
 637                         goto out;
 638                 }
 639
 640                 rb_erase(&ref->rbnode, &preftrees->indirect.root);
 641                 preftrees->indirect.count--;
 642
 643                 if (ref->count == 0) {
 644                         free_pref(ref);
 645                         continue;
 646                 }
 647
 648                 if (sc && sc->root_objectid &&
 649                     ref->root_id != sc->root_objectid) {
 650                         free_pref(ref);
 651                         ret = BACKREF_FOUND_SHARED;
 652                         goto out;
 653                 }
 654                 err = resolve_indirect_ref(fs_info, path, time_seq, ref,
 655                                            parents, extent_item_pos,
 656                                            total_refs, ignore_offset);
 657                 /*
 658                  * we can only tolerate ENOENT,otherwise,we should catch error
 659                  * and return directly.
 660                  */
 661                 if (err == -ENOENT) {
 662                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 663                                           NULL);
 664                         continue;
 665                 } else if (err) {
 666                         free_pref(ref);
 667                         ret = err;
 668                         goto out;
 669                 }
 670
 671                 /* we put the first parent into the ref at hand */
 672                 ULIST_ITER_INIT(&uiter);
 673                 node = ulist_next(parents, &uiter);
 674                 ref->parent = node ? node->val : 0;
 675                 ref->inode_list = unode_aux_to_inode_list(node);
 676
 677                 /* Add a prelim_ref(s) for any other parent(s). */
 678                 while ((node = ulist_next(parents, &uiter))) {
 679                         struct prelim_ref *new_ref;
 680
 681                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 682                                                    GFP_NOFS);
 683                         if (!new_ref) {
 684                                 free_pref(ref);
 685                                 ret = -ENOMEM;
 686                                 goto out;
 687                         }
 688                         memcpy(new_ref, ref, sizeof(*ref));
 689                         new_ref->parent = node->val;
 690                         new_ref->inode_list = unode_aux_to_inode_list(node);
 691                         prelim_ref_insert(fs_info, &preftrees->direct,
 692                                           new_ref, NULL);
 693                 }
 694
 695                 /*
 696                  * Now it's a direct ref, put it in the the direct tree. We must
 697                  * do this last because the ref could be merged/freed here.
 698                  */
 699                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 700
 701                 ulist_reinit(parents);
 702                 cond_resched();
 703         }
 704 out:
 705         ulist_free(parents);
 706         return ret;
 707 }
 708
 709 /*
 710  * read tree blocks and add keys where required.
 711  */
 712 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 713                             struct preftrees *preftrees, bool lock)
 714 {
 715         struct prelim_ref *ref;
 716         struct extent_buffer *eb;
 717         struct preftree *tree = &preftrees->indirect_missing_keys;
 718         struct rb_node *node;
 719
 720         while ((node = rb_first(&tree->root))) {
 721                 ref = rb_entry(node, struct prelim_ref, rbnode);
 722                 rb_erase(node, &tree->root);
 723
 724                 BUG_ON(ref->parent);    /* should not be a direct ref */
 725                 BUG_ON(ref->key_for_search.type);
 726                 BUG_ON(!ref->wanted_disk_byte);
 727
 728                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0,
 729                                      ref->level - 1, NULL);
 730                 if (IS_ERR(eb)) {
 731                         free_pref(ref);
 732                         return PTR_ERR(eb);
 733                 } else if (!extent_buffer_uptodate(eb)) {
 734                         free_pref(ref);
 735                         free_extent_buffer(eb);
 736                         return -EIO;
 737                 }
 738                 if (lock)
 739                         btrfs_tree_read_lock(eb);
 740                 if (btrfs_header_level(eb) == 0)
 741                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 742                 else
 743                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 744                 if (lock)
 745                         btrfs_tree_read_unlock(eb);
 746                 free_extent_buffer(eb);
 747                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 748                 cond_resched();
 749         }
 750         return 0;
 751 }
 752
 753 /*
 754  * add all currently queued delayed refs from this head whose seq nr is
 755  * smaller or equal that seq to the list
 756  */
 757 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 758                             struct btrfs_delayed_ref_head *head, u64 seq,
 759                             struct preftrees *preftrees, u64 *total_refs,
 760                             struct share_check *sc)
 761 {
 762         struct btrfs_delayed_ref_node *node;
 763         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 764         struct btrfs_key key;
 765         struct btrfs_key tmp_op_key;
 766         struct rb_node *n;
 767         int count;
 768         int ret = 0;
 769
 770         if (extent_op && extent_op->update_key)
 771                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 772
 773         spin_lock(&head->lock);
 774         for (n = rb_first(&head->ref_tree); n; n = rb_next(n)) {
 775                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 776                                 ref_node);
 777                 if (node->seq > seq)
 778                         continue;
 779
 780                 switch (node->action) {
 781                 case BTRFS_ADD_DELAYED_EXTENT:
 782                 case BTRFS_UPDATE_DELAYED_HEAD:
 783                         WARN_ON(1);
 784                         continue;
 785                 case BTRFS_ADD_DELAYED_REF:
 786                         count = node->ref_mod;
 787                         break;
 788                 case BTRFS_DROP_DELAYED_REF:
 789                         count = node->ref_mod * -1;
 790                         break;
 791                 default:
 792                         BUG_ON(1);
 793                 }
 794                 *total_refs += count;
 795                 switch (node->type) {
 796                 case BTRFS_TREE_BLOCK_REF_KEY: {
 797                         /* NORMAL INDIRECT METADATA backref */
 798                         struct btrfs_delayed_tree_ref *ref;
 799
 800                         ref = btrfs_delayed_node_to_tree_ref(node);
 801                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 802                                                &tmp_op_key, ref->level + 1,
 803                                                node->bytenr, count, sc,
 804                                                GFP_ATOMIC);
 805                         break;
 806                 }
 807                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 808                         /* SHARED DIRECT METADATA backref */
 809                         struct btrfs_delayed_tree_ref *ref;
 810
 811                         ref = btrfs_delayed_node_to_tree_ref(node);
 812
 813                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 814                                              ref->parent, node->bytenr, count,
 815                                              sc, GFP_ATOMIC);
 816                         break;
 817                 }
 818                 case BTRFS_EXTENT_DATA_REF_KEY: {
 819                         /* NORMAL INDIRECT DATA backref */
 820                         struct btrfs_delayed_data_ref *ref;
 821                         ref = btrfs_delayed_node_to_data_ref(node);
 822
 823                         key.objectid = ref->objectid;
 824                         key.type = BTRFS_EXTENT_DATA_KEY;
 825                         key.offset = ref->offset;
 826
 827                         /*
 828                          * Found a inum that doesn't match our known inum, we
 829                          * know it's shared.
 830                          */
 831                         if (sc && sc->inum && ref->objectid != sc->inum) {
 832                                 ret = BACKREF_FOUND_SHARED;
 833                                 goto out;
 834                         }
 835
 836                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 837                                                &key, 0, node->bytenr, count, sc,
 838                                                GFP_ATOMIC);
 839                         break;
 840                 }
 841                 case BTRFS_SHARED_DATA_REF_KEY: {
 842                         /* SHARED DIRECT FULL backref */
 843                         struct btrfs_delayed_data_ref *ref;
 844
 845                         ref = btrfs_delayed_node_to_data_ref(node);
 846
 847                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 848                                              node->bytenr, count, sc,
 849                                              GFP_ATOMIC);
 850                         break;
 851                 }
 852                 default:
 853                         WARN_ON(1);
 854                 }
 855                 /*
 856                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 857                  * refs have been checked.
 858                  */
 859                 if (ret && (ret != BACKREF_FOUND_SHARED))
 860                         break;
 861         }
 862         if (!ret)
 863                 ret = extent_is_shared(sc);
 864 out:
 865         spin_unlock(&head->lock);
 866         return ret;
 867 }
 868
 869 /*
 870  * add all inline backrefs for bytenr to the list
 871  *
 872  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 873  */
 874 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 875                            struct btrfs_path *path, u64 bytenr,
 876                            int *info_level, struct preftrees *preftrees,
 877                            u64 *total_refs, struct share_check *sc)
 878 {
 879         int ret = 0;
 880         int slot;
 881         struct extent_buffer *leaf;
 882         struct btrfs_key key;
 883         struct btrfs_key found_key;
 884         unsigned long ptr;
 885         unsigned long end;
 886         struct btrfs_extent_item *ei;
 887         u64 flags;
 888         u64 item_size;
 889
 890         /*
 891          * enumerate all inline refs
 892          */
 893         leaf = path->nodes[0];
 894         slot = path->slots[0];
 895
 896         item_size = btrfs_item_size_nr(leaf, slot);
 897         BUG_ON(item_size < sizeof(*ei));
 898
 899         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 900         flags = btrfs_extent_flags(leaf, ei);
 901         *total_refs += btrfs_extent_refs(leaf, ei);
 902         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 903
 904         ptr = (unsigned long)(ei + 1);
 905         end = (unsigned long)ei + item_size;
 906
 907         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 908             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 909                 struct btrfs_tree_block_info *info;
 910
 911                 info = (struct btrfs_tree_block_info *)ptr;
 912                 *info_level = btrfs_tree_block_level(leaf, info);
 913                 ptr += sizeof(struct btrfs_tree_block_info);
 914                 BUG_ON(ptr > end);
 915         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 916                 *info_level = found_key.offset;
 917         } else {
 918                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 919         }
 920
 921         while (ptr < end) {
 922                 struct btrfs_extent_inline_ref *iref;
 923                 u64 offset;
 924                 int type;
 925
 926                 iref = (struct btrfs_extent_inline_ref *)ptr;
 927                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 928                                                         BTRFS_REF_TYPE_ANY);
 929                 if (type == BTRFS_REF_TYPE_INVALID)
 930                         return -EUCLEAN;
 931
 932                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 933
 934                 switch (type) {
 935                 case BTRFS_SHARED_BLOCK_REF_KEY:
 936                         ret = add_direct_ref(fs_info, preftrees,
 937                                              *info_level + 1, offset,
 938                                              bytenr, 1, NULL, GFP_NOFS);
 939                         break;
 940                 case BTRFS_SHARED_DATA_REF_KEY: {
 941                         struct btrfs_shared_data_ref *sdref;
 942                         int count;
 943
 944                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 945                         count = btrfs_shared_data_ref_count(leaf, sdref);
 946
 947                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
 948                                              bytenr, count, sc, GFP_NOFS);
 949                         break;
 950                 }
 951                 case BTRFS_TREE_BLOCK_REF_KEY:
 952                         ret = add_indirect_ref(fs_info, preftrees, offset,
 953                                                NULL, *info_level + 1,
 954                                                bytenr, 1, NULL, GFP_NOFS);
 955                         break;
 956                 case BTRFS_EXTENT_DATA_REF_KEY: {
 957                         struct btrfs_extent_data_ref *dref;
 958                         int count;
 959                         u64 root;
 960
 961                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 962                         count = btrfs_extent_data_ref_count(leaf, dref);
 963                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 964                                                                       dref);
 965                         key.type = BTRFS_EXTENT_DATA_KEY;
 966                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 967
 968                         if (sc && sc->inum && key.objectid != sc->inum) {
 969                                 ret = BACKREF_FOUND_SHARED;
 970                                 break;
 971                         }
 972
 973                         root = btrfs_extent_data_ref_root(leaf, dref);
 974
 975                         ret = add_indirect_ref(fs_info, preftrees, root,
 976                                                &key, 0, bytenr, count,
 977                                                sc, GFP_NOFS);
 978                         break;
 979                 }
 980                 default:
 981                         WARN_ON(1);
 982                 }
 983                 if (ret)
 984                         return ret;
 985                 ptr += btrfs_extent_inline_ref_size(type);
 986         }
 987
 988         return 0;
 989 }
 990
 991 /*
 992  * add all non-inline backrefs for bytenr to the list
 993  *
 994  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 995  */
 996 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
 997                           struct btrfs_path *path, u64 bytenr,
 998                           int info_level, struct preftrees *preftrees,
 999                           struct share_check *sc)
1000 {
1001         struct btrfs_root *extent_root = fs_info->extent_root;
1002         int ret;
1003         int slot;
1004         struct extent_buffer *leaf;
1005         struct btrfs_key key;
1006
1007         while (1) {
1008                 ret = btrfs_next_item(extent_root, path);
1009                 if (ret < 0)
1010                         break;
1011                 if (ret) {
1012                         ret = 0;
1013                         break;
1014                 }
1015
1016                 slot = path->slots[0];
1017                 leaf = path->nodes[0];
1018                 btrfs_item_key_to_cpu(leaf, &key, slot);
1019
1020                 if (key.objectid != bytenr)
1021                         break;
1022                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1023                         continue;
1024                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1025                         break;
1026
1027                 switch (key.type) {
1028                 case BTRFS_SHARED_BLOCK_REF_KEY:
1029                         /* SHARED DIRECT METADATA backref */
1030                         ret = add_direct_ref(fs_info, preftrees,
1031                                              info_level + 1, key.offset,
1032                                              bytenr, 1, NULL, GFP_NOFS);
1033                         break;
1034                 case BTRFS_SHARED_DATA_REF_KEY: {
1035                         /* SHARED DIRECT FULL backref */
1036                         struct btrfs_shared_data_ref *sdref;
1037                         int count;
1038
1039                         sdref = btrfs_item_ptr(leaf, slot,
1040                                               struct btrfs_shared_data_ref);
1041                         count = btrfs_shared_data_ref_count(leaf, sdref);
1042                         ret = add_direct_ref(fs_info, preftrees, 0,
1043                                              key.offset, bytenr, count,
1044                                              sc, GFP_NOFS);
1045                         break;
1046                 }
1047                 case BTRFS_TREE_BLOCK_REF_KEY:
1048                         /* NORMAL INDIRECT METADATA backref */
1049                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1050                                                NULL, info_level + 1, bytenr,
1051                                                1, NULL, GFP_NOFS);
1052                         break;
1053                 case BTRFS_EXTENT_DATA_REF_KEY: {
1054                         /* NORMAL INDIRECT DATA backref */
1055                         struct btrfs_extent_data_ref *dref;
1056                         int count;
1057                         u64 root;
1058
1059                         dref = btrfs_item_ptr(leaf, slot,
1060                                               struct btrfs_extent_data_ref);
1061                         count = btrfs_extent_data_ref_count(leaf, dref);
1062                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1063                                                                       dref);
1064                         key.type = BTRFS_EXTENT_DATA_KEY;
1065                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1066
1067                         if (sc && sc->inum && key.objectid != sc->inum) {
1068                                 ret = BACKREF_FOUND_SHARED;
1069                                 break;
1070                         }
1071
1072                         root = btrfs_extent_data_ref_root(leaf, dref);
1073                         ret = add_indirect_ref(fs_info, preftrees, root,
1074                                                &key, 0, bytenr, count,
1075                                                sc, GFP_NOFS);
1076                         break;
1077                 }
1078                 default:
1079                         WARN_ON(1);
1080                 }
1081                 if (ret)
1082                         return ret;
1083
1084         }
1085
1086         return ret;
1087 }
1088
1089 /*
1090  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1091  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1092  * indirect refs to their parent bytenr.
1093  * When roots are found, they're added to the roots list
1094  *
1095  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1096  * much like trans == NULL case, the difference only lies in it will not
1097  * commit root.
1098  * The special case is for qgroup to search roots in commit_transaction().
1099  *
1100  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1101  * shared extent is detected.
1102  *
1103  * Otherwise this returns 0 for success and <0 for an error.
1104  *
1105  * If ignore_offset is set to false, only extent refs whose offsets match
1106  * extent_item_pos are returned.  If true, every extent ref is returned
1107  * and extent_item_pos is ignored.
1108  *
1109  * FIXME some caching might speed things up
1110  */
1111 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1112                              struct btrfs_fs_info *fs_info, u64 bytenr,
1113                              u64 time_seq, struct ulist *refs,
1114                              struct ulist *roots, const u64 *extent_item_pos,
1115                              struct share_check *sc, bool ignore_offset)
1116 {
1117         struct btrfs_key key;
1118         struct btrfs_path *path;
1119         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1120         struct btrfs_delayed_ref_head *head;
1121         int info_level = 0;
1122         int ret;
1123         struct prelim_ref *ref;
1124         struct rb_node *node;
1125         struct extent_inode_elem *eie = NULL;
1126         /* total of both direct AND indirect refs! */
1127         u64 total_refs = 0;
1128         struct preftrees preftrees = {
1129                 .direct = PREFTREE_INIT,
1130                 .indirect = PREFTREE_INIT,
1131                 .indirect_missing_keys = PREFTREE_INIT
1132         };
1133
1134         key.objectid = bytenr;
1135         key.offset = (u64)-1;
1136         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1137                 key.type = BTRFS_METADATA_ITEM_KEY;
1138         else
1139                 key.type = BTRFS_EXTENT_ITEM_KEY;
1140
1141         path = btrfs_alloc_path();
1142         if (!path)
1143                 return -ENOMEM;
1144         if (!trans) {
1145                 path->search_commit_root = 1;
1146                 path->skip_locking = 1;
1147         }
1148
1149         if (time_seq == SEQ_LAST)
1150                 path->skip_locking = 1;
1151
1152         /*
1153          * grab both a lock on the path and a lock on the delayed ref head.
1154          * We need both to get a consistent picture of how the refs look
1155          * at a specified point in time
1156          */
1157 again:
1158         head = NULL;
1159
1160         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1161         if (ret < 0)
1162                 goto out;
1163         BUG_ON(ret == 0);
1164
1165 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1166         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1167             time_seq != SEQ_LAST) {
1168 #else
1169         if (trans && time_seq != SEQ_LAST) {
1170 #endif
1171                 /*
1172                  * look if there are updates for this ref queued and lock the
1173                  * head
1174                  */
1175                 delayed_refs = &trans->transaction->delayed_refs;
1176                 spin_lock(&delayed_refs->lock);
1177                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1178                 if (head) {
1179                         if (!mutex_trylock(&head->mutex)) {
1180                                 refcount_inc(&head->refs);
1181                                 spin_unlock(&delayed_refs->lock);
1182
1183                                 btrfs_release_path(path);
1184
1185                                 /*
1186                                  * Mutex was contended, block until it's
1187                                  * released and try again
1188                                  */
1189                                 mutex_lock(&head->mutex);
1190                                 mutex_unlock(&head->mutex);
1191                                 btrfs_put_delayed_ref_head(head);
1192                                 goto again;
1193                         }
1194                         spin_unlock(&delayed_refs->lock);
1195                         ret = add_delayed_refs(fs_info, head, time_seq,
1196                                                &preftrees, &total_refs, sc);
1197                         mutex_unlock(&head->mutex);
1198                         if (ret)
1199                                 goto out;
1200                 } else {
1201                         spin_unlock(&delayed_refs->lock);
1202                 }
1203         }
1204
1205         if (path->slots[0]) {
1206                 struct extent_buffer *leaf;
1207                 int slot;
1208
1209                 path->slots[0]--;
1210                 leaf = path->nodes[0];
1211                 slot = path->slots[0];
1212                 btrfs_item_key_to_cpu(leaf, &key, slot);
1213                 if (key.objectid == bytenr &&
1214                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1215                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1216                         ret = add_inline_refs(fs_info, path, bytenr,
1217                                               &info_level, &preftrees,
1218                                               &total_refs, sc);
1219                         if (ret)
1220                                 goto out;
1221                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1222                                              &preftrees, sc);
1223                         if (ret)
1224                                 goto out;
1225                 }
1226         }
1227
1228         btrfs_release_path(path);
1229
1230         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1231         if (ret)
1232                 goto out;
1233
1234         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root));
1235
1236         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1237                                     extent_item_pos, total_refs, sc, ignore_offset);
1238         if (ret)
1239                 goto out;
1240
1241         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root));
1242
1243         /*
1244          * This walks the tree of merged and resolved refs. Tree blocks are
1245          * read in as needed. Unique entries are added to the ulist, and
1246          * the list of found roots is updated.
1247          *
1248          * We release the entire tree in one go before returning.
1249          */
1250         node = rb_first(&preftrees.direct.root);
1251         while (node) {
1252                 ref = rb_entry(node, struct prelim_ref, rbnode);
1253                 node = rb_next(&ref->rbnode);
1254                 /*
1255                  * ref->count < 0 can happen here if there are delayed
1256                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1257                  * prelim_ref_insert() relies on this when merging
1258                  * identical refs to keep the overall count correct.
1259                  * prelim_ref_insert() will merge only those refs
1260                  * which compare identically.  Any refs having
1261                  * e.g. different offsets would not be merged,
1262                  * and would retain their original ref->count < 0.
1263                  */
1264                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1265                         if (sc && sc->root_objectid &&
1266                             ref->root_id != sc->root_objectid) {
1267                                 ret = BACKREF_FOUND_SHARED;
1268                                 goto out;
1269                         }
1270
1271                         /* no parent == root of tree */
1272                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1273                         if (ret < 0)
1274                                 goto out;
1275                 }
1276                 if (ref->count && ref->parent) {
1277                         if (extent_item_pos && !ref->inode_list &&
1278                             ref->level == 0) {
1279                                 struct extent_buffer *eb;
1280
1281                                 eb = read_tree_block(fs_info, ref->parent, 0,
1282                                                      ref->level, NULL);
1283                                 if (IS_ERR(eb)) {
1284                                         ret = PTR_ERR(eb);
1285                                         goto out;
1286                                 } else if (!extent_buffer_uptodate(eb)) {
1287                                         free_extent_buffer(eb);
1288                                         ret = -EIO;
1289                                         goto out;
1290                                 }
1291                                 if (!path->skip_locking) {
1292                                         btrfs_tree_read_lock(eb);
1293                                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1294                                 }
1295                                 ret = find_extent_in_eb(eb, bytenr,
1296                                                         *extent_item_pos, &eie, ignore_offset);
1297                                 if (!path->skip_locking)
1298                                         btrfs_tree_read_unlock_blocking(eb);
1299                                 free_extent_buffer(eb);
1300                                 if (ret < 0)
1301                                         goto out;
1302                                 ref->inode_list = eie;
1303                         }
1304                         ret = ulist_add_merge_ptr(refs, ref->parent,
1305                                                   ref->inode_list,
1306                                                   (void **)&eie, GFP_NOFS);
1307                         if (ret < 0)
1308                                 goto out;
1309                         if (!ret && extent_item_pos) {
1310                                 /*
1311                                  * we've recorded that parent, so we must extend
1312                                  * its inode list here
1313                                  */
1314                                 BUG_ON(!eie);
1315                                 while (eie->next)
1316                                         eie = eie->next;
1317                                 eie->next = ref->inode_list;
1318                         }
1319                         eie = NULL;
1320                 }
1321                 cond_resched();
1322         }
1323
1324 out:
1325         btrfs_free_path(path);
1326
1327         prelim_release(&preftrees.direct);
1328         prelim_release(&preftrees.indirect);
1329         prelim_release(&preftrees.indirect_missing_keys);
1330
1331         if (ret < 0)
1332                 free_inode_elem_list(eie);
1333         return ret;
1334 }
1335
1336 static void free_leaf_list(struct ulist *blocks)
1337 {
1338         struct ulist_node *node = NULL;
1339         struct extent_inode_elem *eie;
1340         struct ulist_iterator uiter;
1341
1342         ULIST_ITER_INIT(&uiter);
1343         while ((node = ulist_next(blocks, &uiter))) {
1344                 if (!node->aux)
1345                         continue;
1346                 eie = unode_aux_to_inode_list(node);
1347                 free_inode_elem_list(eie);
1348                 node->aux = 0;
1349         }
1350
1351         ulist_free(blocks);
1352 }
1353
1354 /*
1355  * Finds all leafs with a reference to the specified combination of bytenr and
1356  * offset. key_list_head will point to a list of corresponding keys (caller must
1357  * free each list element). The leafs will be stored in the leafs ulist, which
1358  * must be freed with ulist_free.
1359  *
1360  * returns 0 on success, <0 on error
1361  */
1362 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1363                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1364                                 u64 time_seq, struct ulist **leafs,
1365                                 const u64 *extent_item_pos, bool ignore_offset)
1366 {
1367         int ret;
1368
1369         *leafs = ulist_alloc(GFP_NOFS);
1370         if (!*leafs)
1371                 return -ENOMEM;
1372
1373         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1374                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1375         if (ret < 0 && ret != -ENOENT) {
1376                 free_leaf_list(*leafs);
1377                 return ret;
1378         }
1379
1380         return 0;
1381 }
1382
1383 /*
1384  * walk all backrefs for a given extent to find all roots that reference this
1385  * extent. Walking a backref means finding all extents that reference this
1386  * extent and in turn walk the backrefs of those, too. Naturally this is a
1387  * recursive process, but here it is implemented in an iterative fashion: We
1388  * find all referencing extents for the extent in question and put them on a
1389  * list. In turn, we find all referencing extents for those, further appending
1390  * to the list. The way we iterate the list allows adding more elements after
1391  * the current while iterating. The process stops when we reach the end of the
1392  * list. Found roots are added to the roots list.
1393  *
1394  * returns 0 on success, < 0 on error.
1395  */
1396 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1397                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1398                                      u64 time_seq, struct ulist **roots,
1399                                      bool ignore_offset)
1400 {
1401         struct ulist *tmp;
1402         struct ulist_node *node = NULL;
1403         struct ulist_iterator uiter;
1404         int ret;
1405
1406         tmp = ulist_alloc(GFP_NOFS);
1407         if (!tmp)
1408                 return -ENOMEM;
1409         *roots = ulist_alloc(GFP_NOFS);
1410         if (!*roots) {
1411                 ulist_free(tmp);
1412                 return -ENOMEM;
1413         }
1414
1415         ULIST_ITER_INIT(&uiter);
1416         while (1) {
1417                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1418                                         tmp, *roots, NULL, NULL, ignore_offset);
1419                 if (ret < 0 && ret != -ENOENT) {
1420                         ulist_free(tmp);
1421                         ulist_free(*roots);
1422                         *roots = NULL;
1423                         return ret;
1424                 }
1425                 node = ulist_next(tmp, &uiter);
1426                 if (!node)
1427                         break;
1428                 bytenr = node->val;
1429                 cond_resched();
1430         }
1431
1432         ulist_free(tmp);
1433         return 0;
1434 }
1435
1436 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1437                          struct btrfs_fs_info *fs_info, u64 bytenr,
1438                          u64 time_seq, struct ulist **roots,
1439                          bool ignore_offset)
1440 {
1441         int ret;
1442
1443         if (!trans)
1444                 down_read(&fs_info->commit_root_sem);
1445         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1446                                         time_seq, roots, ignore_offset);
1447         if (!trans)
1448                 up_read(&fs_info->commit_root_sem);
1449         return ret;
1450 }
1451
1452 /**
1453  * btrfs_check_shared - tell us whether an extent is shared
1454  *
1455  * btrfs_check_shared uses the backref walking code but will short
1456  * circuit as soon as it finds a root or inode that doesn't match the
1457  * one passed in. This provides a significant performance benefit for
1458  * callers (such as fiemap) which want to know whether the extent is
1459  * shared but do not need a ref count.
1460  *
1461  * This attempts to attach to the running transaction in order to account for
1462  * delayed refs, but continues on even when no running transaction exists.
1463  *
1464  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1465  */
1466 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr)
1467 {
1468         struct btrfs_fs_info *fs_info = root->fs_info;
1469         struct btrfs_trans_handle *trans;
1470         struct ulist *tmp = NULL;
1471         struct ulist *roots = NULL;
1472         struct ulist_iterator uiter;
1473         struct ulist_node *node;
1474         struct seq_list elem = SEQ_LIST_INIT(elem);
1475         int ret = 0;
1476         struct share_check shared = {
1477                 .root_objectid = root->objectid,
1478                 .inum = inum,
1479                 .share_count = 0,
1480         };
1481
1482         tmp = ulist_alloc(GFP_NOFS);
1483         roots = ulist_alloc(GFP_NOFS);
1484         if (!tmp || !roots) {
1485                 ret = -ENOMEM;
1486                 goto out;
1487         }
1488
1489         trans = btrfs_join_transaction_nostart(root);
1490         if (IS_ERR(trans)) {
1491                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1492                         ret = PTR_ERR(trans);
1493                         goto out;
1494                 }
1495                 trans = NULL;
1496                 down_read(&fs_info->commit_root_sem);
1497         } else {
1498                 btrfs_get_tree_mod_seq(fs_info, &elem);
1499         }
1500
1501         ULIST_ITER_INIT(&uiter);
1502         while (1) {
1503                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1504                                         roots, NULL, &shared, false);
1505                 if (ret == BACKREF_FOUND_SHARED) {
1506                         /* this is the only condition under which we return 1 */
1507                         ret = 1;
1508                         break;
1509                 }
1510                 if (ret < 0 && ret != -ENOENT)
1511                         break;
1512                 ret = 0;
1513                 node = ulist_next(tmp, &uiter);
1514                 if (!node)
1515                         break;
1516                 bytenr = node->val;
1517                 shared.share_count = 0;
1518                 cond_resched();
1519         }
1520
1521         if (trans) {
1522                 btrfs_put_tree_mod_seq(fs_info, &elem);
1523                 btrfs_end_transaction(trans);
1524         } else {
1525                 up_read(&fs_info->commit_root_sem);
1526         }
1527 out:
1528         ulist_free(tmp);
1529         ulist_free(roots);
1530         return ret;
1531 }
1532
1533 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1534                           u64 start_off, struct btrfs_path *path,
1535                           struct btrfs_inode_extref **ret_extref,
1536                           u64 *found_off)
1537 {
1538         int ret, slot;
1539         struct btrfs_key key;
1540         struct btrfs_key found_key;
1541         struct btrfs_inode_extref *extref;
1542         const struct extent_buffer *leaf;
1543         unsigned long ptr;
1544
1545         key.objectid = inode_objectid;
1546         key.type = BTRFS_INODE_EXTREF_KEY;
1547         key.offset = start_off;
1548
1549         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1550         if (ret < 0)
1551                 return ret;
1552
1553         while (1) {
1554                 leaf = path->nodes[0];
1555                 slot = path->slots[0];
1556                 if (slot >= btrfs_header_nritems(leaf)) {
1557                         /*
1558                          * If the item at offset is not found,
1559                          * btrfs_search_slot will point us to the slot
1560                          * where it should be inserted. In our case
1561                          * that will be the slot directly before the
1562                          * next INODE_REF_KEY_V2 item. In the case
1563                          * that we're pointing to the last slot in a
1564                          * leaf, we must move one leaf over.
1565                          */
1566                         ret = btrfs_next_leaf(root, path);
1567                         if (ret) {
1568                                 if (ret >= 1)
1569                                         ret = -ENOENT;
1570                                 break;
1571                         }
1572                         continue;
1573                 }
1574
1575                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1576
1577                 /*
1578                  * Check that we're still looking at an extended ref key for
1579                  * this particular objectid. If we have different
1580                  * objectid or type then there are no more to be found
1581                  * in the tree and we can exit.
1582                  */
1583                 ret = -ENOENT;
1584                 if (found_key.objectid != inode_objectid)
1585                         break;
1586                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1587                         break;
1588
1589                 ret = 0;
1590                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1591                 extref = (struct btrfs_inode_extref *)ptr;
1592                 *ret_extref = extref;
1593                 if (found_off)
1594                         *found_off = found_key.offset;
1595                 break;
1596         }
1597
1598         return ret;
1599 }
1600
1601 /*
1602  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1603  * Elements of the path are separated by '/' and the path is guaranteed to be
1604  * 0-terminated. the path is only given within the current file system.
1605  * Therefore, it never starts with a '/'. the caller is responsible to provide
1606  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1607  * the start point of the resulting string is returned. this pointer is within
1608  * dest, normally.
1609  * in case the path buffer would overflow, the pointer is decremented further
1610  * as if output was written to the buffer, though no more output is actually
1611  * generated. that way, the caller can determine how much space would be
1612  * required for the path to fit into the buffer. in that case, the returned
1613  * value will be smaller than dest. callers must check this!
1614  */
1615 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1616                         u32 name_len, unsigned long name_off,
1617                         struct extent_buffer *eb_in, u64 parent,
1618                         char *dest, u32 size)
1619 {
1620         int slot;
1621         u64 next_inum;
1622         int ret;
1623         s64 bytes_left = ((s64)size) - 1;
1624         struct extent_buffer *eb = eb_in;
1625         struct btrfs_key found_key;
1626         int leave_spinning = path->leave_spinning;
1627         struct btrfs_inode_ref *iref;
1628
1629         if (bytes_left >= 0)
1630                 dest[bytes_left] = '\0';
1631
1632         path->leave_spinning = 1;
1633         while (1) {
1634                 bytes_left -= name_len;
1635                 if (bytes_left >= 0)
1636                         read_extent_buffer(eb, dest + bytes_left,
1637                                            name_off, name_len);
1638                 if (eb != eb_in) {
1639                         if (!path->skip_locking)
1640                                 btrfs_tree_read_unlock_blocking(eb);
1641                         free_extent_buffer(eb);
1642                 }
1643                 ret = btrfs_find_item(fs_root, path, parent, 0,
1644                                 BTRFS_INODE_REF_KEY, &found_key);
1645                 if (ret > 0)
1646                         ret = -ENOENT;
1647                 if (ret)
1648                         break;
1649
1650                 next_inum = found_key.offset;
1651
1652                 /* regular exit ahead */
1653                 if (parent == next_inum)
1654                         break;
1655
1656                 slot = path->slots[0];
1657                 eb = path->nodes[0];
1658                 /* make sure we can use eb after releasing the path */
1659                 if (eb != eb_in) {
1660                         if (!path->skip_locking)
1661                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1662                         path->nodes[0] = NULL;
1663                         path->locks[0] = 0;
1664                 }
1665                 btrfs_release_path(path);
1666                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1667
1668                 name_len = btrfs_inode_ref_name_len(eb, iref);
1669                 name_off = (unsigned long)(iref + 1);
1670
1671                 parent = next_inum;
1672                 --bytes_left;
1673                 if (bytes_left >= 0)
1674                         dest[bytes_left] = '/';
1675         }
1676
1677         btrfs_release_path(path);
1678         path->leave_spinning = leave_spinning;
1679
1680         if (ret)
1681                 return ERR_PTR(ret);
1682
1683         return dest + bytes_left;
1684 }
1685
1686 /*
1687  * this makes the path point to (logical EXTENT_ITEM *)
1688  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1689  * tree blocks and <0 on error.
1690  */
1691 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1692                         struct btrfs_path *path, struct btrfs_key *found_key,
1693                         u64 *flags_ret)
1694 {
1695         int ret;
1696         u64 flags;
1697         u64 size = 0;
1698         u32 item_size;
1699         const struct extent_buffer *eb;
1700         struct btrfs_extent_item *ei;
1701         struct btrfs_key key;
1702
1703         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1704                 key.type = BTRFS_METADATA_ITEM_KEY;
1705         else
1706                 key.type = BTRFS_EXTENT_ITEM_KEY;
1707         key.objectid = logical;
1708         key.offset = (u64)-1;
1709
1710         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1711         if (ret < 0)
1712                 return ret;
1713
1714         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1715         if (ret) {
1716                 if (ret > 0)
1717                         ret = -ENOENT;
1718                 return ret;
1719         }
1720         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1721         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1722                 size = fs_info->nodesize;
1723         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1724                 size = found_key->offset;
1725
1726         if (found_key->objectid > logical ||
1727             found_key->objectid + size <= logical) {
1728                 btrfs_debug(fs_info,
1729                         "logical %llu is not within any extent", logical);
1730                 return -ENOENT;
1731         }
1732
1733         eb = path->nodes[0];
1734         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1735         BUG_ON(item_size < sizeof(*ei));
1736
1737         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1738         flags = btrfs_extent_flags(eb, ei);
1739
1740         btrfs_debug(fs_info,
1741                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1742                  logical, logical - found_key->objectid, found_key->objectid,
1743                  found_key->offset, flags, item_size);
1744
1745         WARN_ON(!flags_ret);
1746         if (flags_ret) {
1747                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1748                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1749                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1750                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1751                 else
1752                         BUG_ON(1);
1753                 return 0;
1754         }
1755
1756         return -EIO;
1757 }
1758
1759 /*
1760  * helper function to iterate extent inline refs. ptr must point to a 0 value
1761  * for the first call and may be modified. it is used to track state.
1762  * if more refs exist, 0 is returned and the next call to
1763  * get_extent_inline_ref must pass the modified ptr parameter to get the
1764  * next ref. after the last ref was processed, 1 is returned.
1765  * returns <0 on error
1766  */
1767 static int get_extent_inline_ref(unsigned long *ptr,
1768                                  const struct extent_buffer *eb,
1769                                  const struct btrfs_key *key,
1770                                  const struct btrfs_extent_item *ei,
1771                                  u32 item_size,
1772                                  struct btrfs_extent_inline_ref **out_eiref,
1773                                  int *out_type)
1774 {
1775         unsigned long end;
1776         u64 flags;
1777         struct btrfs_tree_block_info *info;
1778
1779         if (!*ptr) {
1780                 /* first call */
1781                 flags = btrfs_extent_flags(eb, ei);
1782                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1783                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1784                                 /* a skinny metadata extent */
1785                                 *out_eiref =
1786                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1787                         } else {
1788                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1789                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1790                                 *out_eiref =
1791                                    (struct btrfs_extent_inline_ref *)(info + 1);
1792                         }
1793                 } else {
1794                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1795                 }
1796                 *ptr = (unsigned long)*out_eiref;
1797                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1798                         return -ENOENT;
1799         }
1800
1801         end = (unsigned long)ei + item_size;
1802         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1803         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1804                                                      BTRFS_REF_TYPE_ANY);
1805         if (*out_type == BTRFS_REF_TYPE_INVALID)
1806                 return -EUCLEAN;
1807
1808         *ptr += btrfs_extent_inline_ref_size(*out_type);
1809         WARN_ON(*ptr > end);
1810         if (*ptr == end)
1811                 return 1; /* last */
1812
1813         return 0;
1814 }
1815
1816 /*
1817  * reads the tree block backref for an extent. tree level and root are returned
1818  * through out_level and out_root. ptr must point to a 0 value for the first
1819  * call and may be modified (see get_extent_inline_ref comment).
1820  * returns 0 if data was provided, 1 if there was no more data to provide or
1821  * <0 on error.
1822  */
1823 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1824                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1825                             u32 item_size, u64 *out_root, u8 *out_level)
1826 {
1827         int ret;
1828         int type;
1829         struct btrfs_extent_inline_ref *eiref;
1830
1831         if (*ptr == (unsigned long)-1)
1832                 return 1;
1833
1834         while (1) {
1835                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1836                                               &eiref, &type);
1837                 if (ret < 0)
1838                         return ret;
1839
1840                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1841                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1842                         break;
1843
1844                 if (ret == 1)
1845                         return 1;
1846         }
1847
1848         /* we can treat both ref types equally here */
1849         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1850
1851         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1852                 struct btrfs_tree_block_info *info;
1853
1854                 info = (struct btrfs_tree_block_info *)(ei + 1);
1855                 *out_level = btrfs_tree_block_level(eb, info);
1856         } else {
1857                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1858                 *out_level = (u8)key->offset;
1859         }
1860
1861         if (ret == 1)
1862                 *ptr = (unsigned long)-1;
1863
1864         return 0;
1865 }
1866
1867 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1868                              struct extent_inode_elem *inode_list,
1869                              u64 root, u64 extent_item_objectid,
1870                              iterate_extent_inodes_t *iterate, void *ctx)
1871 {
1872         struct extent_inode_elem *eie;
1873         int ret = 0;
1874
1875         for (eie = inode_list; eie; eie = eie->next) {
1876                 btrfs_debug(fs_info,
1877                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1878                             extent_item_objectid, eie->inum,
1879                             eie->offset, root);
1880                 ret = iterate(eie->inum, eie->offset, root, ctx);
1881                 if (ret) {
1882                         btrfs_debug(fs_info,
1883                                     "stopping iteration for %llu due to ret=%d",
1884                                     extent_item_objectid, ret);
1885                         break;
1886                 }
1887         }
1888
1889         return ret;
1890 }
1891
1892 /*
1893  * calls iterate() for every inode that references the extent identified by
1894  * the given parameters.
1895  * when the iterator function returns a non-zero value, iteration stops.
1896  */
1897 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1898                                 u64 extent_item_objectid, u64 extent_item_pos,
1899                                 int search_commit_root,
1900                                 iterate_extent_inodes_t *iterate, void *ctx,
1901                                 bool ignore_offset)
1902 {
1903         int ret;
1904         struct btrfs_trans_handle *trans = NULL;
1905         struct ulist *refs = NULL;
1906         struct ulist *roots = NULL;
1907         struct ulist_node *ref_node = NULL;
1908         struct ulist_node *root_node = NULL;
1909         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1910         struct ulist_iterator ref_uiter;
1911         struct ulist_iterator root_uiter;
1912
1913         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1914                         extent_item_objectid);
1915
1916         if (!search_commit_root) {
1917                 trans = btrfs_attach_transaction(fs_info->extent_root);
1918                 if (IS_ERR(trans)) {
1919                         if (PTR_ERR(trans) != -ENOENT &&
1920                             PTR_ERR(trans) != -EROFS)
1921                                 return PTR_ERR(trans);
1922                         trans = NULL;
1923                 }
1924         }
1925
1926         if (trans)
1927                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1928         else
1929                 down_read(&fs_info->commit_root_sem);
1930
1931         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1932                                    tree_mod_seq_elem.seq, &refs,
1933                                    &extent_item_pos, ignore_offset);
1934         if (ret)
1935                 goto out;
1936
1937         ULIST_ITER_INIT(&ref_uiter);
1938         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1939                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1940                                                 tree_mod_seq_elem.seq, &roots,
1941                                                 ignore_offset);
1942                 if (ret)
1943                         break;
1944                 ULIST_ITER_INIT(&root_uiter);
1945                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1946                         btrfs_debug(fs_info,
1947                                     "root %llu references leaf %llu, data list %#llx",
1948                                     root_node->val, ref_node->val,
1949                                     ref_node->aux);
1950                         ret = iterate_leaf_refs(fs_info,
1951                                                 (struct extent_inode_elem *)
1952                                                 (uintptr_t)ref_node->aux,
1953                                                 root_node->val,
1954                                                 extent_item_objectid,
1955                                                 iterate, ctx);
1956                 }
1957                 ulist_free(roots);
1958         }
1959
1960         free_leaf_list(refs);
1961 out:
1962         if (trans) {
1963                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1964                 btrfs_end_transaction(trans);
1965         } else {
1966                 up_read(&fs_info->commit_root_sem);
1967         }
1968
1969         return ret;
1970 }
1971
1972 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
1973                                 struct btrfs_path *path,
1974                                 iterate_extent_inodes_t *iterate, void *ctx,
1975                                 bool ignore_offset)
1976 {
1977         int ret;
1978         u64 extent_item_pos;
1979         u64 flags = 0;
1980         struct btrfs_key found_key;
1981         int search_commit_root = path->search_commit_root;
1982
1983         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
1984         btrfs_release_path(path);
1985         if (ret < 0)
1986                 return ret;
1987         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1988                 return -EINVAL;
1989
1990         extent_item_pos = logical - found_key.objectid;
1991         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1992                                         extent_item_pos, search_commit_root,
1993                                         iterate, ctx, ignore_offset);
1994
1995         return ret;
1996 }
1997
1998 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
1999                               struct extent_buffer *eb, void *ctx);
2000
2001 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2002                               struct btrfs_path *path,
2003                               iterate_irefs_t *iterate, void *ctx)
2004 {
2005         int ret = 0;
2006         int slot;
2007         u32 cur;
2008         u32 len;
2009         u32 name_len;
2010         u64 parent = 0;
2011         int found = 0;
2012         struct extent_buffer *eb;
2013         struct btrfs_item *item;
2014         struct btrfs_inode_ref *iref;
2015         struct btrfs_key found_key;
2016
2017         while (!ret) {
2018                 ret = btrfs_find_item(fs_root, path, inum,
2019                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2020                                 &found_key);
2021
2022                 if (ret < 0)
2023                         break;
2024                 if (ret) {
2025                         ret = found ? 0 : -ENOENT;
2026                         break;
2027                 }
2028                 ++found;
2029
2030                 parent = found_key.offset;
2031                 slot = path->slots[0];
2032                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2033                 if (!eb) {
2034                         ret = -ENOMEM;
2035                         break;
2036                 }
2037                 extent_buffer_get(eb);
2038                 btrfs_tree_read_lock(eb);
2039                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2040                 btrfs_release_path(path);
2041
2042                 item = btrfs_item_nr(slot);
2043                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2044
2045                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2046                         name_len = btrfs_inode_ref_name_len(eb, iref);
2047                         /* path must be released before calling iterate()! */
2048                         btrfs_debug(fs_root->fs_info,
2049                                 "following ref at offset %u for inode %llu in tree %llu",
2050                                 cur, found_key.objectid, fs_root->objectid);
2051                         ret = iterate(parent, name_len,
2052                                       (unsigned long)(iref + 1), eb, ctx);
2053                         if (ret)
2054                                 break;
2055                         len = sizeof(*iref) + name_len;
2056                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2057                 }
2058                 btrfs_tree_read_unlock_blocking(eb);
2059                 free_extent_buffer(eb);
2060         }
2061
2062         btrfs_release_path(path);
2063
2064         return ret;
2065 }
2066
2067 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2068                                  struct btrfs_path *path,
2069                                  iterate_irefs_t *iterate, void *ctx)
2070 {
2071         int ret;
2072         int slot;
2073         u64 offset = 0;
2074         u64 parent;
2075         int found = 0;
2076         struct extent_buffer *eb;
2077         struct btrfs_inode_extref *extref;
2078         u32 item_size;
2079         u32 cur_offset;
2080         unsigned long ptr;
2081
2082         while (1) {
2083                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2084                                             &offset);
2085                 if (ret < 0)
2086                         break;
2087                 if (ret) {
2088                         ret = found ? 0 : -ENOENT;
2089                         break;
2090                 }
2091                 ++found;
2092
2093                 slot = path->slots[0];
2094                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2095                 if (!eb) {
2096                         ret = -ENOMEM;
2097                         break;
2098                 }
2099                 extent_buffer_get(eb);
2100
2101                 btrfs_tree_read_lock(eb);
2102                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2103                 btrfs_release_path(path);
2104
2105                 item_size = btrfs_item_size_nr(eb, slot);
2106                 ptr = btrfs_item_ptr_offset(eb, slot);
2107                 cur_offset = 0;
2108
2109                 while (cur_offset < item_size) {
2110                         u32 name_len;
2111
2112                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2113                         parent = btrfs_inode_extref_parent(eb, extref);
2114                         name_len = btrfs_inode_extref_name_len(eb, extref);
2115                         ret = iterate(parent, name_len,
2116                                       (unsigned long)&extref->name, eb, ctx);
2117                         if (ret)
2118                                 break;
2119
2120                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2121                         cur_offset += sizeof(*extref);
2122                 }
2123                 btrfs_tree_read_unlock_blocking(eb);
2124                 free_extent_buffer(eb);
2125
2126                 offset++;
2127         }
2128
2129         btrfs_release_path(path);
2130
2131         return ret;
2132 }
2133
2134 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2135                          struct btrfs_path *path, iterate_irefs_t *iterate,
2136                          void *ctx)
2137 {
2138         int ret;
2139         int found_refs = 0;
2140
2141         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2142         if (!ret)
2143                 ++found_refs;
2144         else if (ret != -ENOENT)
2145                 return ret;
2146
2147         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2148         if (ret == -ENOENT && found_refs)
2149                 return 0;
2150
2151         return ret;
2152 }
2153
2154 /*
2155  * returns 0 if the path could be dumped (probably truncated)
2156  * returns <0 in case of an error
2157  */
2158 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2159                          struct extent_buffer *eb, void *ctx)
2160 {
2161         struct inode_fs_paths *ipath = ctx;
2162         char *fspath;
2163         char *fspath_min;
2164         int i = ipath->fspath->elem_cnt;
2165         const int s_ptr = sizeof(char *);
2166         u32 bytes_left;
2167
2168         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2169                                         ipath->fspath->bytes_left - s_ptr : 0;
2170
2171         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2172         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2173                                    name_off, eb, inum, fspath_min, bytes_left);
2174         if (IS_ERR(fspath))
2175                 return PTR_ERR(fspath);
2176
2177         if (fspath > fspath_min) {
2178                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2179                 ++ipath->fspath->elem_cnt;
2180                 ipath->fspath->bytes_left = fspath - fspath_min;
2181         } else {
2182                 ++ipath->fspath->elem_missed;
2183                 ipath->fspath->bytes_missing += fspath_min - fspath;
2184                 ipath->fspath->bytes_left = 0;
2185         }
2186
2187         return 0;
2188 }
2189
2190 /*
2191  * this dumps all file system paths to the inode into the ipath struct, provided
2192  * is has been created large enough. each path is zero-terminated and accessed
2193  * from ipath->fspath->val[i].
2194  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2195  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2196  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2197  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2198  * have been needed to return all paths.
2199  */
2200 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2201 {
2202         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2203                              inode_to_path, ipath);
2204 }
2205
2206 struct btrfs_data_container *init_data_container(u32 total_bytes)
2207 {
2208         struct btrfs_data_container *data;
2209         size_t alloc_bytes;
2210
2211         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2212         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2213         if (!data)
2214                 return ERR_PTR(-ENOMEM);
2215
2216         if (total_bytes >= sizeof(*data)) {
2217                 data->bytes_left = total_bytes - sizeof(*data);
2218                 data->bytes_missing = 0;
2219         } else {
2220                 data->bytes_missing = sizeof(*data) - total_bytes;
2221                 data->bytes_left = 0;
2222         }
2223
2224         data->elem_cnt = 0;
2225         data->elem_missed = 0;
2226
2227         return data;
2228 }
2229
2230 /*
2231  * allocates space to return multiple file system paths for an inode.
2232  * total_bytes to allocate are passed, note that space usable for actual path
2233  * information will be total_bytes - sizeof(struct inode_fs_paths).
2234  * the returned pointer must be freed with free_ipath() in the end.
2235  */
2236 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2237                                         struct btrfs_path *path)
2238 {
2239         struct inode_fs_paths *ifp;
2240         struct btrfs_data_container *fspath;
2241
2242         fspath = init_data_container(total_bytes);
2243         if (IS_ERR(fspath))
2244                 return ERR_CAST(fspath);
2245
2246         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2247         if (!ifp) {
2248                 kvfree(fspath);
2249                 return ERR_PTR(-ENOMEM);
2250         }
2251
2252         ifp->btrfs_path = path;
2253         ifp->fspath = fspath;
2254         ifp->fs_root = fs_root;
2255
2256         return ifp;
2257 }
2258
2259 void free_ipath(struct inode_fs_paths *ipath)
2260 {
2261         if (!ipath)
2262                 return;
2263         kvfree(ipath->fspath);
2264         kfree(ipath);
2265 }