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         bool have_delayed_delete_refs;
 140 };
 141
 142 static inline int extent_is_shared(struct share_check *sc)
 143 {
 144         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 145 }
 146
 147 static struct kmem_cache *btrfs_prelim_ref_cache;
 148
 149 int __init btrfs_prelim_ref_init(void)
 150 {
 151         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 152                                         sizeof(struct prelim_ref),
 153                                         0,
 154                                         SLAB_MEM_SPREAD,
 155                                         NULL);
 156         if (!btrfs_prelim_ref_cache)
 157                 return -ENOMEM;
 158         return 0;
 159 }
 160
 161 void __cold btrfs_prelim_ref_exit(void)
 162 {
 163         kmem_cache_destroy(btrfs_prelim_ref_cache);
 164 }
 165
 166 static void free_pref(struct prelim_ref *ref)
 167 {
 168         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 169 }
 170
 171 /*
 172  * Return 0 when both refs are for the same block (and can be merged).
 173  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 174  * indicates a 'higher' block.
 175  */
 176 static int prelim_ref_compare(struct prelim_ref *ref1,
 177                               struct prelim_ref *ref2)
 178 {
 179         if (ref1->level < ref2->level)
 180                 return -1;
 181         if (ref1->level > ref2->level)
 182                 return 1;
 183         if (ref1->root_id < ref2->root_id)
 184                 return -1;
 185         if (ref1->root_id > ref2->root_id)
 186                 return 1;
 187         if (ref1->key_for_search.type < ref2->key_for_search.type)
 188                 return -1;
 189         if (ref1->key_for_search.type > ref2->key_for_search.type)
 190                 return 1;
 191         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 192                 return -1;
 193         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 194                 return 1;
 195         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 196                 return -1;
 197         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 198                 return 1;
 199         if (ref1->parent < ref2->parent)
 200                 return -1;
 201         if (ref1->parent > ref2->parent)
 202                 return 1;
 203
 204         return 0;
 205 }
 206
 207 static void update_share_count(struct share_check *sc, int oldcount,
 208                                int newcount)
 209 {
 210         if ((!sc) || (oldcount == 0 && newcount < 1))
 211                 return;
 212
 213         if (oldcount > 0 && newcount < 1)
 214                 sc->share_count--;
 215         else if (oldcount < 1 && newcount > 0)
 216                 sc->share_count++;
 217 }
 218
 219 /*
 220  * Add @newref to the @root rbtree, merging identical refs.
 221  *
 222  * Callers should assume that newref has been freed after calling.
 223  */
 224 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 225                               struct preftree *preftree,
 226                               struct prelim_ref *newref,
 227                               struct share_check *sc)
 228 {
 229         struct rb_root *root;
 230         struct rb_node **p;
 231         struct rb_node *parent = NULL;
 232         struct prelim_ref *ref;
 233         int result;
 234
 235         root = &preftree->root;
 236         p = &root->rb_node;
 237
 238         while (*p) {
 239                 parent = *p;
 240                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 241                 result = prelim_ref_compare(ref, newref);
 242                 if (result < 0) {
 243                         p = &(*p)->rb_left;
 244                 } else if (result > 0) {
 245                         p = &(*p)->rb_right;
 246                 } else {
 247                         /* Identical refs, merge them and free @newref */
 248                         struct extent_inode_elem *eie = ref->inode_list;
 249
 250                         while (eie && eie->next)
 251                                 eie = eie->next;
 252
 253                         if (!eie)
 254                                 ref->inode_list = newref->inode_list;
 255                         else
 256                                 eie->next = newref->inode_list;
 257                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 258                                                      preftree->count);
 259                         /*
 260                          * A delayed ref can have newref->count < 0.
 261                          * The ref->count is updated to follow any
 262                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 263                          */
 264                         update_share_count(sc, ref->count,
 265                                            ref->count + newref->count);
 266                         ref->count += newref->count;
 267                         free_pref(newref);
 268                         return;
 269                 }
 270         }
 271
 272         update_share_count(sc, 0, newref->count);
 273         preftree->count++;
 274         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 275         rb_link_node(&newref->rbnode, parent, p);
 276         rb_insert_color(&newref->rbnode, root);
 277 }
 278
 279 /*
 280  * Release the entire tree.  We don't care about internal consistency so
 281  * just free everything and then reset the tree root.
 282  */
 283 static void prelim_release(struct preftree *preftree)
 284 {
 285         struct prelim_ref *ref, *next_ref;
 286
 287         rbtree_postorder_for_each_entry_safe(ref, next_ref, &preftree->root,
 288                                              rbnode)
 289                 free_pref(ref);
 290
 291         preftree->root = RB_ROOT;
 292         preftree->count = 0;
 293 }
 294
 295 /*
 296  * the rules for all callers of this function are:
 297  * - obtaining the parent is the goal
 298  * - if you add a key, you must know that it is a correct key
 299  * - if you cannot add the parent or a correct key, then we will look into the
 300  *   block later to set a correct key
 301  *
 302  * delayed refs
 303  * ============
 304  *        backref type | shared | indirect | shared | indirect
 305  * information         |   tree |     tree |   data |     data
 306  * --------------------+--------+----------+--------+----------
 307  *      parent logical |    y   |     -    |    -   |     -
 308  *      key to resolve |    -   |     y    |    y   |     y
 309  *  tree block logical |    -   |     -    |    -   |     -
 310  *  root for resolving |    y   |     y    |    y   |     y
 311  *
 312  * - column 1:       we've the parent -> done
 313  * - column 2, 3, 4: we use the key to find the parent
 314  *
 315  * on disk refs (inline or keyed)
 316  * ==============================
 317  *        backref type | shared | indirect | shared | indirect
 318  * information         |   tree |     tree |   data |     data
 319  * --------------------+--------+----------+--------+----------
 320  *      parent logical |    y   |     -    |    y   |     -
 321  *      key to resolve |    -   |     -    |    -   |     y
 322  *  tree block logical |    y   |     y    |    y   |     y
 323  *  root for resolving |    -   |     y    |    y   |     y
 324  *
 325  * - column 1, 3: we've the parent -> done
 326  * - column 2:    we take the first key from the block to find the parent
 327  *                (see add_missing_keys)
 328  * - column 4:    we use the key to find the parent
 329  *
 330  * additional information that's available but not required to find the parent
 331  * block might help in merging entries to gain some speed.
 332  */
 333 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 334                           struct preftree *preftree, u64 root_id,
 335                           const struct btrfs_key *key, int level, u64 parent,
 336                           u64 wanted_disk_byte, int count,
 337                           struct share_check *sc, gfp_t gfp_mask)
 338 {
 339         struct prelim_ref *ref;
 340
 341         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 342                 return 0;
 343
 344         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 345         if (!ref)
 346                 return -ENOMEM;
 347
 348         ref->root_id = root_id;
 349         if (key) {
 350                 ref->key_for_search = *key;
 351                 /*
 352                  * We can often find data backrefs with an offset that is too
 353                  * large (>= LLONG_MAX, maximum allowed file offset) due to
 354                  * underflows when subtracting a file's offset with the data
 355                  * offset of its corresponding extent data item. This can
 356                  * happen for example in the clone ioctl.
 357                  * So if we detect such case we set the search key's offset to
 358                  * zero to make sure we will find the matching file extent item
 359                  * at add_all_parents(), otherwise we will miss it because the
 360                  * offset taken form the backref is much larger then the offset
 361                  * of the file extent item. This can make us scan a very large
 362                  * number of file extent items, but at least it will not make
 363                  * us miss any.
 364                  * This is an ugly workaround for a behaviour that should have
 365                  * never existed, but it does and a fix for the clone ioctl
 366                  * would touch a lot of places, cause backwards incompatibility
 367                  * and would not fix the problem for extents cloned with older
 368                  * kernels.
 369                  */
 370                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
 371                     ref->key_for_search.offset >= LLONG_MAX)
 372                         ref->key_for_search.offset = 0;
 373         } else {
 374                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 375         }
 376
 377         ref->inode_list = NULL;
 378         ref->level = level;
 379         ref->count = count;
 380         ref->parent = parent;
 381         ref->wanted_disk_byte = wanted_disk_byte;
 382         prelim_ref_insert(fs_info, preftree, ref, sc);
 383         return extent_is_shared(sc);
 384 }
 385
 386 /* direct refs use root == 0, key == NULL */
 387 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 388                           struct preftrees *preftrees, int level, u64 parent,
 389                           u64 wanted_disk_byte, int count,
 390                           struct share_check *sc, gfp_t gfp_mask)
 391 {
 392         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 393                               parent, wanted_disk_byte, count, sc, gfp_mask);
 394 }
 395
 396 /* indirect refs use parent == 0 */
 397 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 398                             struct preftrees *preftrees, u64 root_id,
 399                             const struct btrfs_key *key, int level,
 400                             u64 wanted_disk_byte, int count,
 401                             struct share_check *sc, gfp_t gfp_mask)
 402 {
 403         struct preftree *tree = &preftrees->indirect;
 404
 405         if (!key)
 406                 tree = &preftrees->indirect_missing_keys;
 407         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 408                               wanted_disk_byte, count, sc, gfp_mask);
 409 }
 410
 411 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 412                            struct ulist *parents, struct prelim_ref *ref,
 413                            int level, u64 time_seq, const u64 *extent_item_pos,
 414                            u64 total_refs, bool ignore_offset)
 415 {
 416         int ret = 0;
 417         int slot;
 418         struct extent_buffer *eb;
 419         struct btrfs_key key;
 420         struct btrfs_key *key_for_search = &ref->key_for_search;
 421         struct btrfs_file_extent_item *fi;
 422         struct extent_inode_elem *eie = NULL, *old = NULL;
 423         u64 disk_byte;
 424         u64 wanted_disk_byte = ref->wanted_disk_byte;
 425         u64 count = 0;
 426
 427         if (level != 0) {
 428                 eb = path->nodes[level];
 429                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 430                 if (ret < 0)
 431                         return ret;
 432                 return 0;
 433         }
 434
 435         /*
 436          * We normally enter this function with the path already pointing to
 437          * the first item to check. But sometimes, we may enter it with
 438          * slot==nritems. In that case, go to the next leaf before we continue.
 439          */
 440         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 441                 if (time_seq == SEQ_LAST)
 442                         ret = btrfs_next_leaf(root, path);
 443                 else
 444                         ret = btrfs_next_old_leaf(root, path, time_seq);
 445         }
 446
 447         while (!ret && count < total_refs) {
 448                 eb = path->nodes[0];
 449                 slot = path->slots[0];
 450
 451                 btrfs_item_key_to_cpu(eb, &key, slot);
 452
 453                 if (key.objectid != key_for_search->objectid ||
 454                     key.type != BTRFS_EXTENT_DATA_KEY)
 455                         break;
 456
 457                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 458                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 459
 460                 if (disk_byte == wanted_disk_byte) {
 461                         eie = NULL;
 462                         old = NULL;
 463                         count++;
 464                         if (extent_item_pos) {
 465                                 ret = check_extent_in_eb(&key, eb, fi,
 466                                                 *extent_item_pos,
 467                                                 &eie, ignore_offset);
 468                                 if (ret < 0)
 469                                         break;
 470                         }
 471                         if (ret > 0)
 472                                 goto next;
 473                         ret = ulist_add_merge_ptr(parents, eb->start,
 474                                                   eie, (void **)&old, GFP_NOFS);
 475                         if (ret < 0)
 476                                 break;
 477                         if (!ret && extent_item_pos) {
 478                                 while (old->next)
 479                                         old = old->next;
 480                                 old->next = eie;
 481                         }
 482                         eie = NULL;
 483                 }
 484 next:
 485                 if (time_seq == SEQ_LAST)
 486                         ret = btrfs_next_item(root, path);
 487                 else
 488                         ret = btrfs_next_old_item(root, path, time_seq);
 489         }
 490
 491         if (ret > 0)
 492                 ret = 0;
 493         else if (ret < 0)
 494                 free_inode_elem_list(eie);
 495         return ret;
 496 }
 497
 498 /*
 499  * resolve an indirect backref in the form (root_id, key, level)
 500  * to a logical address
 501  */
 502 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 503                                 struct btrfs_path *path, u64 time_seq,
 504                                 struct prelim_ref *ref, struct ulist *parents,
 505                                 const u64 *extent_item_pos, u64 total_refs,
 506                                 bool ignore_offset)
 507 {
 508         struct btrfs_root *root;
 509         struct btrfs_key root_key;
 510         struct extent_buffer *eb;
 511         int ret = 0;
 512         int root_level;
 513         int level = ref->level;
 514         int index;
 515
 516         root_key.objectid = ref->root_id;
 517         root_key.type = BTRFS_ROOT_ITEM_KEY;
 518         root_key.offset = (u64)-1;
 519
 520         index = srcu_read_lock(&fs_info->subvol_srcu);
 521
 522         root = btrfs_get_fs_root(fs_info, &root_key, false);
 523         if (IS_ERR(root)) {
 524                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 525                 ret = PTR_ERR(root);
 526                 goto out;
 527         }
 528
 529         if (btrfs_is_testing(fs_info)) {
 530                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 531                 ret = -ENOENT;
 532                 goto out;
 533         }
 534
 535         if (path->search_commit_root)
 536                 root_level = btrfs_header_level(root->commit_root);
 537         else if (time_seq == SEQ_LAST)
 538                 root_level = btrfs_header_level(root->node);
 539         else
 540                 root_level = btrfs_old_root_level(root, time_seq);
 541
 542         if (root_level + 1 == level) {
 543                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 544                 goto out;
 545         }
 546
 547         path->lowest_level = level;
 548         if (time_seq == SEQ_LAST)
 549                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
 550                                         0, 0);
 551         else
 552                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
 553                                             time_seq);
 554
 555         /* root node has been locked, we can release @subvol_srcu safely here */
 556         srcu_read_unlock(&fs_info->subvol_srcu, index);
 557
 558         btrfs_debug(fs_info,
 559                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 560                  ref->root_id, level, ref->count, ret,
 561                  ref->key_for_search.objectid, ref->key_for_search.type,
 562                  ref->key_for_search.offset);
 563         if (ret < 0)
 564                 goto out;
 565
 566         eb = path->nodes[level];
 567         while (!eb) {
 568                 if (WARN_ON(!level)) {
 569                         ret = 1;
 570                         goto out;
 571                 }
 572                 level--;
 573                 eb = path->nodes[level];
 574         }
 575
 576         ret = add_all_parents(root, path, parents, ref, level, time_seq,
 577                               extent_item_pos, total_refs, ignore_offset);
 578 out:
 579         path->lowest_level = 0;
 580         btrfs_release_path(path);
 581         return ret;
 582 }
 583
 584 static struct extent_inode_elem *
 585 unode_aux_to_inode_list(struct ulist_node *node)
 586 {
 587         if (!node)
 588                 return NULL;
 589         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 590 }
 591
 592 /*
 593  * We maintain three seperate rbtrees: one for direct refs, one for
 594  * indirect refs which have a key, and one for indirect refs which do not
 595  * have a key. Each tree does merge on insertion.
 596  *
 597  * Once all of the references are located, we iterate over the tree of
 598  * indirect refs with missing keys. An appropriate key is located and
 599  * the ref is moved onto the tree for indirect refs. After all missing
 600  * keys are thus located, we iterate over the indirect ref tree, resolve
 601  * each reference, and then insert the resolved reference onto the
 602  * direct tree (merging there too).
 603  *
 604  * New backrefs (i.e., for parent nodes) are added to the appropriate
 605  * rbtree as they are encountered. The new backrefs are subsequently
 606  * resolved as above.
 607  */
 608 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 609                                  struct btrfs_path *path, u64 time_seq,
 610                                  struct preftrees *preftrees,
 611                                  const u64 *extent_item_pos, u64 total_refs,
 612                                  struct share_check *sc, bool ignore_offset)
 613 {
 614         int err;
 615         int ret = 0;
 616         struct ulist *parents;
 617         struct ulist_node *node;
 618         struct ulist_iterator uiter;
 619         struct rb_node *rnode;
 620
 621         parents = ulist_alloc(GFP_NOFS);
 622         if (!parents)
 623                 return -ENOMEM;
 624
 625         /*
 626          * We could trade memory usage for performance here by iterating
 627          * the tree, allocating new refs for each insertion, and then
 628          * freeing the entire indirect tree when we're done.  In some test
 629          * cases, the tree can grow quite large (~200k objects).
 630          */
 631         while ((rnode = rb_first(&preftrees->indirect.root))) {
 632                 struct prelim_ref *ref;
 633
 634                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 635                 if (WARN(ref->parent,
 636                          "BUG: direct ref found in indirect tree")) {
 637                         ret = -EINVAL;
 638                         goto out;
 639                 }
 640
 641                 rb_erase(&ref->rbnode, &preftrees->indirect.root);
 642                 preftrees->indirect.count--;
 643
 644                 if (ref->count == 0) {
 645                         free_pref(ref);
 646                         continue;
 647                 }
 648
 649                 if (sc && sc->root_objectid &&
 650                     ref->root_id != sc->root_objectid) {
 651                         free_pref(ref);
 652                         ret = BACKREF_FOUND_SHARED;
 653                         goto out;
 654                 }
 655                 err = resolve_indirect_ref(fs_info, path, time_seq, ref,
 656                                            parents, extent_item_pos,
 657                                            total_refs, ignore_offset);
 658                 /*
 659                  * we can only tolerate ENOENT,otherwise,we should catch error
 660                  * and return directly.
 661                  */
 662                 if (err == -ENOENT) {
 663                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 664                                           NULL);
 665                         continue;
 666                 } else if (err) {
 667                         free_pref(ref);
 668                         ret = err;
 669                         goto out;
 670                 }
 671
 672                 /* we put the first parent into the ref at hand */
 673                 ULIST_ITER_INIT(&uiter);
 674                 node = ulist_next(parents, &uiter);
 675                 ref->parent = node ? node->val : 0;
 676                 ref->inode_list = unode_aux_to_inode_list(node);
 677
 678                 /* Add a prelim_ref(s) for any other parent(s). */
 679                 while ((node = ulist_next(parents, &uiter))) {
 680                         struct prelim_ref *new_ref;
 681
 682                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 683                                                    GFP_NOFS);
 684                         if (!new_ref) {
 685                                 free_pref(ref);
 686                                 ret = -ENOMEM;
 687                                 goto out;
 688                         }
 689                         memcpy(new_ref, ref, sizeof(*ref));
 690                         new_ref->parent = node->val;
 691                         new_ref->inode_list = unode_aux_to_inode_list(node);
 692                         prelim_ref_insert(fs_info, &preftrees->direct,
 693                                           new_ref, NULL);
 694                 }
 695
 696                 /*
 697                  * Now it's a direct ref, put it in the the direct tree. We must
 698                  * do this last because the ref could be merged/freed here.
 699                  */
 700                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 701
 702                 ulist_reinit(parents);
 703                 cond_resched();
 704         }
 705 out:
 706         ulist_free(parents);
 707         return ret;
 708 }
 709
 710 /*
 711  * read tree blocks and add keys where required.
 712  */
 713 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 714                             struct preftrees *preftrees, bool lock)
 715 {
 716         struct prelim_ref *ref;
 717         struct extent_buffer *eb;
 718         struct preftree *tree = &preftrees->indirect_missing_keys;
 719         struct rb_node *node;
 720
 721         while ((node = rb_first(&tree->root))) {
 722                 ref = rb_entry(node, struct prelim_ref, rbnode);
 723                 rb_erase(node, &tree->root);
 724
 725                 BUG_ON(ref->parent);    /* should not be a direct ref */
 726                 BUG_ON(ref->key_for_search.type);
 727                 BUG_ON(!ref->wanted_disk_byte);
 728
 729                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0,
 730                                      ref->level - 1, NULL);
 731                 if (IS_ERR(eb)) {
 732                         free_pref(ref);
 733                         return PTR_ERR(eb);
 734                 } else if (!extent_buffer_uptodate(eb)) {
 735                         free_pref(ref);
 736                         free_extent_buffer(eb);
 737                         return -EIO;
 738                 }
 739                 if (lock)
 740                         btrfs_tree_read_lock(eb);
 741                 if (btrfs_header_level(eb) == 0)
 742                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 743                 else
 744                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 745                 if (lock)
 746                         btrfs_tree_read_unlock(eb);
 747                 free_extent_buffer(eb);
 748                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 749                 cond_resched();
 750         }
 751         return 0;
 752 }
 753
 754 /*
 755  * add all currently queued delayed refs from this head whose seq nr is
 756  * smaller or equal that seq to the list
 757  */
 758 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 759                             struct btrfs_delayed_ref_head *head, u64 seq,
 760                             struct preftrees *preftrees, u64 *total_refs,
 761                             struct share_check *sc)
 762 {
 763         struct btrfs_delayed_ref_node *node;
 764         struct btrfs_key key;
 765         struct rb_node *n;
 766         int count;
 767         int ret = 0;
 768
 769         spin_lock(&head->lock);
 770         for (n = rb_first(&head->ref_tree); n; n = rb_next(n)) {
 771                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 772                                 ref_node);
 773                 if (node->seq > seq)
 774                         continue;
 775
 776                 switch (node->action) {
 777                 case BTRFS_ADD_DELAYED_EXTENT:
 778                 case BTRFS_UPDATE_DELAYED_HEAD:
 779                         WARN_ON(1);
 780                         continue;
 781                 case BTRFS_ADD_DELAYED_REF:
 782                         count = node->ref_mod;
 783                         break;
 784                 case BTRFS_DROP_DELAYED_REF:
 785                         count = node->ref_mod * -1;
 786                         break;
 787                 default:
 788                         BUG_ON(1);
 789                 }
 790                 *total_refs += count;
 791                 switch (node->type) {
 792                 case BTRFS_TREE_BLOCK_REF_KEY: {
 793                         /* NORMAL INDIRECT METADATA backref */
 794                         struct btrfs_delayed_tree_ref *ref;
 795                         struct btrfs_key *key_ptr = NULL;
 796
 797                         if (head->extent_op && head->extent_op->update_key) {
 798                                 btrfs_disk_key_to_cpu(&key, &head->extent_op->key);
 799                                 key_ptr = &key;
 800                         }
 801
 802                         ref = btrfs_delayed_node_to_tree_ref(node);
 803                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 804                                                key_ptr, ref->level + 1,
 805                                                node->bytenr, count, sc,
 806                                                GFP_ATOMIC);
 807                         break;
 808                 }
 809                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 810                         /* SHARED DIRECT METADATA backref */
 811                         struct btrfs_delayed_tree_ref *ref;
 812
 813                         ref = btrfs_delayed_node_to_tree_ref(node);
 814
 815                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 816                                              ref->parent, node->bytenr, count,
 817                                              sc, GFP_ATOMIC);
 818                         break;
 819                 }
 820                 case BTRFS_EXTENT_DATA_REF_KEY: {
 821                         /* NORMAL INDIRECT DATA backref */
 822                         struct btrfs_delayed_data_ref *ref;
 823                         ref = btrfs_delayed_node_to_data_ref(node);
 824
 825                         key.objectid = ref->objectid;
 826                         key.type = BTRFS_EXTENT_DATA_KEY;
 827                         key.offset = ref->offset;
 828
 829                         /*
 830                          * If we have a share check context and a reference for
 831                          * another inode, we can't exit immediately. This is
 832                          * because even if this is a BTRFS_ADD_DELAYED_REF
 833                          * reference we may find next a BTRFS_DROP_DELAYED_REF
 834                          * which cancels out this ADD reference.
 835                          *
 836                          * If this is a DROP reference and there was no previous
 837                          * ADD reference, then we need to signal that when we
 838                          * process references from the extent tree (through
 839                          * add_inline_refs() and add_keyed_refs()), we should
 840                          * not exit early if we find a reference for another
 841                          * inode, because one of the delayed DROP references
 842                          * may cancel that reference in the extent tree.
 843                          */
 844                         if (sc && count < 0)
 845                                 sc->have_delayed_delete_refs = true;
 846
 847                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 848                                                &key, 0, node->bytenr, count, sc,
 849                                                GFP_ATOMIC);
 850                         break;
 851                 }
 852                 case BTRFS_SHARED_DATA_REF_KEY: {
 853                         /* SHARED DIRECT FULL backref */
 854                         struct btrfs_delayed_data_ref *ref;
 855
 856                         ref = btrfs_delayed_node_to_data_ref(node);
 857
 858                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 859                                              node->bytenr, count, sc,
 860                                              GFP_ATOMIC);
 861                         break;
 862                 }
 863                 default:
 864                         WARN_ON(1);
 865                 }
 866                 /*
 867                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 868                  * refs have been checked.
 869                  */
 870                 if (ret && (ret != BACKREF_FOUND_SHARED))
 871                         break;
 872         }
 873         if (!ret)
 874                 ret = extent_is_shared(sc);
 875
 876         spin_unlock(&head->lock);
 877         return ret;
 878 }
 879
 880 /*
 881  * add all inline backrefs for bytenr to the list
 882  *
 883  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 884  */
 885 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 886                            struct btrfs_path *path, u64 bytenr,
 887                            int *info_level, struct preftrees *preftrees,
 888                            u64 *total_refs, struct share_check *sc)
 889 {
 890         int ret = 0;
 891         int slot;
 892         struct extent_buffer *leaf;
 893         struct btrfs_key key;
 894         struct btrfs_key found_key;
 895         unsigned long ptr;
 896         unsigned long end;
 897         struct btrfs_extent_item *ei;
 898         u64 flags;
 899         u64 item_size;
 900
 901         /*
 902          * enumerate all inline refs
 903          */
 904         leaf = path->nodes[0];
 905         slot = path->slots[0];
 906
 907         item_size = btrfs_item_size_nr(leaf, slot);
 908         BUG_ON(item_size < sizeof(*ei));
 909
 910         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 911         flags = btrfs_extent_flags(leaf, ei);
 912         *total_refs += btrfs_extent_refs(leaf, ei);
 913         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 914
 915         ptr = (unsigned long)(ei + 1);
 916         end = (unsigned long)ei + item_size;
 917
 918         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 919             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 920                 struct btrfs_tree_block_info *info;
 921
 922                 info = (struct btrfs_tree_block_info *)ptr;
 923                 *info_level = btrfs_tree_block_level(leaf, info);
 924                 ptr += sizeof(struct btrfs_tree_block_info);
 925                 BUG_ON(ptr > end);
 926         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 927                 *info_level = found_key.offset;
 928         } else {
 929                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 930         }
 931
 932         while (ptr < end) {
 933                 struct btrfs_extent_inline_ref *iref;
 934                 u64 offset;
 935                 int type;
 936
 937                 iref = (struct btrfs_extent_inline_ref *)ptr;
 938                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 939                                                         BTRFS_REF_TYPE_ANY);
 940                 if (type == BTRFS_REF_TYPE_INVALID)
 941                         return -EUCLEAN;
 942
 943                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 944
 945                 switch (type) {
 946                 case BTRFS_SHARED_BLOCK_REF_KEY:
 947                         ret = add_direct_ref(fs_info, preftrees,
 948                                              *info_level + 1, offset,
 949                                              bytenr, 1, NULL, GFP_NOFS);
 950                         break;
 951                 case BTRFS_SHARED_DATA_REF_KEY: {
 952                         struct btrfs_shared_data_ref *sdref;
 953                         int count;
 954
 955                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 956                         count = btrfs_shared_data_ref_count(leaf, sdref);
 957
 958                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
 959                                              bytenr, count, sc, GFP_NOFS);
 960                         break;
 961                 }
 962                 case BTRFS_TREE_BLOCK_REF_KEY:
 963                         ret = add_indirect_ref(fs_info, preftrees, offset,
 964                                                NULL, *info_level + 1,
 965                                                bytenr, 1, NULL, GFP_NOFS);
 966                         break;
 967                 case BTRFS_EXTENT_DATA_REF_KEY: {
 968                         struct btrfs_extent_data_ref *dref;
 969                         int count;
 970                         u64 root;
 971
 972                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 973                         count = btrfs_extent_data_ref_count(leaf, dref);
 974                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 975                                                                       dref);
 976                         key.type = BTRFS_EXTENT_DATA_KEY;
 977                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 978
 979                         if (sc && sc->inum && key.objectid != sc->inum &&
 980                             !sc->have_delayed_delete_refs) {
 981                                 ret = BACKREF_FOUND_SHARED;
 982                                 break;
 983                         }
 984
 985                         root = btrfs_extent_data_ref_root(leaf, dref);
 986
 987                         ret = add_indirect_ref(fs_info, preftrees, root,
 988                                                &key, 0, bytenr, count,
 989                                                sc, GFP_NOFS);
 990
 991                         break;
 992                 }
 993                 default:
 994                         WARN_ON(1);
 995                 }
 996                 if (ret)
 997                         return ret;
 998                 ptr += btrfs_extent_inline_ref_size(type);
 999         }
1000
1001         return 0;
1002 }
1003
1004 /*
1005  * add all non-inline backrefs for bytenr to the list
1006  *
1007  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1008  */
1009 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1010                           struct btrfs_path *path, u64 bytenr,
1011                           int info_level, struct preftrees *preftrees,
1012                           struct share_check *sc)
1013 {
1014         struct btrfs_root *extent_root = fs_info->extent_root;
1015         int ret;
1016         int slot;
1017         struct extent_buffer *leaf;
1018         struct btrfs_key key;
1019
1020         while (1) {
1021                 ret = btrfs_next_item(extent_root, path);
1022                 if (ret < 0)
1023                         break;
1024                 if (ret) {
1025                         ret = 0;
1026                         break;
1027                 }
1028
1029                 slot = path->slots[0];
1030                 leaf = path->nodes[0];
1031                 btrfs_item_key_to_cpu(leaf, &key, slot);
1032
1033                 if (key.objectid != bytenr)
1034                         break;
1035                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1036                         continue;
1037                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1038                         break;
1039
1040                 switch (key.type) {
1041                 case BTRFS_SHARED_BLOCK_REF_KEY:
1042                         /* SHARED DIRECT METADATA backref */
1043                         ret = add_direct_ref(fs_info, preftrees,
1044                                              info_level + 1, key.offset,
1045                                              bytenr, 1, NULL, GFP_NOFS);
1046                         break;
1047                 case BTRFS_SHARED_DATA_REF_KEY: {
1048                         /* SHARED DIRECT FULL backref */
1049                         struct btrfs_shared_data_ref *sdref;
1050                         int count;
1051
1052                         sdref = btrfs_item_ptr(leaf, slot,
1053                                               struct btrfs_shared_data_ref);
1054                         count = btrfs_shared_data_ref_count(leaf, sdref);
1055                         ret = add_direct_ref(fs_info, preftrees, 0,
1056                                              key.offset, bytenr, count,
1057                                              sc, GFP_NOFS);
1058                         break;
1059                 }
1060                 case BTRFS_TREE_BLOCK_REF_KEY:
1061                         /* NORMAL INDIRECT METADATA backref */
1062                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1063                                                NULL, info_level + 1, bytenr,
1064                                                1, NULL, GFP_NOFS);
1065                         break;
1066                 case BTRFS_EXTENT_DATA_REF_KEY: {
1067                         /* NORMAL INDIRECT DATA backref */
1068                         struct btrfs_extent_data_ref *dref;
1069                         int count;
1070                         u64 root;
1071
1072                         dref = btrfs_item_ptr(leaf, slot,
1073                                               struct btrfs_extent_data_ref);
1074                         count = btrfs_extent_data_ref_count(leaf, dref);
1075                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1076                                                                       dref);
1077                         key.type = BTRFS_EXTENT_DATA_KEY;
1078                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1079
1080                         if (sc && sc->inum && key.objectid != sc->inum &&
1081                             !sc->have_delayed_delete_refs) {
1082                                 ret = BACKREF_FOUND_SHARED;
1083                                 break;
1084                         }
1085
1086                         root = btrfs_extent_data_ref_root(leaf, dref);
1087                         ret = add_indirect_ref(fs_info, preftrees, root,
1088                                                &key, 0, bytenr, count,
1089                                                sc, GFP_NOFS);
1090                         break;
1091                 }
1092                 default:
1093                         WARN_ON(1);
1094                 }
1095                 if (ret)
1096                         return ret;
1097
1098         }
1099
1100         return ret;
1101 }
1102
1103 /*
1104  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1105  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1106  * indirect refs to their parent bytenr.
1107  * When roots are found, they're added to the roots list
1108  *
1109  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1110  * much like trans == NULL case, the difference only lies in it will not
1111  * commit root.
1112  * The special case is for qgroup to search roots in commit_transaction().
1113  *
1114  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1115  * shared extent is detected.
1116  *
1117  * Otherwise this returns 0 for success and <0 for an error.
1118  *
1119  * If ignore_offset is set to false, only extent refs whose offsets match
1120  * extent_item_pos are returned.  If true, every extent ref is returned
1121  * and extent_item_pos is ignored.
1122  *
1123  * FIXME some caching might speed things up
1124  */
1125 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1126                              struct btrfs_fs_info *fs_info, u64 bytenr,
1127                              u64 time_seq, struct ulist *refs,
1128                              struct ulist *roots, const u64 *extent_item_pos,
1129                              struct share_check *sc, bool ignore_offset)
1130 {
1131         struct btrfs_key key;
1132         struct btrfs_path *path;
1133         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1134         struct btrfs_delayed_ref_head *head;
1135         int info_level = 0;
1136         int ret;
1137         struct prelim_ref *ref;
1138         struct rb_node *node;
1139         struct extent_inode_elem *eie = NULL;
1140         /* total of both direct AND indirect refs! */
1141         u64 total_refs = 0;
1142         struct preftrees preftrees = {
1143                 .direct = PREFTREE_INIT,
1144                 .indirect = PREFTREE_INIT,
1145                 .indirect_missing_keys = PREFTREE_INIT
1146         };
1147
1148         key.objectid = bytenr;
1149         key.offset = (u64)-1;
1150         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1151                 key.type = BTRFS_METADATA_ITEM_KEY;
1152         else
1153                 key.type = BTRFS_EXTENT_ITEM_KEY;
1154
1155         path = btrfs_alloc_path();
1156         if (!path)
1157                 return -ENOMEM;
1158         if (!trans) {
1159                 path->search_commit_root = 1;
1160                 path->skip_locking = 1;
1161         }
1162
1163         if (time_seq == SEQ_LAST)
1164                 path->skip_locking = 1;
1165
1166         /*
1167          * grab both a lock on the path and a lock on the delayed ref head.
1168          * We need both to get a consistent picture of how the refs look
1169          * at a specified point in time
1170          */
1171 again:
1172         head = NULL;
1173
1174         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1175         if (ret < 0)
1176                 goto out;
1177         if (ret == 0) {
1178                 /* This shouldn't happen, indicates a bug or fs corruption. */
1179                 ASSERT(ret != 0);
1180                 ret = -EUCLEAN;
1181                 goto out;
1182         }
1183
1184 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1185         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1186             time_seq != SEQ_LAST) {
1187 #else
1188         if (trans && time_seq != SEQ_LAST) {
1189 #endif
1190                 /*
1191                  * look if there are updates for this ref queued and lock the
1192                  * head
1193                  */
1194                 delayed_refs = &trans->transaction->delayed_refs;
1195                 spin_lock(&delayed_refs->lock);
1196                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1197                 if (head) {
1198                         if (!mutex_trylock(&head->mutex)) {
1199                                 refcount_inc(&head->refs);
1200                                 spin_unlock(&delayed_refs->lock);
1201
1202                                 btrfs_release_path(path);
1203
1204                                 /*
1205                                  * Mutex was contended, block until it's
1206                                  * released and try again
1207                                  */
1208                                 mutex_lock(&head->mutex);
1209                                 mutex_unlock(&head->mutex);
1210                                 btrfs_put_delayed_ref_head(head);
1211                                 goto again;
1212                         }
1213                         spin_unlock(&delayed_refs->lock);
1214                         ret = add_delayed_refs(fs_info, head, time_seq,
1215                                                &preftrees, &total_refs, sc);
1216                         mutex_unlock(&head->mutex);
1217                         if (ret)
1218                                 goto out;
1219                 } else {
1220                         spin_unlock(&delayed_refs->lock);
1221                 }
1222         }
1223
1224         if (path->slots[0]) {
1225                 struct extent_buffer *leaf;
1226                 int slot;
1227
1228                 path->slots[0]--;
1229                 leaf = path->nodes[0];
1230                 slot = path->slots[0];
1231                 btrfs_item_key_to_cpu(leaf, &key, slot);
1232                 if (key.objectid == bytenr &&
1233                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1234                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1235                         ret = add_inline_refs(fs_info, path, bytenr,
1236                                               &info_level, &preftrees,
1237                                               &total_refs, sc);
1238                         if (ret)
1239                                 goto out;
1240                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1241                                              &preftrees, sc);
1242                         if (ret)
1243                                 goto out;
1244                 }
1245         }
1246
1247         btrfs_release_path(path);
1248
1249         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1250         if (ret)
1251                 goto out;
1252
1253         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root));
1254
1255         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1256                                     extent_item_pos, total_refs, sc, ignore_offset);
1257         if (ret)
1258                 goto out;
1259
1260         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root));
1261
1262         /*
1263          * This walks the tree of merged and resolved refs. Tree blocks are
1264          * read in as needed. Unique entries are added to the ulist, and
1265          * the list of found roots is updated.
1266          *
1267          * We release the entire tree in one go before returning.
1268          */
1269         node = rb_first(&preftrees.direct.root);
1270         while (node) {
1271                 ref = rb_entry(node, struct prelim_ref, rbnode);
1272                 node = rb_next(&ref->rbnode);
1273                 /*
1274                  * ref->count < 0 can happen here if there are delayed
1275                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1276                  * prelim_ref_insert() relies on this when merging
1277                  * identical refs to keep the overall count correct.
1278                  * prelim_ref_insert() will merge only those refs
1279                  * which compare identically.  Any refs having
1280                  * e.g. different offsets would not be merged,
1281                  * and would retain their original ref->count < 0.
1282                  */
1283                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1284                         if (sc && sc->root_objectid &&
1285                             ref->root_id != sc->root_objectid) {
1286                                 ret = BACKREF_FOUND_SHARED;
1287                                 goto out;
1288                         }
1289
1290                         /* no parent == root of tree */
1291                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1292                         if (ret < 0)
1293                                 goto out;
1294                 }
1295                 if (ref->count && ref->parent) {
1296                         if (extent_item_pos && !ref->inode_list &&
1297                             ref->level == 0) {
1298                                 struct extent_buffer *eb;
1299
1300                                 eb = read_tree_block(fs_info, ref->parent, 0,
1301                                                      ref->level, NULL);
1302                                 if (IS_ERR(eb)) {
1303                                         ret = PTR_ERR(eb);
1304                                         goto out;
1305                                 } else if (!extent_buffer_uptodate(eb)) {
1306                                         free_extent_buffer(eb);
1307                                         ret = -EIO;
1308                                         goto out;
1309                                 }
1310                                 if (!path->skip_locking) {
1311                                         btrfs_tree_read_lock(eb);
1312                                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1313                                 }
1314                                 ret = find_extent_in_eb(eb, bytenr,
1315                                                         *extent_item_pos, &eie, ignore_offset);
1316                                 if (!path->skip_locking)
1317                                         btrfs_tree_read_unlock_blocking(eb);
1318                                 free_extent_buffer(eb);
1319                                 if (ret < 0)
1320                                         goto out;
1321                                 ref->inode_list = eie;
1322                         }
1323                         ret = ulist_add_merge_ptr(refs, ref->parent,
1324                                                   ref->inode_list,
1325                                                   (void **)&eie, GFP_NOFS);
1326                         if (ret < 0)
1327                                 goto out;
1328                         if (!ret && extent_item_pos) {
1329                                 /*
1330                                  * We've recorded that parent, so we must extend
1331                                  * its inode list here.
1332                                  *
1333                                  * However if there was corruption we may not
1334                                  * have found an eie, return an error in this
1335                                  * case.
1336                                  */
1337                                 ASSERT(eie);
1338                                 if (!eie) {
1339                                         ret = -EUCLEAN;
1340                                         goto out;
1341                                 }
1342                                 while (eie->next)
1343                                         eie = eie->next;
1344                                 eie->next = ref->inode_list;
1345                         }
1346                         eie = NULL;
1347                 }
1348                 cond_resched();
1349         }
1350
1351 out:
1352         btrfs_free_path(path);
1353
1354         prelim_release(&preftrees.direct);
1355         prelim_release(&preftrees.indirect);
1356         prelim_release(&preftrees.indirect_missing_keys);
1357
1358         if (ret < 0)
1359                 free_inode_elem_list(eie);
1360         return ret;
1361 }
1362
1363 static void free_leaf_list(struct ulist *blocks)
1364 {
1365         struct ulist_node *node = NULL;
1366         struct extent_inode_elem *eie;
1367         struct ulist_iterator uiter;
1368
1369         ULIST_ITER_INIT(&uiter);
1370         while ((node = ulist_next(blocks, &uiter))) {
1371                 if (!node->aux)
1372                         continue;
1373                 eie = unode_aux_to_inode_list(node);
1374                 free_inode_elem_list(eie);
1375                 node->aux = 0;
1376         }
1377
1378         ulist_free(blocks);
1379 }
1380
1381 /*
1382  * Finds all leafs with a reference to the specified combination of bytenr and
1383  * offset. key_list_head will point to a list of corresponding keys (caller must
1384  * free each list element). The leafs will be stored in the leafs ulist, which
1385  * must be freed with ulist_free.
1386  *
1387  * returns 0 on success, <0 on error
1388  */
1389 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1390                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1391                                 u64 time_seq, struct ulist **leafs,
1392                                 const u64 *extent_item_pos, bool ignore_offset)
1393 {
1394         int ret;
1395
1396         *leafs = ulist_alloc(GFP_NOFS);
1397         if (!*leafs)
1398                 return -ENOMEM;
1399
1400         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1401                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1402         if (ret < 0 && ret != -ENOENT) {
1403                 free_leaf_list(*leafs);
1404                 return ret;
1405         }
1406
1407         return 0;
1408 }
1409
1410 /*
1411  * walk all backrefs for a given extent to find all roots that reference this
1412  * extent. Walking a backref means finding all extents that reference this
1413  * extent and in turn walk the backrefs of those, too. Naturally this is a
1414  * recursive process, but here it is implemented in an iterative fashion: We
1415  * find all referencing extents for the extent in question and put them on a
1416  * list. In turn, we find all referencing extents for those, further appending
1417  * to the list. The way we iterate the list allows adding more elements after
1418  * the current while iterating. The process stops when we reach the end of the
1419  * list. Found roots are added to the roots list.
1420  *
1421  * returns 0 on success, < 0 on error.
1422  */
1423 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1424                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1425                                      u64 time_seq, struct ulist **roots,
1426                                      bool ignore_offset)
1427 {
1428         struct ulist *tmp;
1429         struct ulist_node *node = NULL;
1430         struct ulist_iterator uiter;
1431         int ret;
1432
1433         tmp = ulist_alloc(GFP_NOFS);
1434         if (!tmp)
1435                 return -ENOMEM;
1436         *roots = ulist_alloc(GFP_NOFS);
1437         if (!*roots) {
1438                 ulist_free(tmp);
1439                 return -ENOMEM;
1440         }
1441
1442         ULIST_ITER_INIT(&uiter);
1443         while (1) {
1444                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1445                                         tmp, *roots, NULL, NULL, ignore_offset);
1446                 if (ret < 0 && ret != -ENOENT) {
1447                         ulist_free(tmp);
1448                         ulist_free(*roots);
1449                         *roots = NULL;
1450                         return ret;
1451                 }
1452                 node = ulist_next(tmp, &uiter);
1453                 if (!node)
1454                         break;
1455                 bytenr = node->val;
1456                 cond_resched();
1457         }
1458
1459         ulist_free(tmp);
1460         return 0;
1461 }
1462
1463 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1464                          struct btrfs_fs_info *fs_info, u64 bytenr,
1465                          u64 time_seq, struct ulist **roots,
1466                          bool ignore_offset)
1467 {
1468         int ret;
1469
1470         if (!trans)
1471                 down_read(&fs_info->commit_root_sem);
1472         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1473                                         time_seq, roots, ignore_offset);
1474         if (!trans)
1475                 up_read(&fs_info->commit_root_sem);
1476         return ret;
1477 }
1478
1479 /**
1480  * btrfs_check_shared - tell us whether an extent is shared
1481  *
1482  * btrfs_check_shared uses the backref walking code but will short
1483  * circuit as soon as it finds a root or inode that doesn't match the
1484  * one passed in. This provides a significant performance benefit for
1485  * callers (such as fiemap) which want to know whether the extent is
1486  * shared but do not need a ref count.
1487  *
1488  * This attempts to attach to the running transaction in order to account for
1489  * delayed refs, but continues on even when no running transaction exists.
1490  *
1491  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1492  */
1493 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr)
1494 {
1495         struct btrfs_fs_info *fs_info = root->fs_info;
1496         struct btrfs_trans_handle *trans;
1497         struct ulist *tmp = NULL;
1498         struct ulist *roots = NULL;
1499         struct ulist_iterator uiter;
1500         struct ulist_node *node;
1501         struct seq_list elem = SEQ_LIST_INIT(elem);
1502         int ret = 0;
1503         struct share_check shared = {
1504                 .root_objectid = root->objectid,
1505                 .inum = inum,
1506                 .share_count = 0,
1507                 .have_delayed_delete_refs = false,
1508         };
1509
1510         tmp = ulist_alloc(GFP_NOFS);
1511         roots = ulist_alloc(GFP_NOFS);
1512         if (!tmp || !roots) {
1513                 ret = -ENOMEM;
1514                 goto out;
1515         }
1516
1517         trans = btrfs_join_transaction_nostart(root);
1518         if (IS_ERR(trans)) {
1519                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1520                         ret = PTR_ERR(trans);
1521                         goto out;
1522                 }
1523                 trans = NULL;
1524                 down_read(&fs_info->commit_root_sem);
1525         } else {
1526                 btrfs_get_tree_mod_seq(fs_info, &elem);
1527         }
1528
1529         ULIST_ITER_INIT(&uiter);
1530         while (1) {
1531                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1532                                         roots, NULL, &shared, false);
1533                 if (ret == BACKREF_FOUND_SHARED) {
1534                         /* this is the only condition under which we return 1 */
1535                         ret = 1;
1536                         break;
1537                 }
1538                 if (ret < 0 && ret != -ENOENT)
1539                         break;
1540                 ret = 0;
1541                 node = ulist_next(tmp, &uiter);
1542                 if (!node)
1543                         break;
1544                 bytenr = node->val;
1545                 shared.share_count = 0;
1546                 shared.have_delayed_delete_refs = false;
1547                 cond_resched();
1548         }
1549
1550         if (trans) {
1551                 btrfs_put_tree_mod_seq(fs_info, &elem);
1552                 btrfs_end_transaction(trans);
1553         } else {
1554                 up_read(&fs_info->commit_root_sem);
1555         }
1556 out:
1557         ulist_free(tmp);
1558         ulist_free(roots);
1559         return ret;
1560 }
1561
1562 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1563                           u64 start_off, struct btrfs_path *path,
1564                           struct btrfs_inode_extref **ret_extref,
1565                           u64 *found_off)
1566 {
1567         int ret, slot;
1568         struct btrfs_key key;
1569         struct btrfs_key found_key;
1570         struct btrfs_inode_extref *extref;
1571         const struct extent_buffer *leaf;
1572         unsigned long ptr;
1573
1574         key.objectid = inode_objectid;
1575         key.type = BTRFS_INODE_EXTREF_KEY;
1576         key.offset = start_off;
1577
1578         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1579         if (ret < 0)
1580                 return ret;
1581
1582         while (1) {
1583                 leaf = path->nodes[0];
1584                 slot = path->slots[0];
1585                 if (slot >= btrfs_header_nritems(leaf)) {
1586                         /*
1587                          * If the item at offset is not found,
1588                          * btrfs_search_slot will point us to the slot
1589                          * where it should be inserted. In our case
1590                          * that will be the slot directly before the
1591                          * next INODE_REF_KEY_V2 item. In the case
1592                          * that we're pointing to the last slot in a
1593                          * leaf, we must move one leaf over.
1594                          */
1595                         ret = btrfs_next_leaf(root, path);
1596                         if (ret) {
1597                                 if (ret >= 1)
1598                                         ret = -ENOENT;
1599                                 break;
1600                         }
1601                         continue;
1602                 }
1603
1604                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1605
1606                 /*
1607                  * Check that we're still looking at an extended ref key for
1608                  * this particular objectid. If we have different
1609                  * objectid or type then there are no more to be found
1610                  * in the tree and we can exit.
1611                  */
1612                 ret = -ENOENT;
1613                 if (found_key.objectid != inode_objectid)
1614                         break;
1615                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1616                         break;
1617
1618                 ret = 0;
1619                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1620                 extref = (struct btrfs_inode_extref *)ptr;
1621                 *ret_extref = extref;
1622                 if (found_off)
1623                         *found_off = found_key.offset;
1624                 break;
1625         }
1626
1627         return ret;
1628 }
1629
1630 /*
1631  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1632  * Elements of the path are separated by '/' and the path is guaranteed to be
1633  * 0-terminated. the path is only given within the current file system.
1634  * Therefore, it never starts with a '/'. the caller is responsible to provide
1635  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1636  * the start point of the resulting string is returned. this pointer is within
1637  * dest, normally.
1638  * in case the path buffer would overflow, the pointer is decremented further
1639  * as if output was written to the buffer, though no more output is actually
1640  * generated. that way, the caller can determine how much space would be
1641  * required for the path to fit into the buffer. in that case, the returned
1642  * value will be smaller than dest. callers must check this!
1643  */
1644 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1645                         u32 name_len, unsigned long name_off,
1646                         struct extent_buffer *eb_in, u64 parent,
1647                         char *dest, u32 size)
1648 {
1649         int slot;
1650         u64 next_inum;
1651         int ret;
1652         s64 bytes_left = ((s64)size) - 1;
1653         struct extent_buffer *eb = eb_in;
1654         struct btrfs_key found_key;
1655         int leave_spinning = path->leave_spinning;
1656         struct btrfs_inode_ref *iref;
1657
1658         if (bytes_left >= 0)
1659                 dest[bytes_left] = '\0';
1660
1661         path->leave_spinning = 1;
1662         while (1) {
1663                 bytes_left -= name_len;
1664                 if (bytes_left >= 0)
1665                         read_extent_buffer(eb, dest + bytes_left,
1666                                            name_off, name_len);
1667                 if (eb != eb_in) {
1668                         if (!path->skip_locking)
1669                                 btrfs_tree_read_unlock_blocking(eb);
1670                         free_extent_buffer(eb);
1671                 }
1672                 ret = btrfs_find_item(fs_root, path, parent, 0,
1673                                 BTRFS_INODE_REF_KEY, &found_key);
1674                 if (ret > 0)
1675                         ret = -ENOENT;
1676                 if (ret)
1677                         break;
1678
1679                 next_inum = found_key.offset;
1680
1681                 /* regular exit ahead */
1682                 if (parent == next_inum)
1683                         break;
1684
1685                 slot = path->slots[0];
1686                 eb = path->nodes[0];
1687                 /* make sure we can use eb after releasing the path */
1688                 if (eb != eb_in) {
1689                         if (!path->skip_locking)
1690                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1691                         path->nodes[0] = NULL;
1692                         path->locks[0] = 0;
1693                 }
1694                 btrfs_release_path(path);
1695                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1696
1697                 name_len = btrfs_inode_ref_name_len(eb, iref);
1698                 name_off = (unsigned long)(iref + 1);
1699
1700                 parent = next_inum;
1701                 --bytes_left;
1702                 if (bytes_left >= 0)
1703                         dest[bytes_left] = '/';
1704         }
1705
1706         btrfs_release_path(path);
1707         path->leave_spinning = leave_spinning;
1708
1709         if (ret)
1710                 return ERR_PTR(ret);
1711
1712         return dest + bytes_left;
1713 }
1714
1715 /*
1716  * this makes the path point to (logical EXTENT_ITEM *)
1717  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1718  * tree blocks and <0 on error.
1719  */
1720 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1721                         struct btrfs_path *path, struct btrfs_key *found_key,
1722                         u64 *flags_ret)
1723 {
1724         int ret;
1725         u64 flags;
1726         u64 size = 0;
1727         u32 item_size;
1728         const struct extent_buffer *eb;
1729         struct btrfs_extent_item *ei;
1730         struct btrfs_key key;
1731
1732         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1733                 key.type = BTRFS_METADATA_ITEM_KEY;
1734         else
1735                 key.type = BTRFS_EXTENT_ITEM_KEY;
1736         key.objectid = logical;
1737         key.offset = (u64)-1;
1738
1739         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1740         if (ret < 0)
1741                 return ret;
1742
1743         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1744         if (ret) {
1745                 if (ret > 0)
1746                         ret = -ENOENT;
1747                 return ret;
1748         }
1749         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1750         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1751                 size = fs_info->nodesize;
1752         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1753                 size = found_key->offset;
1754
1755         if (found_key->objectid > logical ||
1756             found_key->objectid + size <= logical) {
1757                 btrfs_debug(fs_info,
1758                         "logical %llu is not within any extent", logical);
1759                 return -ENOENT;
1760         }
1761
1762         eb = path->nodes[0];
1763         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1764         BUG_ON(item_size < sizeof(*ei));
1765
1766         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1767         flags = btrfs_extent_flags(eb, ei);
1768
1769         btrfs_debug(fs_info,
1770                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1771                  logical, logical - found_key->objectid, found_key->objectid,
1772                  found_key->offset, flags, item_size);
1773
1774         WARN_ON(!flags_ret);
1775         if (flags_ret) {
1776                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1777                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1778                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1779                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1780                 else
1781                         BUG_ON(1);
1782                 return 0;
1783         }
1784
1785         return -EIO;
1786 }
1787
1788 /*
1789  * helper function to iterate extent inline refs. ptr must point to a 0 value
1790  * for the first call and may be modified. it is used to track state.
1791  * if more refs exist, 0 is returned and the next call to
1792  * get_extent_inline_ref must pass the modified ptr parameter to get the
1793  * next ref. after the last ref was processed, 1 is returned.
1794  * returns <0 on error
1795  */
1796 static int get_extent_inline_ref(unsigned long *ptr,
1797                                  const struct extent_buffer *eb,
1798                                  const struct btrfs_key *key,
1799                                  const struct btrfs_extent_item *ei,
1800                                  u32 item_size,
1801                                  struct btrfs_extent_inline_ref **out_eiref,
1802                                  int *out_type)
1803 {
1804         unsigned long end;
1805         u64 flags;
1806         struct btrfs_tree_block_info *info;
1807
1808         if (!*ptr) {
1809                 /* first call */
1810                 flags = btrfs_extent_flags(eb, ei);
1811                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1812                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1813                                 /* a skinny metadata extent */
1814                                 *out_eiref =
1815                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1816                         } else {
1817                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1818                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1819                                 *out_eiref =
1820                                    (struct btrfs_extent_inline_ref *)(info + 1);
1821                         }
1822                 } else {
1823                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1824                 }
1825                 *ptr = (unsigned long)*out_eiref;
1826                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1827                         return -ENOENT;
1828         }
1829
1830         end = (unsigned long)ei + item_size;
1831         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1832         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1833                                                      BTRFS_REF_TYPE_ANY);
1834         if (*out_type == BTRFS_REF_TYPE_INVALID)
1835                 return -EUCLEAN;
1836
1837         *ptr += btrfs_extent_inline_ref_size(*out_type);
1838         WARN_ON(*ptr > end);
1839         if (*ptr == end)
1840                 return 1; /* last */
1841
1842         return 0;
1843 }
1844
1845 /*
1846  * reads the tree block backref for an extent. tree level and root are returned
1847  * through out_level and out_root. ptr must point to a 0 value for the first
1848  * call and may be modified (see get_extent_inline_ref comment).
1849  * returns 0 if data was provided, 1 if there was no more data to provide or
1850  * <0 on error.
1851  */
1852 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1853                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1854                             u32 item_size, u64 *out_root, u8 *out_level)
1855 {
1856         int ret;
1857         int type;
1858         struct btrfs_extent_inline_ref *eiref;
1859
1860         if (*ptr == (unsigned long)-1)
1861                 return 1;
1862
1863         while (1) {
1864                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1865                                               &eiref, &type);
1866                 if (ret < 0)
1867                         return ret;
1868
1869                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1870                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1871                         break;
1872
1873                 if (ret == 1)
1874                         return 1;
1875         }
1876
1877         /* we can treat both ref types equally here */
1878         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1879
1880         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1881                 struct btrfs_tree_block_info *info;
1882
1883                 info = (struct btrfs_tree_block_info *)(ei + 1);
1884                 *out_level = btrfs_tree_block_level(eb, info);
1885         } else {
1886                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1887                 *out_level = (u8)key->offset;
1888         }
1889
1890         if (ret == 1)
1891                 *ptr = (unsigned long)-1;
1892
1893         return 0;
1894 }
1895
1896 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1897                              struct extent_inode_elem *inode_list,
1898                              u64 root, u64 extent_item_objectid,
1899                              iterate_extent_inodes_t *iterate, void *ctx)
1900 {
1901         struct extent_inode_elem *eie;
1902         int ret = 0;
1903
1904         for (eie = inode_list; eie; eie = eie->next) {
1905                 btrfs_debug(fs_info,
1906                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1907                             extent_item_objectid, eie->inum,
1908                             eie->offset, root);
1909                 ret = iterate(eie->inum, eie->offset, root, ctx);
1910                 if (ret) {
1911                         btrfs_debug(fs_info,
1912                                     "stopping iteration for %llu due to ret=%d",
1913                                     extent_item_objectid, ret);
1914                         break;
1915                 }
1916         }
1917
1918         return ret;
1919 }
1920
1921 /*
1922  * calls iterate() for every inode that references the extent identified by
1923  * the given parameters.
1924  * when the iterator function returns a non-zero value, iteration stops.
1925  */
1926 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1927                                 u64 extent_item_objectid, u64 extent_item_pos,
1928                                 int search_commit_root,
1929                                 iterate_extent_inodes_t *iterate, void *ctx,
1930                                 bool ignore_offset)
1931 {
1932         int ret;
1933         struct btrfs_trans_handle *trans = NULL;
1934         struct ulist *refs = NULL;
1935         struct ulist *roots = NULL;
1936         struct ulist_node *ref_node = NULL;
1937         struct ulist_node *root_node = NULL;
1938         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1939         struct ulist_iterator ref_uiter;
1940         struct ulist_iterator root_uiter;
1941
1942         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1943                         extent_item_objectid);
1944
1945         if (!search_commit_root) {
1946                 trans = btrfs_attach_transaction(fs_info->extent_root);
1947                 if (IS_ERR(trans)) {
1948                         if (PTR_ERR(trans) != -ENOENT &&
1949                             PTR_ERR(trans) != -EROFS)
1950                                 return PTR_ERR(trans);
1951                         trans = NULL;
1952                 }
1953         }
1954
1955         if (trans)
1956                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1957         else
1958                 down_read(&fs_info->commit_root_sem);
1959
1960         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1961                                    tree_mod_seq_elem.seq, &refs,
1962                                    &extent_item_pos, ignore_offset);
1963         if (ret)
1964                 goto out;
1965
1966         ULIST_ITER_INIT(&ref_uiter);
1967         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1968                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1969                                                 tree_mod_seq_elem.seq, &roots,
1970                                                 ignore_offset);
1971                 if (ret)
1972                         break;
1973                 ULIST_ITER_INIT(&root_uiter);
1974                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1975                         btrfs_debug(fs_info,
1976                                     "root %llu references leaf %llu, data list %#llx",
1977                                     root_node->val, ref_node->val,
1978                                     ref_node->aux);
1979                         ret = iterate_leaf_refs(fs_info,
1980                                                 (struct extent_inode_elem *)
1981                                                 (uintptr_t)ref_node->aux,
1982                                                 root_node->val,
1983                                                 extent_item_objectid,
1984                                                 iterate, ctx);
1985                 }
1986                 ulist_free(roots);
1987         }
1988
1989         free_leaf_list(refs);
1990 out:
1991         if (trans) {
1992                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1993                 btrfs_end_transaction(trans);
1994         } else {
1995                 up_read(&fs_info->commit_root_sem);
1996         }
1997
1998         return ret;
1999 }
2000
2001 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2002                                 struct btrfs_path *path,
2003                                 iterate_extent_inodes_t *iterate, void *ctx,
2004                                 bool ignore_offset)
2005 {
2006         int ret;
2007         u64 extent_item_pos;
2008         u64 flags = 0;
2009         struct btrfs_key found_key;
2010         int search_commit_root = path->search_commit_root;
2011
2012         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2013         btrfs_release_path(path);
2014         if (ret < 0)
2015                 return ret;
2016         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2017                 return -EINVAL;
2018
2019         extent_item_pos = logical - found_key.objectid;
2020         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2021                                         extent_item_pos, search_commit_root,
2022                                         iterate, ctx, ignore_offset);
2023
2024         return ret;
2025 }
2026
2027 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2028                               struct extent_buffer *eb, void *ctx);
2029
2030 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2031                               struct btrfs_path *path,
2032                               iterate_irefs_t *iterate, void *ctx)
2033 {
2034         int ret = 0;
2035         int slot;
2036         u32 cur;
2037         u32 len;
2038         u32 name_len;
2039         u64 parent = 0;
2040         int found = 0;
2041         struct extent_buffer *eb;
2042         struct btrfs_item *item;
2043         struct btrfs_inode_ref *iref;
2044         struct btrfs_key found_key;
2045
2046         while (!ret) {
2047                 ret = btrfs_find_item(fs_root, path, inum,
2048                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2049                                 &found_key);
2050
2051                 if (ret < 0)
2052                         break;
2053                 if (ret) {
2054                         ret = found ? 0 : -ENOENT;
2055                         break;
2056                 }
2057                 ++found;
2058
2059                 parent = found_key.offset;
2060                 slot = path->slots[0];
2061                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2062                 if (!eb) {
2063                         ret = -ENOMEM;
2064                         break;
2065                 }
2066                 extent_buffer_get(eb);
2067                 btrfs_tree_read_lock(eb);
2068                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2069                 btrfs_release_path(path);
2070
2071                 item = btrfs_item_nr(slot);
2072                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2073
2074                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2075                         name_len = btrfs_inode_ref_name_len(eb, iref);
2076                         /* path must be released before calling iterate()! */
2077                         btrfs_debug(fs_root->fs_info,
2078                                 "following ref at offset %u for inode %llu in tree %llu",
2079                                 cur, found_key.objectid, fs_root->objectid);
2080                         ret = iterate(parent, name_len,
2081                                       (unsigned long)(iref + 1), eb, ctx);
2082                         if (ret)
2083                                 break;
2084                         len = sizeof(*iref) + name_len;
2085                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2086                 }
2087                 btrfs_tree_read_unlock_blocking(eb);
2088                 free_extent_buffer(eb);
2089         }
2090
2091         btrfs_release_path(path);
2092
2093         return ret;
2094 }
2095
2096 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2097                                  struct btrfs_path *path,
2098                                  iterate_irefs_t *iterate, void *ctx)
2099 {
2100         int ret;
2101         int slot;
2102         u64 offset = 0;
2103         u64 parent;
2104         int found = 0;
2105         struct extent_buffer *eb;
2106         struct btrfs_inode_extref *extref;
2107         u32 item_size;
2108         u32 cur_offset;
2109         unsigned long ptr;
2110
2111         while (1) {
2112                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2113                                             &offset);
2114                 if (ret < 0)
2115                         break;
2116                 if (ret) {
2117                         ret = found ? 0 : -ENOENT;
2118                         break;
2119                 }
2120                 ++found;
2121
2122                 slot = path->slots[0];
2123                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2124                 if (!eb) {
2125                         ret = -ENOMEM;
2126                         break;
2127                 }
2128                 extent_buffer_get(eb);
2129
2130                 btrfs_tree_read_lock(eb);
2131                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2132                 btrfs_release_path(path);
2133
2134                 item_size = btrfs_item_size_nr(eb, slot);
2135                 ptr = btrfs_item_ptr_offset(eb, slot);
2136                 cur_offset = 0;
2137
2138                 while (cur_offset < item_size) {
2139                         u32 name_len;
2140
2141                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2142                         parent = btrfs_inode_extref_parent(eb, extref);
2143                         name_len = btrfs_inode_extref_name_len(eb, extref);
2144                         ret = iterate(parent, name_len,
2145                                       (unsigned long)&extref->name, eb, ctx);
2146                         if (ret)
2147                                 break;
2148
2149                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2150                         cur_offset += sizeof(*extref);
2151                 }
2152                 btrfs_tree_read_unlock_blocking(eb);
2153                 free_extent_buffer(eb);
2154
2155                 offset++;
2156         }
2157
2158         btrfs_release_path(path);
2159
2160         return ret;
2161 }
2162
2163 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2164                          struct btrfs_path *path, iterate_irefs_t *iterate,
2165                          void *ctx)
2166 {
2167         int ret;
2168         int found_refs = 0;
2169
2170         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2171         if (!ret)
2172                 ++found_refs;
2173         else if (ret != -ENOENT)
2174                 return ret;
2175
2176         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2177         if (ret == -ENOENT && found_refs)
2178                 return 0;
2179
2180         return ret;
2181 }
2182
2183 /*
2184  * returns 0 if the path could be dumped (probably truncated)
2185  * returns <0 in case of an error
2186  */
2187 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2188                          struct extent_buffer *eb, void *ctx)
2189 {
2190         struct inode_fs_paths *ipath = ctx;
2191         char *fspath;
2192         char *fspath_min;
2193         int i = ipath->fspath->elem_cnt;
2194         const int s_ptr = sizeof(char *);
2195         u32 bytes_left;
2196
2197         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2198                                         ipath->fspath->bytes_left - s_ptr : 0;
2199
2200         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2201         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2202                                    name_off, eb, inum, fspath_min, bytes_left);
2203         if (IS_ERR(fspath))
2204                 return PTR_ERR(fspath);
2205
2206         if (fspath > fspath_min) {
2207                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2208                 ++ipath->fspath->elem_cnt;
2209                 ipath->fspath->bytes_left = fspath - fspath_min;
2210         } else {
2211                 ++ipath->fspath->elem_missed;
2212                 ipath->fspath->bytes_missing += fspath_min - fspath;
2213                 ipath->fspath->bytes_left = 0;
2214         }
2215
2216         return 0;
2217 }
2218
2219 /*
2220  * this dumps all file system paths to the inode into the ipath struct, provided
2221  * is has been created large enough. each path is zero-terminated and accessed
2222  * from ipath->fspath->val[i].
2223  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2224  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2225  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2226  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2227  * have been needed to return all paths.
2228  */
2229 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2230 {
2231         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2232                              inode_to_path, ipath);
2233 }
2234
2235 struct btrfs_data_container *init_data_container(u32 total_bytes)
2236 {
2237         struct btrfs_data_container *data;
2238         size_t alloc_bytes;
2239
2240         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2241         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2242         if (!data)
2243                 return ERR_PTR(-ENOMEM);
2244
2245         if (total_bytes >= sizeof(*data)) {
2246                 data->bytes_left = total_bytes - sizeof(*data);
2247                 data->bytes_missing = 0;
2248         } else {
2249                 data->bytes_missing = sizeof(*data) - total_bytes;
2250                 data->bytes_left = 0;
2251         }
2252
2253         data->elem_cnt = 0;
2254         data->elem_missed = 0;
2255
2256         return data;
2257 }
2258
2259 /*
2260  * allocates space to return multiple file system paths for an inode.
2261  * total_bytes to allocate are passed, note that space usable for actual path
2262  * information will be total_bytes - sizeof(struct inode_fs_paths).
2263  * the returned pointer must be freed with free_ipath() in the end.
2264  */
2265 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2266                                         struct btrfs_path *path)
2267 {
2268         struct inode_fs_paths *ifp;
2269         struct btrfs_data_container *fspath;
2270
2271         fspath = init_data_container(total_bytes);
2272         if (IS_ERR(fspath))
2273                 return ERR_CAST(fspath);
2274
2275         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2276         if (!ifp) {
2277                 kvfree(fspath);
2278                 return ERR_PTR(-ENOMEM);
2279         }
2280
2281         ifp->btrfs_path = path;
2282         ifp->fspath = fspath;
2283         ifp->fs_root = fs_root;
2284
2285         return ifp;
2286 }
2287
2288 void free_ipath(struct inode_fs_paths *ipath)
2289 {
2290         if (!ipath)
2291                 return;
2292         kvfree(ipath->fspath);
2293         kfree(ipath);
2294 }