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_cached root;
 116         unsigned int count;
 117 };
 118
 119 #define PREFTREE_INIT   { .root = RB_ROOT_CACHED, .count = 0 }
 120
 121 struct preftrees {
 122         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 123         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 124         struct preftree indirect_missing_keys;
 125 };
 126
 127 /*
 128  * Checks for a shared extent during backref search.
 129  *
 130  * The share_count tracks prelim_refs (direct and indirect) having a
 131  * ref->count >0:
 132  *  - incremented when a ref->count transitions to >0
 133  *  - decremented when a ref->count transitions to <1
 134  */
 135 struct share_check {
 136         u64 root_objectid;
 137         u64 inum;
 138         int share_count;
 139 };
 140
 141 static inline int extent_is_shared(struct share_check *sc)
 142 {
 143         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 144 }
 145
 146 static struct kmem_cache *btrfs_prelim_ref_cache;
 147
 148 int __init btrfs_prelim_ref_init(void)
 149 {
 150         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 151                                         sizeof(struct prelim_ref),
 152                                         0,
 153                                         SLAB_MEM_SPREAD,
 154                                         NULL);
 155         if (!btrfs_prelim_ref_cache)
 156                 return -ENOMEM;
 157         return 0;
 158 }
 159
 160 void __cold btrfs_prelim_ref_exit(void)
 161 {
 162         kmem_cache_destroy(btrfs_prelim_ref_cache);
 163 }
 164
 165 static void free_pref(struct prelim_ref *ref)
 166 {
 167         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 168 }
 169
 170 /*
 171  * Return 0 when both refs are for the same block (and can be merged).
 172  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 173  * indicates a 'higher' block.
 174  */
 175 static int prelim_ref_compare(struct prelim_ref *ref1,
 176                               struct prelim_ref *ref2)
 177 {
 178         if (ref1->level < ref2->level)
 179                 return -1;
 180         if (ref1->level > ref2->level)
 181                 return 1;
 182         if (ref1->root_id < ref2->root_id)
 183                 return -1;
 184         if (ref1->root_id > ref2->root_id)
 185                 return 1;
 186         if (ref1->key_for_search.type < ref2->key_for_search.type)
 187                 return -1;
 188         if (ref1->key_for_search.type > ref2->key_for_search.type)
 189                 return 1;
 190         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 191                 return -1;
 192         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 193                 return 1;
 194         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 195                 return -1;
 196         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 197                 return 1;
 198         if (ref1->parent < ref2->parent)
 199                 return -1;
 200         if (ref1->parent > ref2->parent)
 201                 return 1;
 202
 203         return 0;
 204 }
 205
 206 static void update_share_count(struct share_check *sc, int oldcount,
 207                                int newcount)
 208 {
 209         if ((!sc) || (oldcount == 0 && newcount < 1))
 210                 return;
 211
 212         if (oldcount > 0 && newcount < 1)
 213                 sc->share_count--;
 214         else if (oldcount < 1 && newcount > 0)
 215                 sc->share_count++;
 216 }
 217
 218 /*
 219  * Add @newref to the @root rbtree, merging identical refs.
 220  *
 221  * Callers should assume that newref has been freed after calling.
 222  */
 223 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 224                               struct preftree *preftree,
 225                               struct prelim_ref *newref,
 226                               struct share_check *sc)
 227 {
 228         struct rb_root_cached *root;
 229         struct rb_node **p;
 230         struct rb_node *parent = NULL;
 231         struct prelim_ref *ref;
 232         int result;
 233         bool leftmost = true;
 234
 235         root = &preftree->root;
 236         p = &root->rb_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                         leftmost = false;
 247                 } else {
 248                         /* Identical refs, merge them and free @newref */
 249                         struct extent_inode_elem *eie = ref->inode_list;
 250
 251                         while (eie && eie->next)
 252                                 eie = eie->next;
 253
 254                         if (!eie)
 255                                 ref->inode_list = newref->inode_list;
 256                         else
 257                                 eie->next = newref->inode_list;
 258                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 259                                                      preftree->count);
 260                         /*
 261                          * A delayed ref can have newref->count < 0.
 262                          * The ref->count is updated to follow any
 263                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 264                          */
 265                         update_share_count(sc, ref->count,
 266                                            ref->count + newref->count);
 267                         ref->count += newref->count;
 268                         free_pref(newref);
 269                         return;
 270                 }
 271         }
 272
 273         update_share_count(sc, 0, newref->count);
 274         preftree->count++;
 275         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 276         rb_link_node(&newref->rbnode, parent, p);
 277         rb_insert_color_cached(&newref->rbnode, root, leftmost);
 278 }
 279
 280 /*
 281  * Release the entire tree.  We don't care about internal consistency so
 282  * just free everything and then reset the tree root.
 283  */
 284 static void prelim_release(struct preftree *preftree)
 285 {
 286         struct prelim_ref *ref, *next_ref;
 287
 288         rbtree_postorder_for_each_entry_safe(ref, next_ref,
 289                                              &preftree->root.rb_root, rbnode)
 290                 free_pref(ref);
 291
 292         preftree->root = RB_ROOT_CACHED;
 293         preftree->count = 0;
 294 }
 295
 296 /*
 297  * the rules for all callers of this function are:
 298  * - obtaining the parent is the goal
 299  * - if you add a key, you must know that it is a correct key
 300  * - if you cannot add the parent or a correct key, then we will look into the
 301  *   block later to set a correct key
 302  *
 303  * delayed refs
 304  * ============
 305  *        backref type | shared | indirect | shared | indirect
 306  * information         |   tree |     tree |   data |     data
 307  * --------------------+--------+----------+--------+----------
 308  *      parent logical |    y   |     -    |    -   |     -
 309  *      key to resolve |    -   |     y    |    y   |     y
 310  *  tree block logical |    -   |     -    |    -   |     -
 311  *  root for resolving |    y   |     y    |    y   |     y
 312  *
 313  * - column 1:       we've the parent -> done
 314  * - column 2, 3, 4: we use the key to find the parent
 315  *
 316  * on disk refs (inline or keyed)
 317  * ==============================
 318  *        backref type | shared | indirect | shared | indirect
 319  * information         |   tree |     tree |   data |     data
 320  * --------------------+--------+----------+--------+----------
 321  *      parent logical |    y   |     -    |    y   |     -
 322  *      key to resolve |    -   |     -    |    -   |     y
 323  *  tree block logical |    y   |     y    |    y   |     y
 324  *  root for resolving |    -   |     y    |    y   |     y
 325  *
 326  * - column 1, 3: we've the parent -> done
 327  * - column 2:    we take the first key from the block to find the parent
 328  *                (see add_missing_keys)
 329  * - column 4:    we use the key to find the parent
 330  *
 331  * additional information that's available but not required to find the parent
 332  * block might help in merging entries to gain some speed.
 333  */
 334 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 335                           struct preftree *preftree, u64 root_id,
 336                           const struct btrfs_key *key, int level, u64 parent,
 337                           u64 wanted_disk_byte, int count,
 338                           struct share_check *sc, gfp_t gfp_mask)
 339 {
 340         struct prelim_ref *ref;
 341
 342         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 343                 return 0;
 344
 345         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 346         if (!ref)
 347                 return -ENOMEM;
 348
 349         ref->root_id = root_id;
 350         if (key)
 351                 ref->key_for_search = *key;
 352         else
 353                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 354
 355         ref->inode_list = NULL;
 356         ref->level = level;
 357         ref->count = count;
 358         ref->parent = parent;
 359         ref->wanted_disk_byte = wanted_disk_byte;
 360         prelim_ref_insert(fs_info, preftree, ref, sc);
 361         return extent_is_shared(sc);
 362 }
 363
 364 /* direct refs use root == 0, key == NULL */
 365 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 366                           struct preftrees *preftrees, int level, u64 parent,
 367                           u64 wanted_disk_byte, int count,
 368                           struct share_check *sc, gfp_t gfp_mask)
 369 {
 370         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 371                               parent, wanted_disk_byte, count, sc, gfp_mask);
 372 }
 373
 374 /* indirect refs use parent == 0 */
 375 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 376                             struct preftrees *preftrees, u64 root_id,
 377                             const struct btrfs_key *key, int level,
 378                             u64 wanted_disk_byte, int count,
 379                             struct share_check *sc, gfp_t gfp_mask)
 380 {
 381         struct preftree *tree = &preftrees->indirect;
 382
 383         if (!key)
 384                 tree = &preftrees->indirect_missing_keys;
 385         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 386                               wanted_disk_byte, count, sc, gfp_mask);
 387 }
 388
 389 static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
 390 {
 391         struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
 392         struct rb_node *parent = NULL;
 393         struct prelim_ref *ref = NULL;
 394         struct prelim_ref target = {0};
 395         int result;
 396
 397         target.parent = bytenr;
 398
 399         while (*p) {
 400                 parent = *p;
 401                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 402                 result = prelim_ref_compare(ref, &target);
 403
 404                 if (result < 0)
 405                         p = &(*p)->rb_left;
 406                 else if (result > 0)
 407                         p = &(*p)->rb_right;
 408                 else
 409                         return 1;
 410         }
 411         return 0;
 412 }
 413
 414 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 415                            struct ulist *parents,
 416                            struct preftrees *preftrees, struct prelim_ref *ref,
 417                            int level, u64 time_seq, const u64 *extent_item_pos,
 418                            bool ignore_offset)
 419 {
 420         int ret = 0;
 421         int slot;
 422         struct extent_buffer *eb;
 423         struct btrfs_key key;
 424         struct btrfs_key *key_for_search = &ref->key_for_search;
 425         struct btrfs_file_extent_item *fi;
 426         struct extent_inode_elem *eie = NULL, *old = NULL;
 427         u64 disk_byte;
 428         u64 wanted_disk_byte = ref->wanted_disk_byte;
 429         u64 count = 0;
 430         u64 data_offset;
 431
 432         if (level != 0) {
 433                 eb = path->nodes[level];
 434                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 435                 if (ret < 0)
 436                         return ret;
 437                 return 0;
 438         }
 439
 440         /*
 441          * 1. We normally enter this function with the path already pointing to
 442          *    the first item to check. But sometimes, we may enter it with
 443          *    slot == nritems.
 444          * 2. We are searching for normal backref but bytenr of this leaf
 445          *    matches shared data backref
 446          * 3. The leaf owner is not equal to the root we are searching
 447          *
 448          * For these cases, go to the next leaf before we continue.
 449          */
 450         eb = path->nodes[0];
 451         if (path->slots[0] >= btrfs_header_nritems(eb) ||
 452             is_shared_data_backref(preftrees, eb->start) ||
 453             ref->root_id != btrfs_header_owner(eb)) {
 454                 if (time_seq == SEQ_LAST)
 455                         ret = btrfs_next_leaf(root, path);
 456                 else
 457                         ret = btrfs_next_old_leaf(root, path, time_seq);
 458         }
 459
 460         while (!ret && count < ref->count) {
 461                 eb = path->nodes[0];
 462                 slot = path->slots[0];
 463
 464                 btrfs_item_key_to_cpu(eb, &key, slot);
 465
 466                 if (key.objectid != key_for_search->objectid ||
 467                     key.type != BTRFS_EXTENT_DATA_KEY)
 468                         break;
 469
 470                 /*
 471                  * We are searching for normal backref but bytenr of this leaf
 472                  * matches shared data backref, OR
 473                  * the leaf owner is not equal to the root we are searching for
 474                  */
 475                 if (slot == 0 &&
 476                     (is_shared_data_backref(preftrees, eb->start) ||
 477                      ref->root_id != btrfs_header_owner(eb))) {
 478                         if (time_seq == SEQ_LAST)
 479                                 ret = btrfs_next_leaf(root, path);
 480                         else
 481                                 ret = btrfs_next_old_leaf(root, path, time_seq);
 482                         continue;
 483                 }
 484                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 485                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 486                 data_offset = btrfs_file_extent_offset(eb, fi);
 487
 488                 if (disk_byte == wanted_disk_byte) {
 489                         eie = NULL;
 490                         old = NULL;
 491                         if (ref->key_for_search.offset == key.offset - data_offset)
 492                                 count++;
 493                         else
 494                                 goto next;
 495                         if (extent_item_pos) {
 496                                 ret = check_extent_in_eb(&key, eb, fi,
 497                                                 *extent_item_pos,
 498                                                 &eie, ignore_offset);
 499                                 if (ret < 0)
 500                                         break;
 501                         }
 502                         if (ret > 0)
 503                                 goto next;
 504                         ret = ulist_add_merge_ptr(parents, eb->start,
 505                                                   eie, (void **)&old, GFP_NOFS);
 506                         if (ret < 0)
 507                                 break;
 508                         if (!ret && extent_item_pos) {
 509                                 while (old->next)
 510                                         old = old->next;
 511                                 old->next = eie;
 512                         }
 513                         eie = NULL;
 514                 }
 515 next:
 516                 if (time_seq == SEQ_LAST)
 517                         ret = btrfs_next_item(root, path);
 518                 else
 519                         ret = btrfs_next_old_item(root, path, time_seq);
 520         }
 521
 522         if (ret > 0)
 523                 ret = 0;
 524         else if (ret < 0)
 525                 free_inode_elem_list(eie);
 526         return ret;
 527 }
 528
 529 /*
 530  * resolve an indirect backref in the form (root_id, key, level)
 531  * to a logical address
 532  */
 533 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 534                                 struct btrfs_path *path, u64 time_seq,
 535                                 struct preftrees *preftrees,
 536                                 struct prelim_ref *ref, struct ulist *parents,
 537                                 const u64 *extent_item_pos, bool ignore_offset)
 538 {
 539         struct btrfs_root *root;
 540         struct btrfs_key root_key;
 541         struct extent_buffer *eb;
 542         int ret = 0;
 543         int root_level;
 544         int level = ref->level;
 545         struct btrfs_key search_key = ref->key_for_search;
 546
 547         root_key.objectid = ref->root_id;
 548         root_key.type = BTRFS_ROOT_ITEM_KEY;
 549         root_key.offset = (u64)-1;
 550
 551         root = btrfs_get_fs_root(fs_info, &root_key, false);
 552         if (IS_ERR(root)) {
 553                 ret = PTR_ERR(root);
 554                 goto out_free;
 555         }
 556
 557         if (btrfs_is_testing(fs_info)) {
 558                 ret = -ENOENT;
 559                 goto out;
 560         }
 561
 562         if (path->search_commit_root)
 563                 root_level = btrfs_header_level(root->commit_root);
 564         else if (time_seq == SEQ_LAST)
 565                 root_level = btrfs_header_level(root->node);
 566         else
 567                 root_level = btrfs_old_root_level(root, time_seq);
 568
 569         if (root_level + 1 == level)
 570                 goto out;
 571
 572         /*
 573          * We can often find data backrefs with an offset that is too large
 574          * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
 575          * subtracting a file's offset with the data offset of its
 576          * corresponding extent data item. This can happen for example in the
 577          * clone ioctl.
 578          *
 579          * So if we detect such case we set the search key's offset to zero to
 580          * make sure we will find the matching file extent item at
 581          * add_all_parents(), otherwise we will miss it because the offset
 582          * taken form the backref is much larger then the offset of the file
 583          * extent item. This can make us scan a very large number of file
 584          * extent items, but at least it will not make us miss any.
 585          *
 586          * This is an ugly workaround for a behaviour that should have never
 587          * existed, but it does and a fix for the clone ioctl would touch a lot
 588          * of places, cause backwards incompatibility and would not fix the
 589          * problem for extents cloned with older kernels.
 590          */
 591         if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
 592             search_key.offset >= LLONG_MAX)
 593                 search_key.offset = 0;
 594         path->lowest_level = level;
 595         if (time_seq == SEQ_LAST)
 596                 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 597         else
 598                 ret = btrfs_search_old_slot(root, &search_key, path, time_seq);
 599
 600         btrfs_debug(fs_info,
 601                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 602                  ref->root_id, level, ref->count, ret,
 603                  ref->key_for_search.objectid, ref->key_for_search.type,
 604                  ref->key_for_search.offset);
 605         if (ret < 0)
 606                 goto out;
 607
 608         eb = path->nodes[level];
 609         while (!eb) {
 610                 if (WARN_ON(!level)) {
 611                         ret = 1;
 612                         goto out;
 613                 }
 614                 level--;
 615                 eb = path->nodes[level];
 616         }
 617
 618         ret = add_all_parents(root, path, parents, preftrees, ref, level,
 619                               time_seq, extent_item_pos, ignore_offset);
 620 out:
 621         btrfs_put_root(root);
 622 out_free:
 623         path->lowest_level = 0;
 624         btrfs_release_path(path);
 625         return ret;
 626 }
 627
 628 static struct extent_inode_elem *
 629 unode_aux_to_inode_list(struct ulist_node *node)
 630 {
 631         if (!node)
 632                 return NULL;
 633         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 634 }
 635
 636 /*
 637  * We maintain three separate rbtrees: one for direct refs, one for
 638  * indirect refs which have a key, and one for indirect refs which do not
 639  * have a key. Each tree does merge on insertion.
 640  *
 641  * Once all of the references are located, we iterate over the tree of
 642  * indirect refs with missing keys. An appropriate key is located and
 643  * the ref is moved onto the tree for indirect refs. After all missing
 644  * keys are thus located, we iterate over the indirect ref tree, resolve
 645  * each reference, and then insert the resolved reference onto the
 646  * direct tree (merging there too).
 647  *
 648  * New backrefs (i.e., for parent nodes) are added to the appropriate
 649  * rbtree as they are encountered. The new backrefs are subsequently
 650  * resolved as above.
 651  */
 652 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 653                                  struct btrfs_path *path, u64 time_seq,
 654                                  struct preftrees *preftrees,
 655                                  const u64 *extent_item_pos,
 656                                  struct share_check *sc, bool ignore_offset)
 657 {
 658         int err;
 659         int ret = 0;
 660         struct ulist *parents;
 661         struct ulist_node *node;
 662         struct ulist_iterator uiter;
 663         struct rb_node *rnode;
 664
 665         parents = ulist_alloc(GFP_NOFS);
 666         if (!parents)
 667                 return -ENOMEM;
 668
 669         /*
 670          * We could trade memory usage for performance here by iterating
 671          * the tree, allocating new refs for each insertion, and then
 672          * freeing the entire indirect tree when we're done.  In some test
 673          * cases, the tree can grow quite large (~200k objects).
 674          */
 675         while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
 676                 struct prelim_ref *ref;
 677
 678                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 679                 if (WARN(ref->parent,
 680                          "BUG: direct ref found in indirect tree")) {
 681                         ret = -EINVAL;
 682                         goto out;
 683                 }
 684
 685                 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
 686                 preftrees->indirect.count--;
 687
 688                 if (ref->count == 0) {
 689                         free_pref(ref);
 690                         continue;
 691                 }
 692
 693                 if (sc && sc->root_objectid &&
 694                     ref->root_id != sc->root_objectid) {
 695                         free_pref(ref);
 696                         ret = BACKREF_FOUND_SHARED;
 697                         goto out;
 698                 }
 699                 err = resolve_indirect_ref(fs_info, path, time_seq, preftrees,
 700                                            ref, parents, extent_item_pos,
 701                                            ignore_offset);
 702                 /*
 703                  * we can only tolerate ENOENT,otherwise,we should catch error
 704                  * and return directly.
 705                  */
 706                 if (err == -ENOENT) {
 707                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 708                                           NULL);
 709                         continue;
 710                 } else if (err) {
 711                         free_pref(ref);
 712                         ret = err;
 713                         goto out;
 714                 }
 715
 716                 /* we put the first parent into the ref at hand */
 717                 ULIST_ITER_INIT(&uiter);
 718                 node = ulist_next(parents, &uiter);
 719                 ref->parent = node ? node->val : 0;
 720                 ref->inode_list = unode_aux_to_inode_list(node);
 721
 722                 /* Add a prelim_ref(s) for any other parent(s). */
 723                 while ((node = ulist_next(parents, &uiter))) {
 724                         struct prelim_ref *new_ref;
 725
 726                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 727                                                    GFP_NOFS);
 728                         if (!new_ref) {
 729                                 free_pref(ref);
 730                                 ret = -ENOMEM;
 731                                 goto out;
 732                         }
 733                         memcpy(new_ref, ref, sizeof(*ref));
 734                         new_ref->parent = node->val;
 735                         new_ref->inode_list = unode_aux_to_inode_list(node);
 736                         prelim_ref_insert(fs_info, &preftrees->direct,
 737                                           new_ref, NULL);
 738                 }
 739
 740                 /*
 741                  * Now it's a direct ref, put it in the direct tree. We must
 742                  * do this last because the ref could be merged/freed here.
 743                  */
 744                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 745
 746                 ulist_reinit(parents);
 747                 cond_resched();
 748         }
 749 out:
 750         ulist_free(parents);
 751         return ret;
 752 }
 753
 754 /*
 755  * read tree blocks and add keys where required.
 756  */
 757 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 758                             struct preftrees *preftrees, bool lock)
 759 {
 760         struct prelim_ref *ref;
 761         struct extent_buffer *eb;
 762         struct preftree *tree = &preftrees->indirect_missing_keys;
 763         struct rb_node *node;
 764
 765         while ((node = rb_first_cached(&tree->root))) {
 766                 ref = rb_entry(node, struct prelim_ref, rbnode);
 767                 rb_erase_cached(node, &tree->root);
 768
 769                 BUG_ON(ref->parent);    /* should not be a direct ref */
 770                 BUG_ON(ref->key_for_search.type);
 771                 BUG_ON(!ref->wanted_disk_byte);
 772
 773                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0,
 774                                      ref->level - 1, NULL);
 775                 if (IS_ERR(eb)) {
 776                         free_pref(ref);
 777                         return PTR_ERR(eb);
 778                 } else if (!extent_buffer_uptodate(eb)) {
 779                         free_pref(ref);
 780                         free_extent_buffer(eb);
 781                         return -EIO;
 782                 }
 783                 if (lock)
 784                         btrfs_tree_read_lock(eb);
 785                 if (btrfs_header_level(eb) == 0)
 786                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 787                 else
 788                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 789                 if (lock)
 790                         btrfs_tree_read_unlock(eb);
 791                 free_extent_buffer(eb);
 792                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 793                 cond_resched();
 794         }
 795         return 0;
 796 }
 797
 798 /*
 799  * add all currently queued delayed refs from this head whose seq nr is
 800  * smaller or equal that seq to the list
 801  */
 802 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 803                             struct btrfs_delayed_ref_head *head, u64 seq,
 804                             struct preftrees *preftrees, struct share_check *sc)
 805 {
 806         struct btrfs_delayed_ref_node *node;
 807         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 808         struct btrfs_key key;
 809         struct btrfs_key tmp_op_key;
 810         struct rb_node *n;
 811         int count;
 812         int ret = 0;
 813
 814         if (extent_op && extent_op->update_key)
 815                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 816
 817         spin_lock(&head->lock);
 818         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 819                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 820                                 ref_node);
 821                 if (node->seq > seq)
 822                         continue;
 823
 824                 switch (node->action) {
 825                 case BTRFS_ADD_DELAYED_EXTENT:
 826                 case BTRFS_UPDATE_DELAYED_HEAD:
 827                         WARN_ON(1);
 828                         continue;
 829                 case BTRFS_ADD_DELAYED_REF:
 830                         count = node->ref_mod;
 831                         break;
 832                 case BTRFS_DROP_DELAYED_REF:
 833                         count = node->ref_mod * -1;
 834                         break;
 835                 default:
 836                         BUG();
 837                 }
 838                 switch (node->type) {
 839                 case BTRFS_TREE_BLOCK_REF_KEY: {
 840                         /* NORMAL INDIRECT METADATA backref */
 841                         struct btrfs_delayed_tree_ref *ref;
 842
 843                         ref = btrfs_delayed_node_to_tree_ref(node);
 844                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 845                                                &tmp_op_key, ref->level + 1,
 846                                                node->bytenr, count, sc,
 847                                                GFP_ATOMIC);
 848                         break;
 849                 }
 850                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 851                         /* SHARED DIRECT METADATA backref */
 852                         struct btrfs_delayed_tree_ref *ref;
 853
 854                         ref = btrfs_delayed_node_to_tree_ref(node);
 855
 856                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 857                                              ref->parent, node->bytenr, count,
 858                                              sc, GFP_ATOMIC);
 859                         break;
 860                 }
 861                 case BTRFS_EXTENT_DATA_REF_KEY: {
 862                         /* NORMAL INDIRECT DATA backref */
 863                         struct btrfs_delayed_data_ref *ref;
 864                         ref = btrfs_delayed_node_to_data_ref(node);
 865
 866                         key.objectid = ref->objectid;
 867                         key.type = BTRFS_EXTENT_DATA_KEY;
 868                         key.offset = ref->offset;
 869
 870                         /*
 871                          * Found a inum that doesn't match our known inum, we
 872                          * know it's shared.
 873                          */
 874                         if (sc && sc->inum && ref->objectid != sc->inum) {
 875                                 ret = BACKREF_FOUND_SHARED;
 876                                 goto out;
 877                         }
 878
 879                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 880                                                &key, 0, node->bytenr, count, sc,
 881                                                GFP_ATOMIC);
 882                         break;
 883                 }
 884                 case BTRFS_SHARED_DATA_REF_KEY: {
 885                         /* SHARED DIRECT FULL backref */
 886                         struct btrfs_delayed_data_ref *ref;
 887
 888                         ref = btrfs_delayed_node_to_data_ref(node);
 889
 890                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 891                                              node->bytenr, count, sc,
 892                                              GFP_ATOMIC);
 893                         break;
 894                 }
 895                 default:
 896                         WARN_ON(1);
 897                 }
 898                 /*
 899                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 900                  * refs have been checked.
 901                  */
 902                 if (ret && (ret != BACKREF_FOUND_SHARED))
 903                         break;
 904         }
 905         if (!ret)
 906                 ret = extent_is_shared(sc);
 907 out:
 908         spin_unlock(&head->lock);
 909         return ret;
 910 }
 911
 912 /*
 913  * add all inline backrefs for bytenr to the list
 914  *
 915  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 916  */
 917 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 918                            struct btrfs_path *path, u64 bytenr,
 919                            int *info_level, struct preftrees *preftrees,
 920                            struct share_check *sc)
 921 {
 922         int ret = 0;
 923         int slot;
 924         struct extent_buffer *leaf;
 925         struct btrfs_key key;
 926         struct btrfs_key found_key;
 927         unsigned long ptr;
 928         unsigned long end;
 929         struct btrfs_extent_item *ei;
 930         u64 flags;
 931         u64 item_size;
 932
 933         /*
 934          * enumerate all inline refs
 935          */
 936         leaf = path->nodes[0];
 937         slot = path->slots[0];
 938
 939         item_size = btrfs_item_size_nr(leaf, slot);
 940         BUG_ON(item_size < sizeof(*ei));
 941
 942         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 943         flags = btrfs_extent_flags(leaf, ei);
 944         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 945
 946         ptr = (unsigned long)(ei + 1);
 947         end = (unsigned long)ei + item_size;
 948
 949         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 950             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 951                 struct btrfs_tree_block_info *info;
 952
 953                 info = (struct btrfs_tree_block_info *)ptr;
 954                 *info_level = btrfs_tree_block_level(leaf, info);
 955                 ptr += sizeof(struct btrfs_tree_block_info);
 956                 BUG_ON(ptr > end);
 957         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 958                 *info_level = found_key.offset;
 959         } else {
 960                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 961         }
 962
 963         while (ptr < end) {
 964                 struct btrfs_extent_inline_ref *iref;
 965                 u64 offset;
 966                 int type;
 967
 968                 iref = (struct btrfs_extent_inline_ref *)ptr;
 969                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 970                                                         BTRFS_REF_TYPE_ANY);
 971                 if (type == BTRFS_REF_TYPE_INVALID)
 972                         return -EUCLEAN;
 973
 974                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 975
 976                 switch (type) {
 977                 case BTRFS_SHARED_BLOCK_REF_KEY:
 978                         ret = add_direct_ref(fs_info, preftrees,
 979                                              *info_level + 1, offset,
 980                                              bytenr, 1, NULL, GFP_NOFS);
 981                         break;
 982                 case BTRFS_SHARED_DATA_REF_KEY: {
 983                         struct btrfs_shared_data_ref *sdref;
 984                         int count;
 985
 986                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 987                         count = btrfs_shared_data_ref_count(leaf, sdref);
 988
 989                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
 990                                              bytenr, count, sc, GFP_NOFS);
 991                         break;
 992                 }
 993                 case BTRFS_TREE_BLOCK_REF_KEY:
 994                         ret = add_indirect_ref(fs_info, preftrees, offset,
 995                                                NULL, *info_level + 1,
 996                                                bytenr, 1, NULL, GFP_NOFS);
 997                         break;
 998                 case BTRFS_EXTENT_DATA_REF_KEY: {
 999                         struct btrfs_extent_data_ref *dref;
1000                         int count;
1001                         u64 root;
1002
1003                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1004                         count = btrfs_extent_data_ref_count(leaf, dref);
1005                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1006                                                                       dref);
1007                         key.type = BTRFS_EXTENT_DATA_KEY;
1008                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1009
1010                         if (sc && sc->inum && key.objectid != sc->inum) {
1011                                 ret = BACKREF_FOUND_SHARED;
1012                                 break;
1013                         }
1014
1015                         root = btrfs_extent_data_ref_root(leaf, dref);
1016
1017                         ret = add_indirect_ref(fs_info, preftrees, root,
1018                                                &key, 0, bytenr, count,
1019                                                sc, GFP_NOFS);
1020                         break;
1021                 }
1022                 default:
1023                         WARN_ON(1);
1024                 }
1025                 if (ret)
1026                         return ret;
1027                 ptr += btrfs_extent_inline_ref_size(type);
1028         }
1029
1030         return 0;
1031 }
1032
1033 /*
1034  * add all non-inline backrefs for bytenr to the list
1035  *
1036  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1037  */
1038 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1039                           struct btrfs_path *path, u64 bytenr,
1040                           int info_level, struct preftrees *preftrees,
1041                           struct share_check *sc)
1042 {
1043         struct btrfs_root *extent_root = fs_info->extent_root;
1044         int ret;
1045         int slot;
1046         struct extent_buffer *leaf;
1047         struct btrfs_key key;
1048
1049         while (1) {
1050                 ret = btrfs_next_item(extent_root, path);
1051                 if (ret < 0)
1052                         break;
1053                 if (ret) {
1054                         ret = 0;
1055                         break;
1056                 }
1057
1058                 slot = path->slots[0];
1059                 leaf = path->nodes[0];
1060                 btrfs_item_key_to_cpu(leaf, &key, slot);
1061
1062                 if (key.objectid != bytenr)
1063                         break;
1064                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1065                         continue;
1066                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1067                         break;
1068
1069                 switch (key.type) {
1070                 case BTRFS_SHARED_BLOCK_REF_KEY:
1071                         /* SHARED DIRECT METADATA backref */
1072                         ret = add_direct_ref(fs_info, preftrees,
1073                                              info_level + 1, key.offset,
1074                                              bytenr, 1, NULL, GFP_NOFS);
1075                         break;
1076                 case BTRFS_SHARED_DATA_REF_KEY: {
1077                         /* SHARED DIRECT FULL backref */
1078                         struct btrfs_shared_data_ref *sdref;
1079                         int count;
1080
1081                         sdref = btrfs_item_ptr(leaf, slot,
1082                                               struct btrfs_shared_data_ref);
1083                         count = btrfs_shared_data_ref_count(leaf, sdref);
1084                         ret = add_direct_ref(fs_info, preftrees, 0,
1085                                              key.offset, bytenr, count,
1086                                              sc, GFP_NOFS);
1087                         break;
1088                 }
1089                 case BTRFS_TREE_BLOCK_REF_KEY:
1090                         /* NORMAL INDIRECT METADATA backref */
1091                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1092                                                NULL, info_level + 1, bytenr,
1093                                                1, NULL, GFP_NOFS);
1094                         break;
1095                 case BTRFS_EXTENT_DATA_REF_KEY: {
1096                         /* NORMAL INDIRECT DATA backref */
1097                         struct btrfs_extent_data_ref *dref;
1098                         int count;
1099                         u64 root;
1100
1101                         dref = btrfs_item_ptr(leaf, slot,
1102                                               struct btrfs_extent_data_ref);
1103                         count = btrfs_extent_data_ref_count(leaf, dref);
1104                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1105                                                                       dref);
1106                         key.type = BTRFS_EXTENT_DATA_KEY;
1107                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1108
1109                         if (sc && sc->inum && key.objectid != sc->inum) {
1110                                 ret = BACKREF_FOUND_SHARED;
1111                                 break;
1112                         }
1113
1114                         root = btrfs_extent_data_ref_root(leaf, dref);
1115                         ret = add_indirect_ref(fs_info, preftrees, root,
1116                                                &key, 0, bytenr, count,
1117                                                sc, GFP_NOFS);
1118                         break;
1119                 }
1120                 default:
1121                         WARN_ON(1);
1122                 }
1123                 if (ret)
1124                         return ret;
1125
1126         }
1127
1128         return ret;
1129 }
1130
1131 /*
1132  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1133  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1134  * indirect refs to their parent bytenr.
1135  * When roots are found, they're added to the roots list
1136  *
1137  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1138  * much like trans == NULL case, the difference only lies in it will not
1139  * commit root.
1140  * The special case is for qgroup to search roots in commit_transaction().
1141  *
1142  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1143  * shared extent is detected.
1144  *
1145  * Otherwise this returns 0 for success and <0 for an error.
1146  *
1147  * If ignore_offset is set to false, only extent refs whose offsets match
1148  * extent_item_pos are returned.  If true, every extent ref is returned
1149  * and extent_item_pos is ignored.
1150  *
1151  * FIXME some caching might speed things up
1152  */
1153 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1154                              struct btrfs_fs_info *fs_info, u64 bytenr,
1155                              u64 time_seq, struct ulist *refs,
1156                              struct ulist *roots, const u64 *extent_item_pos,
1157                              struct share_check *sc, bool ignore_offset)
1158 {
1159         struct btrfs_key key;
1160         struct btrfs_path *path;
1161         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1162         struct btrfs_delayed_ref_head *head;
1163         int info_level = 0;
1164         int ret;
1165         struct prelim_ref *ref;
1166         struct rb_node *node;
1167         struct extent_inode_elem *eie = NULL;
1168         struct preftrees preftrees = {
1169                 .direct = PREFTREE_INIT,
1170                 .indirect = PREFTREE_INIT,
1171                 .indirect_missing_keys = PREFTREE_INIT
1172         };
1173
1174         key.objectid = bytenr;
1175         key.offset = (u64)-1;
1176         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1177                 key.type = BTRFS_METADATA_ITEM_KEY;
1178         else
1179                 key.type = BTRFS_EXTENT_ITEM_KEY;
1180
1181         path = btrfs_alloc_path();
1182         if (!path)
1183                 return -ENOMEM;
1184         if (!trans) {
1185                 path->search_commit_root = 1;
1186                 path->skip_locking = 1;
1187         }
1188
1189         if (time_seq == SEQ_LAST)
1190                 path->skip_locking = 1;
1191
1192         /*
1193          * grab both a lock on the path and a lock on the delayed ref head.
1194          * We need both to get a consistent picture of how the refs look
1195          * at a specified point in time
1196          */
1197 again:
1198         head = NULL;
1199
1200         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1201         if (ret < 0)
1202                 goto out;
1203         BUG_ON(ret == 0);
1204
1205 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1206         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1207             time_seq != SEQ_LAST) {
1208 #else
1209         if (trans && time_seq != SEQ_LAST) {
1210 #endif
1211                 /*
1212                  * look if there are updates for this ref queued and lock the
1213                  * head
1214                  */
1215                 delayed_refs = &trans->transaction->delayed_refs;
1216                 spin_lock(&delayed_refs->lock);
1217                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1218                 if (head) {
1219                         if (!mutex_trylock(&head->mutex)) {
1220                                 refcount_inc(&head->refs);
1221                                 spin_unlock(&delayed_refs->lock);
1222
1223                                 btrfs_release_path(path);
1224
1225                                 /*
1226                                  * Mutex was contended, block until it's
1227                                  * released and try again
1228                                  */
1229                                 mutex_lock(&head->mutex);
1230                                 mutex_unlock(&head->mutex);
1231                                 btrfs_put_delayed_ref_head(head);
1232                                 goto again;
1233                         }
1234                         spin_unlock(&delayed_refs->lock);
1235                         ret = add_delayed_refs(fs_info, head, time_seq,
1236                                                &preftrees, sc);
1237                         mutex_unlock(&head->mutex);
1238                         if (ret)
1239                                 goto out;
1240                 } else {
1241                         spin_unlock(&delayed_refs->lock);
1242                 }
1243         }
1244
1245         if (path->slots[0]) {
1246                 struct extent_buffer *leaf;
1247                 int slot;
1248
1249                 path->slots[0]--;
1250                 leaf = path->nodes[0];
1251                 slot = path->slots[0];
1252                 btrfs_item_key_to_cpu(leaf, &key, slot);
1253                 if (key.objectid == bytenr &&
1254                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1255                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1256                         ret = add_inline_refs(fs_info, path, bytenr,
1257                                               &info_level, &preftrees, sc);
1258                         if (ret)
1259                                 goto out;
1260                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1261                                              &preftrees, sc);
1262                         if (ret)
1263                                 goto out;
1264                 }
1265         }
1266
1267         btrfs_release_path(path);
1268
1269         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1270         if (ret)
1271                 goto out;
1272
1273         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1274
1275         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1276                                     extent_item_pos, sc, ignore_offset);
1277         if (ret)
1278                 goto out;
1279
1280         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1281
1282         /*
1283          * This walks the tree of merged and resolved refs. Tree blocks are
1284          * read in as needed. Unique entries are added to the ulist, and
1285          * the list of found roots is updated.
1286          *
1287          * We release the entire tree in one go before returning.
1288          */
1289         node = rb_first_cached(&preftrees.direct.root);
1290         while (node) {
1291                 ref = rb_entry(node, struct prelim_ref, rbnode);
1292                 node = rb_next(&ref->rbnode);
1293                 /*
1294                  * ref->count < 0 can happen here if there are delayed
1295                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1296                  * prelim_ref_insert() relies on this when merging
1297                  * identical refs to keep the overall count correct.
1298                  * prelim_ref_insert() will merge only those refs
1299                  * which compare identically.  Any refs having
1300                  * e.g. different offsets would not be merged,
1301                  * and would retain their original ref->count < 0.
1302                  */
1303                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1304                         if (sc && sc->root_objectid &&
1305                             ref->root_id != sc->root_objectid) {
1306                                 ret = BACKREF_FOUND_SHARED;
1307                                 goto out;
1308                         }
1309
1310                         /* no parent == root of tree */
1311                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1312                         if (ret < 0)
1313                                 goto out;
1314                 }
1315                 if (ref->count && ref->parent) {
1316                         if (extent_item_pos && !ref->inode_list &&
1317                             ref->level == 0) {
1318                                 struct extent_buffer *eb;
1319
1320                                 eb = read_tree_block(fs_info, ref->parent, 0,
1321                                                      ref->level, NULL);
1322                                 if (IS_ERR(eb)) {
1323                                         ret = PTR_ERR(eb);
1324                                         goto out;
1325                                 } else if (!extent_buffer_uptodate(eb)) {
1326                                         free_extent_buffer(eb);
1327                                         ret = -EIO;
1328                                         goto out;
1329                                 }
1330
1331                                 if (!path->skip_locking) {
1332                                         btrfs_tree_read_lock(eb);
1333                                         btrfs_set_lock_blocking_read(eb);
1334                                 }
1335                                 ret = find_extent_in_eb(eb, bytenr,
1336                                                         *extent_item_pos, &eie, ignore_offset);
1337                                 if (!path->skip_locking)
1338                                         btrfs_tree_read_unlock_blocking(eb);
1339                                 free_extent_buffer(eb);
1340                                 if (ret < 0)
1341                                         goto out;
1342                                 ref->inode_list = eie;
1343                         }
1344                         ret = ulist_add_merge_ptr(refs, ref->parent,
1345                                                   ref->inode_list,
1346                                                   (void **)&eie, GFP_NOFS);
1347                         if (ret < 0)
1348                                 goto out;
1349                         if (!ret && extent_item_pos) {
1350                                 /*
1351                                  * we've recorded that parent, so we must extend
1352                                  * its inode list here
1353                                  */
1354                                 BUG_ON(!eie);
1355                                 while (eie->next)
1356                                         eie = eie->next;
1357                                 eie->next = ref->inode_list;
1358                         }
1359                         eie = NULL;
1360                 }
1361                 cond_resched();
1362         }
1363
1364 out:
1365         btrfs_free_path(path);
1366
1367         prelim_release(&preftrees.direct);
1368         prelim_release(&preftrees.indirect);
1369         prelim_release(&preftrees.indirect_missing_keys);
1370
1371         if (ret < 0)
1372                 free_inode_elem_list(eie);
1373         return ret;
1374 }
1375
1376 static void free_leaf_list(struct ulist *blocks)
1377 {
1378         struct ulist_node *node = NULL;
1379         struct extent_inode_elem *eie;
1380         struct ulist_iterator uiter;
1381
1382         ULIST_ITER_INIT(&uiter);
1383         while ((node = ulist_next(blocks, &uiter))) {
1384                 if (!node->aux)
1385                         continue;
1386                 eie = unode_aux_to_inode_list(node);
1387                 free_inode_elem_list(eie);
1388                 node->aux = 0;
1389         }
1390
1391         ulist_free(blocks);
1392 }
1393
1394 /*
1395  * Finds all leafs with a reference to the specified combination of bytenr and
1396  * offset. key_list_head will point to a list of corresponding keys (caller must
1397  * free each list element). The leafs will be stored in the leafs ulist, which
1398  * must be freed with ulist_free.
1399  *
1400  * returns 0 on success, <0 on error
1401  */
1402 int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1403                          struct btrfs_fs_info *fs_info, u64 bytenr,
1404                          u64 time_seq, struct ulist **leafs,
1405                          const u64 *extent_item_pos, bool ignore_offset)
1406 {
1407         int ret;
1408
1409         *leafs = ulist_alloc(GFP_NOFS);
1410         if (!*leafs)
1411                 return -ENOMEM;
1412
1413         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1414                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1415         if (ret < 0 && ret != -ENOENT) {
1416                 free_leaf_list(*leafs);
1417                 return ret;
1418         }
1419
1420         return 0;
1421 }
1422
1423 /*
1424  * walk all backrefs for a given extent to find all roots that reference this
1425  * extent. Walking a backref means finding all extents that reference this
1426  * extent and in turn walk the backrefs of those, too. Naturally this is a
1427  * recursive process, but here it is implemented in an iterative fashion: We
1428  * find all referencing extents for the extent in question and put them on a
1429  * list. In turn, we find all referencing extents for those, further appending
1430  * to the list. The way we iterate the list allows adding more elements after
1431  * the current while iterating. The process stops when we reach the end of the
1432  * list. Found roots are added to the roots list.
1433  *
1434  * returns 0 on success, < 0 on error.
1435  */
1436 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1437                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1438                                      u64 time_seq, struct ulist **roots,
1439                                      bool ignore_offset)
1440 {
1441         struct ulist *tmp;
1442         struct ulist_node *node = NULL;
1443         struct ulist_iterator uiter;
1444         int ret;
1445
1446         tmp = ulist_alloc(GFP_NOFS);
1447         if (!tmp)
1448                 return -ENOMEM;
1449         *roots = ulist_alloc(GFP_NOFS);
1450         if (!*roots) {
1451                 ulist_free(tmp);
1452                 return -ENOMEM;
1453         }
1454
1455         ULIST_ITER_INIT(&uiter);
1456         while (1) {
1457                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1458                                         tmp, *roots, NULL, NULL, ignore_offset);
1459                 if (ret < 0 && ret != -ENOENT) {
1460                         ulist_free(tmp);
1461                         ulist_free(*roots);
1462                         return ret;
1463                 }
1464                 node = ulist_next(tmp, &uiter);
1465                 if (!node)
1466                         break;
1467                 bytenr = node->val;
1468                 cond_resched();
1469         }
1470
1471         ulist_free(tmp);
1472         return 0;
1473 }
1474
1475 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1476                          struct btrfs_fs_info *fs_info, u64 bytenr,
1477                          u64 time_seq, struct ulist **roots,
1478                          bool ignore_offset)
1479 {
1480         int ret;
1481
1482         if (!trans)
1483                 down_read(&fs_info->commit_root_sem);
1484         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1485                                         time_seq, roots, ignore_offset);
1486         if (!trans)
1487                 up_read(&fs_info->commit_root_sem);
1488         return ret;
1489 }
1490
1491 /**
1492  * btrfs_check_shared - tell us whether an extent is shared
1493  *
1494  * btrfs_check_shared uses the backref walking code but will short
1495  * circuit as soon as it finds a root or inode that doesn't match the
1496  * one passed in. This provides a significant performance benefit for
1497  * callers (such as fiemap) which want to know whether the extent is
1498  * shared but do not need a ref count.
1499  *
1500  * This attempts to attach to the running transaction in order to account for
1501  * delayed refs, but continues on even when no running transaction exists.
1502  *
1503  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1504  */
1505 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr,
1506                 struct ulist *roots, struct ulist *tmp)
1507 {
1508         struct btrfs_fs_info *fs_info = root->fs_info;
1509         struct btrfs_trans_handle *trans;
1510         struct ulist_iterator uiter;
1511         struct ulist_node *node;
1512         struct seq_list elem = SEQ_LIST_INIT(elem);
1513         int ret = 0;
1514         struct share_check shared = {
1515                 .root_objectid = root->root_key.objectid,
1516                 .inum = inum,
1517                 .share_count = 0,
1518         };
1519
1520         ulist_init(roots);
1521         ulist_init(tmp);
1522
1523         trans = btrfs_join_transaction_nostart(root);
1524         if (IS_ERR(trans)) {
1525                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1526                         ret = PTR_ERR(trans);
1527                         goto out;
1528                 }
1529                 trans = NULL;
1530                 down_read(&fs_info->commit_root_sem);
1531         } else {
1532                 btrfs_get_tree_mod_seq(fs_info, &elem);
1533         }
1534
1535         ULIST_ITER_INIT(&uiter);
1536         while (1) {
1537                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1538                                         roots, NULL, &shared, false);
1539                 if (ret == BACKREF_FOUND_SHARED) {
1540                         /* this is the only condition under which we return 1 */
1541                         ret = 1;
1542                         break;
1543                 }
1544                 if (ret < 0 && ret != -ENOENT)
1545                         break;
1546                 ret = 0;
1547                 node = ulist_next(tmp, &uiter);
1548                 if (!node)
1549                         break;
1550                 bytenr = node->val;
1551                 shared.share_count = 0;
1552                 cond_resched();
1553         }
1554
1555         if (trans) {
1556                 btrfs_put_tree_mod_seq(fs_info, &elem);
1557                 btrfs_end_transaction(trans);
1558         } else {
1559                 up_read(&fs_info->commit_root_sem);
1560         }
1561 out:
1562         ulist_release(roots);
1563         ulist_release(tmp);
1564         return ret;
1565 }
1566
1567 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1568                           u64 start_off, struct btrfs_path *path,
1569                           struct btrfs_inode_extref **ret_extref,
1570                           u64 *found_off)
1571 {
1572         int ret, slot;
1573         struct btrfs_key key;
1574         struct btrfs_key found_key;
1575         struct btrfs_inode_extref *extref;
1576         const struct extent_buffer *leaf;
1577         unsigned long ptr;
1578
1579         key.objectid = inode_objectid;
1580         key.type = BTRFS_INODE_EXTREF_KEY;
1581         key.offset = start_off;
1582
1583         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1584         if (ret < 0)
1585                 return ret;
1586
1587         while (1) {
1588                 leaf = path->nodes[0];
1589                 slot = path->slots[0];
1590                 if (slot >= btrfs_header_nritems(leaf)) {
1591                         /*
1592                          * If the item at offset is not found,
1593                          * btrfs_search_slot will point us to the slot
1594                          * where it should be inserted. In our case
1595                          * that will be the slot directly before the
1596                          * next INODE_REF_KEY_V2 item. In the case
1597                          * that we're pointing to the last slot in a
1598                          * leaf, we must move one leaf over.
1599                          */
1600                         ret = btrfs_next_leaf(root, path);
1601                         if (ret) {
1602                                 if (ret >= 1)
1603                                         ret = -ENOENT;
1604                                 break;
1605                         }
1606                         continue;
1607                 }
1608
1609                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1610
1611                 /*
1612                  * Check that we're still looking at an extended ref key for
1613                  * this particular objectid. If we have different
1614                  * objectid or type then there are no more to be found
1615                  * in the tree and we can exit.
1616                  */
1617                 ret = -ENOENT;
1618                 if (found_key.objectid != inode_objectid)
1619                         break;
1620                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1621                         break;
1622
1623                 ret = 0;
1624                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1625                 extref = (struct btrfs_inode_extref *)ptr;
1626                 *ret_extref = extref;
1627                 if (found_off)
1628                         *found_off = found_key.offset;
1629                 break;
1630         }
1631
1632         return ret;
1633 }
1634
1635 /*
1636  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1637  * Elements of the path are separated by '/' and the path is guaranteed to be
1638  * 0-terminated. the path is only given within the current file system.
1639  * Therefore, it never starts with a '/'. the caller is responsible to provide
1640  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1641  * the start point of the resulting string is returned. this pointer is within
1642  * dest, normally.
1643  * in case the path buffer would overflow, the pointer is decremented further
1644  * as if output was written to the buffer, though no more output is actually
1645  * generated. that way, the caller can determine how much space would be
1646  * required for the path to fit into the buffer. in that case, the returned
1647  * value will be smaller than dest. callers must check this!
1648  */
1649 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1650                         u32 name_len, unsigned long name_off,
1651                         struct extent_buffer *eb_in, u64 parent,
1652                         char *dest, u32 size)
1653 {
1654         int slot;
1655         u64 next_inum;
1656         int ret;
1657         s64 bytes_left = ((s64)size) - 1;
1658         struct extent_buffer *eb = eb_in;
1659         struct btrfs_key found_key;
1660         int leave_spinning = path->leave_spinning;
1661         struct btrfs_inode_ref *iref;
1662
1663         if (bytes_left >= 0)
1664                 dest[bytes_left] = '\0';
1665
1666         path->leave_spinning = 1;
1667         while (1) {
1668                 bytes_left -= name_len;
1669                 if (bytes_left >= 0)
1670                         read_extent_buffer(eb, dest + bytes_left,
1671                                            name_off, name_len);
1672                 if (eb != eb_in) {
1673                         if (!path->skip_locking)
1674                                 btrfs_tree_read_unlock_blocking(eb);
1675                         free_extent_buffer(eb);
1676                 }
1677                 ret = btrfs_find_item(fs_root, path, parent, 0,
1678                                 BTRFS_INODE_REF_KEY, &found_key);
1679                 if (ret > 0)
1680                         ret = -ENOENT;
1681                 if (ret)
1682                         break;
1683
1684                 next_inum = found_key.offset;
1685
1686                 /* regular exit ahead */
1687                 if (parent == next_inum)
1688                         break;
1689
1690                 slot = path->slots[0];
1691                 eb = path->nodes[0];
1692                 /* make sure we can use eb after releasing the path */
1693                 if (eb != eb_in) {
1694                         if (!path->skip_locking)
1695                                 btrfs_set_lock_blocking_read(eb);
1696                         path->nodes[0] = NULL;
1697                         path->locks[0] = 0;
1698                 }
1699                 btrfs_release_path(path);
1700                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1701
1702                 name_len = btrfs_inode_ref_name_len(eb, iref);
1703                 name_off = (unsigned long)(iref + 1);
1704
1705                 parent = next_inum;
1706                 --bytes_left;
1707                 if (bytes_left >= 0)
1708                         dest[bytes_left] = '/';
1709         }
1710
1711         btrfs_release_path(path);
1712         path->leave_spinning = leave_spinning;
1713
1714         if (ret)
1715                 return ERR_PTR(ret);
1716
1717         return dest + bytes_left;
1718 }
1719
1720 /*
1721  * this makes the path point to (logical EXTENT_ITEM *)
1722  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1723  * tree blocks and <0 on error.
1724  */
1725 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1726                         struct btrfs_path *path, struct btrfs_key *found_key,
1727                         u64 *flags_ret)
1728 {
1729         int ret;
1730         u64 flags;
1731         u64 size = 0;
1732         u32 item_size;
1733         const struct extent_buffer *eb;
1734         struct btrfs_extent_item *ei;
1735         struct btrfs_key key;
1736
1737         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1738                 key.type = BTRFS_METADATA_ITEM_KEY;
1739         else
1740                 key.type = BTRFS_EXTENT_ITEM_KEY;
1741         key.objectid = logical;
1742         key.offset = (u64)-1;
1743
1744         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1745         if (ret < 0)
1746                 return ret;
1747
1748         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1749         if (ret) {
1750                 if (ret > 0)
1751                         ret = -ENOENT;
1752                 return ret;
1753         }
1754         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1755         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1756                 size = fs_info->nodesize;
1757         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1758                 size = found_key->offset;
1759
1760         if (found_key->objectid > logical ||
1761             found_key->objectid + size <= logical) {
1762                 btrfs_debug(fs_info,
1763                         "logical %llu is not within any extent", logical);
1764                 return -ENOENT;
1765         }
1766
1767         eb = path->nodes[0];
1768         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1769         BUG_ON(item_size < sizeof(*ei));
1770
1771         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1772         flags = btrfs_extent_flags(eb, ei);
1773
1774         btrfs_debug(fs_info,
1775                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1776                  logical, logical - found_key->objectid, found_key->objectid,
1777                  found_key->offset, flags, item_size);
1778
1779         WARN_ON(!flags_ret);
1780         if (flags_ret) {
1781                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1782                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1783                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1784                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1785                 else
1786                         BUG();
1787                 return 0;
1788         }
1789
1790         return -EIO;
1791 }
1792
1793 /*
1794  * helper function to iterate extent inline refs. ptr must point to a 0 value
1795  * for the first call and may be modified. it is used to track state.
1796  * if more refs exist, 0 is returned and the next call to
1797  * get_extent_inline_ref must pass the modified ptr parameter to get the
1798  * next ref. after the last ref was processed, 1 is returned.
1799  * returns <0 on error
1800  */
1801 static int get_extent_inline_ref(unsigned long *ptr,
1802                                  const struct extent_buffer *eb,
1803                                  const struct btrfs_key *key,
1804                                  const struct btrfs_extent_item *ei,
1805                                  u32 item_size,
1806                                  struct btrfs_extent_inline_ref **out_eiref,
1807                                  int *out_type)
1808 {
1809         unsigned long end;
1810         u64 flags;
1811         struct btrfs_tree_block_info *info;
1812
1813         if (!*ptr) {
1814                 /* first call */
1815                 flags = btrfs_extent_flags(eb, ei);
1816                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1817                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1818                                 /* a skinny metadata extent */
1819                                 *out_eiref =
1820                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1821                         } else {
1822                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1823                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1824                                 *out_eiref =
1825                                    (struct btrfs_extent_inline_ref *)(info + 1);
1826                         }
1827                 } else {
1828                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1829                 }
1830                 *ptr = (unsigned long)*out_eiref;
1831                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1832                         return -ENOENT;
1833         }
1834
1835         end = (unsigned long)ei + item_size;
1836         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1837         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1838                                                      BTRFS_REF_TYPE_ANY);
1839         if (*out_type == BTRFS_REF_TYPE_INVALID)
1840                 return -EUCLEAN;
1841
1842         *ptr += btrfs_extent_inline_ref_size(*out_type);
1843         WARN_ON(*ptr > end);
1844         if (*ptr == end)
1845                 return 1; /* last */
1846
1847         return 0;
1848 }
1849
1850 /*
1851  * reads the tree block backref for an extent. tree level and root are returned
1852  * through out_level and out_root. ptr must point to a 0 value for the first
1853  * call and may be modified (see get_extent_inline_ref comment).
1854  * returns 0 if data was provided, 1 if there was no more data to provide or
1855  * <0 on error.
1856  */
1857 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1858                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1859                             u32 item_size, u64 *out_root, u8 *out_level)
1860 {
1861         int ret;
1862         int type;
1863         struct btrfs_extent_inline_ref *eiref;
1864
1865         if (*ptr == (unsigned long)-1)
1866                 return 1;
1867
1868         while (1) {
1869                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1870                                               &eiref, &type);
1871                 if (ret < 0)
1872                         return ret;
1873
1874                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1875                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1876                         break;
1877
1878                 if (ret == 1)
1879                         return 1;
1880         }
1881
1882         /* we can treat both ref types equally here */
1883         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1884
1885         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1886                 struct btrfs_tree_block_info *info;
1887
1888                 info = (struct btrfs_tree_block_info *)(ei + 1);
1889                 *out_level = btrfs_tree_block_level(eb, info);
1890         } else {
1891                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1892                 *out_level = (u8)key->offset;
1893         }
1894
1895         if (ret == 1)
1896                 *ptr = (unsigned long)-1;
1897
1898         return 0;
1899 }
1900
1901 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1902                              struct extent_inode_elem *inode_list,
1903                              u64 root, u64 extent_item_objectid,
1904                              iterate_extent_inodes_t *iterate, void *ctx)
1905 {
1906         struct extent_inode_elem *eie;
1907         int ret = 0;
1908
1909         for (eie = inode_list; eie; eie = eie->next) {
1910                 btrfs_debug(fs_info,
1911                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1912                             extent_item_objectid, eie->inum,
1913                             eie->offset, root);
1914                 ret = iterate(eie->inum, eie->offset, root, ctx);
1915                 if (ret) {
1916                         btrfs_debug(fs_info,
1917                                     "stopping iteration for %llu due to ret=%d",
1918                                     extent_item_objectid, ret);
1919                         break;
1920                 }
1921         }
1922
1923         return ret;
1924 }
1925
1926 /*
1927  * calls iterate() for every inode that references the extent identified by
1928  * the given parameters.
1929  * when the iterator function returns a non-zero value, iteration stops.
1930  */
1931 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1932                                 u64 extent_item_objectid, u64 extent_item_pos,
1933                                 int search_commit_root,
1934                                 iterate_extent_inodes_t *iterate, void *ctx,
1935                                 bool ignore_offset)
1936 {
1937         int ret;
1938         struct btrfs_trans_handle *trans = NULL;
1939         struct ulist *refs = NULL;
1940         struct ulist *roots = NULL;
1941         struct ulist_node *ref_node = NULL;
1942         struct ulist_node *root_node = NULL;
1943         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1944         struct ulist_iterator ref_uiter;
1945         struct ulist_iterator root_uiter;
1946
1947         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1948                         extent_item_objectid);
1949
1950         if (!search_commit_root) {
1951                 trans = btrfs_attach_transaction(fs_info->extent_root);
1952                 if (IS_ERR(trans)) {
1953                         if (PTR_ERR(trans) != -ENOENT &&
1954                             PTR_ERR(trans) != -EROFS)
1955                                 return PTR_ERR(trans);
1956                         trans = NULL;
1957                 }
1958         }
1959
1960         if (trans)
1961                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1962         else
1963                 down_read(&fs_info->commit_root_sem);
1964
1965         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1966                                    tree_mod_seq_elem.seq, &refs,
1967                                    &extent_item_pos, ignore_offset);
1968         if (ret)
1969                 goto out;
1970
1971         ULIST_ITER_INIT(&ref_uiter);
1972         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1973                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1974                                                 tree_mod_seq_elem.seq, &roots,
1975                                                 ignore_offset);
1976                 if (ret)
1977                         break;
1978                 ULIST_ITER_INIT(&root_uiter);
1979                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1980                         btrfs_debug(fs_info,
1981                                     "root %llu references leaf %llu, data list %#llx",
1982                                     root_node->val, ref_node->val,
1983                                     ref_node->aux);
1984                         ret = iterate_leaf_refs(fs_info,
1985                                                 (struct extent_inode_elem *)
1986                                                 (uintptr_t)ref_node->aux,
1987                                                 root_node->val,
1988                                                 extent_item_objectid,
1989                                                 iterate, ctx);
1990                 }
1991                 ulist_free(roots);
1992         }
1993
1994         free_leaf_list(refs);
1995 out:
1996         if (trans) {
1997                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1998                 btrfs_end_transaction(trans);
1999         } else {
2000                 up_read(&fs_info->commit_root_sem);
2001         }
2002
2003         return ret;
2004 }
2005
2006 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2007                                 struct btrfs_path *path,
2008                                 iterate_extent_inodes_t *iterate, void *ctx,
2009                                 bool ignore_offset)
2010 {
2011         int ret;
2012         u64 extent_item_pos;
2013         u64 flags = 0;
2014         struct btrfs_key found_key;
2015         int search_commit_root = path->search_commit_root;
2016
2017         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2018         btrfs_release_path(path);
2019         if (ret < 0)
2020                 return ret;
2021         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2022                 return -EINVAL;
2023
2024         extent_item_pos = logical - found_key.objectid;
2025         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2026                                         extent_item_pos, search_commit_root,
2027                                         iterate, ctx, ignore_offset);
2028
2029         return ret;
2030 }
2031
2032 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2033                               struct extent_buffer *eb, void *ctx);
2034
2035 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2036                               struct btrfs_path *path,
2037                               iterate_irefs_t *iterate, void *ctx)
2038 {
2039         int ret = 0;
2040         int slot;
2041         u32 cur;
2042         u32 len;
2043         u32 name_len;
2044         u64 parent = 0;
2045         int found = 0;
2046         struct extent_buffer *eb;
2047         struct btrfs_item *item;
2048         struct btrfs_inode_ref *iref;
2049         struct btrfs_key found_key;
2050
2051         while (!ret) {
2052                 ret = btrfs_find_item(fs_root, path, inum,
2053                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2054                                 &found_key);
2055
2056                 if (ret < 0)
2057                         break;
2058                 if (ret) {
2059                         ret = found ? 0 : -ENOENT;
2060                         break;
2061                 }
2062                 ++found;
2063
2064                 parent = found_key.offset;
2065                 slot = path->slots[0];
2066                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2067                 if (!eb) {
2068                         ret = -ENOMEM;
2069                         break;
2070                 }
2071                 btrfs_release_path(path);
2072
2073                 item = btrfs_item_nr(slot);
2074                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2075
2076                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2077                         name_len = btrfs_inode_ref_name_len(eb, iref);
2078                         /* path must be released before calling iterate()! */
2079                         btrfs_debug(fs_root->fs_info,
2080                                 "following ref at offset %u for inode %llu in tree %llu",
2081                                 cur, found_key.objectid,
2082                                 fs_root->root_key.objectid);
2083                         ret = iterate(parent, name_len,
2084                                       (unsigned long)(iref + 1), eb, ctx);
2085                         if (ret)
2086                                 break;
2087                         len = sizeof(*iref) + name_len;
2088                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2089                 }
2090                 free_extent_buffer(eb);
2091         }
2092
2093         btrfs_release_path(path);
2094
2095         return ret;
2096 }
2097
2098 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2099                                  struct btrfs_path *path,
2100                                  iterate_irefs_t *iterate, void *ctx)
2101 {
2102         int ret;
2103         int slot;
2104         u64 offset = 0;
2105         u64 parent;
2106         int found = 0;
2107         struct extent_buffer *eb;
2108         struct btrfs_inode_extref *extref;
2109         u32 item_size;
2110         u32 cur_offset;
2111         unsigned long ptr;
2112
2113         while (1) {
2114                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2115                                             &offset);
2116                 if (ret < 0)
2117                         break;
2118                 if (ret) {
2119                         ret = found ? 0 : -ENOENT;
2120                         break;
2121                 }
2122                 ++found;
2123
2124                 slot = path->slots[0];
2125                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2126                 if (!eb) {
2127                         ret = -ENOMEM;
2128                         break;
2129                 }
2130                 btrfs_release_path(path);
2131
2132                 item_size = btrfs_item_size_nr(eb, slot);
2133                 ptr = btrfs_item_ptr_offset(eb, slot);
2134                 cur_offset = 0;
2135
2136                 while (cur_offset < item_size) {
2137                         u32 name_len;
2138
2139                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2140                         parent = btrfs_inode_extref_parent(eb, extref);
2141                         name_len = btrfs_inode_extref_name_len(eb, extref);
2142                         ret = iterate(parent, name_len,
2143                                       (unsigned long)&extref->name, eb, ctx);
2144                         if (ret)
2145                                 break;
2146
2147                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2148                         cur_offset += sizeof(*extref);
2149                 }
2150                 free_extent_buffer(eb);
2151
2152                 offset++;
2153         }
2154
2155         btrfs_release_path(path);
2156
2157         return ret;
2158 }
2159
2160 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2161                          struct btrfs_path *path, iterate_irefs_t *iterate,
2162                          void *ctx)
2163 {
2164         int ret;
2165         int found_refs = 0;
2166
2167         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2168         if (!ret)
2169                 ++found_refs;
2170         else if (ret != -ENOENT)
2171                 return ret;
2172
2173         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2174         if (ret == -ENOENT && found_refs)
2175                 return 0;
2176
2177         return ret;
2178 }
2179
2180 /*
2181  * returns 0 if the path could be dumped (probably truncated)
2182  * returns <0 in case of an error
2183  */
2184 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2185                          struct extent_buffer *eb, void *ctx)
2186 {
2187         struct inode_fs_paths *ipath = ctx;
2188         char *fspath;
2189         char *fspath_min;
2190         int i = ipath->fspath->elem_cnt;
2191         const int s_ptr = sizeof(char *);
2192         u32 bytes_left;
2193
2194         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2195                                         ipath->fspath->bytes_left - s_ptr : 0;
2196
2197         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2198         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2199                                    name_off, eb, inum, fspath_min, bytes_left);
2200         if (IS_ERR(fspath))
2201                 return PTR_ERR(fspath);
2202
2203         if (fspath > fspath_min) {
2204                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2205                 ++ipath->fspath->elem_cnt;
2206                 ipath->fspath->bytes_left = fspath - fspath_min;
2207         } else {
2208                 ++ipath->fspath->elem_missed;
2209                 ipath->fspath->bytes_missing += fspath_min - fspath;
2210                 ipath->fspath->bytes_left = 0;
2211         }
2212
2213         return 0;
2214 }
2215
2216 /*
2217  * this dumps all file system paths to the inode into the ipath struct, provided
2218  * is has been created large enough. each path is zero-terminated and accessed
2219  * from ipath->fspath->val[i].
2220  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2221  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2222  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2223  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2224  * have been needed to return all paths.
2225  */
2226 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2227 {
2228         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2229                              inode_to_path, ipath);
2230 }
2231
2232 struct btrfs_data_container *init_data_container(u32 total_bytes)
2233 {
2234         struct btrfs_data_container *data;
2235         size_t alloc_bytes;
2236
2237         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2238         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2239         if (!data)
2240                 return ERR_PTR(-ENOMEM);
2241
2242         if (total_bytes >= sizeof(*data)) {
2243                 data->bytes_left = total_bytes - sizeof(*data);
2244                 data->bytes_missing = 0;
2245         } else {
2246                 data->bytes_missing = sizeof(*data) - total_bytes;
2247                 data->bytes_left = 0;
2248         }
2249
2250         data->elem_cnt = 0;
2251         data->elem_missed = 0;
2252
2253         return data;
2254 }
2255
2256 /*
2257  * allocates space to return multiple file system paths for an inode.
2258  * total_bytes to allocate are passed, note that space usable for actual path
2259  * information will be total_bytes - sizeof(struct inode_fs_paths).
2260  * the returned pointer must be freed with free_ipath() in the end.
2261  */
2262 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2263                                         struct btrfs_path *path)
2264 {
2265         struct inode_fs_paths *ifp;
2266         struct btrfs_data_container *fspath;
2267
2268         fspath = init_data_container(total_bytes);
2269         if (IS_ERR(fspath))
2270                 return ERR_CAST(fspath);
2271
2272         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2273         if (!ifp) {
2274                 kvfree(fspath);
2275                 return ERR_PTR(-ENOMEM);
2276         }
2277
2278         ifp->btrfs_path = path;
2279         ifp->fspath = fspath;
2280         ifp->fs_root = fs_root;
2281
2282         return ifp;
2283 }
2284
2285 void free_ipath(struct inode_fs_paths *ipath)
2286 {
2287         if (!ipath)
2288                 return;
2289         kvfree(ipath->fspath);
2290         kfree(ipath);
2291 }