backref.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2011 STRATO.  All rights reserved.
   4 */
   5
   6#include <linux/mm.h>
   7#include <linux/rbtree.h>
   8#include <trace/events/btrfs.h>
   9#include "ctree.h"
  10#include "disk-io.h"
  11#include "backref.h"
  12#include "ulist.h"
  13#include "transaction.h"
  14#include "delayed-ref.h"
  15#include "locking.h"
  16#include "misc.h"
  17#include "tree-mod-log.h"
  18#include "fs.h"
  19#include "accessors.h"
  20#include "extent-tree.h"
  21#include "relocation.h"
  22#include "tree-checker.h"
  23
  24/* Just arbitrary numbers so we can be sure one of these happened. */
  25#define BACKREF_FOUND_SHARED     6
  26#define BACKREF_FOUND_NOT_SHARED 7
  27
  28struct extent_inode_elem {
  29	u64 inum;
  30	u64 offset;
  31	u64 num_bytes;
  32	struct extent_inode_elem *next;
  33};
  34
  35static int check_extent_in_eb(struct btrfs_backref_walk_ctx *ctx,
  36			      const struct btrfs_key *key,
  37			      const struct extent_buffer *eb,
  38			      const struct btrfs_file_extent_item *fi,
  39			      struct extent_inode_elem **eie)
  40{
  41	const u64 data_len = btrfs_file_extent_num_bytes(eb, fi);
  42	u64 offset = key->offset;
  43	struct extent_inode_elem *e;
  44	const u64 *root_ids;
  45	int root_count;
  46	bool cached;
  47
  48	if (!ctx->ignore_extent_item_pos &&
  49	    !btrfs_file_extent_compression(eb, fi) &&
  50	    !btrfs_file_extent_encryption(eb, fi) &&
  51	    !btrfs_file_extent_other_encoding(eb, fi)) {
  52		u64 data_offset;
  53
  54		data_offset = btrfs_file_extent_offset(eb, fi);
  55
  56		if (ctx->extent_item_pos < data_offset ||
  57		    ctx->extent_item_pos >= data_offset + data_len)
  58			return 1;
  59		offset += ctx->extent_item_pos - data_offset;
  60	}
  61
  62	if (!ctx->indirect_ref_iterator || !ctx->cache_lookup)
  63		goto add_inode_elem;
  64
  65	cached = ctx->cache_lookup(eb->start, ctx->user_ctx, &root_ids,
  66				   &root_count);
  67	if (!cached)
  68		goto add_inode_elem;
  69
  70	for (int i = 0; i < root_count; i++) {
  71		int ret;
  72
  73		ret = ctx->indirect_ref_iterator(key->objectid, offset,
  74						 data_len, root_ids[i],
  75						 ctx->user_ctx);
  76		if (ret)
  77			return ret;
  78	}
  79
  80add_inode_elem:
  81	e = kmalloc(sizeof(*e), GFP_NOFS);
  82	if (!e)
  83		return -ENOMEM;
  84
  85	e->next = *eie;
  86	e->inum = key->objectid;
  87	e->offset = offset;
  88	e->num_bytes = data_len;
  89	*eie = e;
  90
  91	return 0;
  92}
  93
  94static void free_inode_elem_list(struct extent_inode_elem *eie)
  95{
  96	struct extent_inode_elem *eie_next;
  97
  98	for (; eie; eie = eie_next) {
  99		eie_next = eie->next;
 100		kfree(eie);
 101	}
 102}
 103
 104static int find_extent_in_eb(struct btrfs_backref_walk_ctx *ctx,
 105			     const struct extent_buffer *eb,
 106			     struct extent_inode_elem **eie)
 107{
 108	u64 disk_byte;
 109	struct btrfs_key key;
 110	struct btrfs_file_extent_item *fi;
 111	int slot;
 112	int nritems;
 113	int extent_type;
 114	int ret;
 115
 116	/*
 117	 * from the shared data ref, we only have the leaf but we need
 118	 * the key. thus, we must look into all items and see that we
 119	 * find one (some) with a reference to our extent item.
 120	 */
 121	nritems = btrfs_header_nritems(eb);
 122	for (slot = 0; slot < nritems; ++slot) {
 123		btrfs_item_key_to_cpu(eb, &key, slot);
 124		if (key.type != BTRFS_EXTENT_DATA_KEY)
 125			continue;
 126		fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 127		extent_type = btrfs_file_extent_type(eb, fi);
 128		if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 129			continue;
 130		/* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 131		disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 132		if (disk_byte != ctx->bytenr)
 133			continue;
 134
 135		ret = check_extent_in_eb(ctx, &key, eb, fi, eie);
 136		if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
 137			return ret;
 138	}
 139
 140	return 0;
 141}
 142
 143struct preftree {
 144	struct rb_root_cached root;
 145	unsigned int count;
 146};
 147
 148#define PREFTREE_INIT	{ .root = RB_ROOT_CACHED, .count = 0 }
 149
 150struct preftrees {
 151	struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 152	struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 153	struct preftree indirect_missing_keys;
 154};
 155
 156/*
 157 * Checks for a shared extent during backref search.
 158 *
 159 * The share_count tracks prelim_refs (direct and indirect) having a
 160 * ref->count >0:
 161 *  - incremented when a ref->count transitions to >0
 162 *  - decremented when a ref->count transitions to <1
 163 */
 164struct share_check {
 165	struct btrfs_backref_share_check_ctx *ctx;
 166	struct btrfs_root *root;
 167	u64 inum;
 168	u64 data_bytenr;
 169	u64 data_extent_gen;
 170	/*
 171	 * Counts number of inodes that refer to an extent (different inodes in
 172	 * the same root or different roots) that we could find. The sharedness
 173	 * check typically stops once this counter gets greater than 1, so it
 174	 * may not reflect the total number of inodes.
 175	 */
 176	int share_count;
 177	/*
 178	 * The number of times we found our inode refers to the data extent we
 179	 * are determining the sharedness. In other words, how many file extent
 180	 * items we could find for our inode that point to our target data
 181	 * extent. The value we get here after finishing the extent sharedness
 182	 * check may be smaller than reality, but if it ends up being greater
 183	 * than 1, then we know for sure the inode has multiple file extent
 184	 * items that point to our inode, and we can safely assume it's useful
 185	 * to cache the sharedness check result.
 186	 */
 187	int self_ref_count;
 188	bool have_delayed_delete_refs;
 189};
 190
 191static inline int extent_is_shared(struct share_check *sc)
 192{
 193	return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 194}
 195
 196static struct kmem_cache *btrfs_prelim_ref_cache;
 197
 198int __init btrfs_prelim_ref_init(void)
 199{
 200	btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 201					sizeof(struct prelim_ref), 0, 0, NULL);
 202	if (!btrfs_prelim_ref_cache)
 203		return -ENOMEM;
 204	return 0;
 205}
 206
 207void __cold btrfs_prelim_ref_exit(void)
 208{
 209	kmem_cache_destroy(btrfs_prelim_ref_cache);
 210}
 211
 212static void free_pref(struct prelim_ref *ref)
 213{
 214	kmem_cache_free(btrfs_prelim_ref_cache, ref);
 215}
 216
 217/*
 218 * Return 0 when both refs are for the same block (and can be merged).
 219 * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 220 * indicates a 'higher' block.
 221 */
 222static int prelim_ref_compare(const struct prelim_ref *ref1,
 223			      const struct prelim_ref *ref2)
 224{
 225	if (ref1->level < ref2->level)
 226		return -1;
 227	if (ref1->level > ref2->level)
 228		return 1;
 229	if (ref1->root_id < ref2->root_id)
 230		return -1;
 231	if (ref1->root_id > ref2->root_id)
 232		return 1;
 233	if (ref1->key_for_search.type < ref2->key_for_search.type)
 234		return -1;
 235	if (ref1->key_for_search.type > ref2->key_for_search.type)
 236		return 1;
 237	if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 238		return -1;
 239	if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 240		return 1;
 241	if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 242		return -1;
 243	if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 244		return 1;
 245	if (ref1->parent < ref2->parent)
 246		return -1;
 247	if (ref1->parent > ref2->parent)
 248		return 1;
 249
 250	return 0;
 251}
 252
 253static int prelim_ref_rb_add_cmp(const struct rb_node *new,
 254				 const struct rb_node *exist)
 255{
 256	const struct prelim_ref *ref_new =
 257		rb_entry(new, struct prelim_ref, rbnode);
 258	const struct prelim_ref *ref_exist =
 259		rb_entry(exist, struct prelim_ref, rbnode);
 260
 261	/*
 262	 * prelim_ref_compare() expects the first parameter as the existing one,
 263	 * different from the rb_find_add_cached() order.
 264	 */
 265	return prelim_ref_compare(ref_exist, ref_new);
 266}
 267
 268static void update_share_count(struct share_check *sc, int oldcount,
 269			       int newcount, const struct prelim_ref *newref)
 270{
 271	if ((!sc) || (oldcount == 0 && newcount < 1))
 272		return;
 273
 274	if (oldcount > 0 && newcount < 1)
 275		sc->share_count--;
 276	else if (oldcount < 1 && newcount > 0)
 277		sc->share_count++;
 278
 279	if (newref->root_id == btrfs_root_id(sc->root) &&
 280	    newref->wanted_disk_byte == sc->data_bytenr &&
 281	    newref->key_for_search.objectid == sc->inum)
 282		sc->self_ref_count += newref->count;
 283}
 284
 285/*
 286 * Add @newref to the @root rbtree, merging identical refs.
 287 *
 288 * Callers should assume that newref has been freed after calling.
 289 */
 290static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 291			      struct preftree *preftree,
 292			      struct prelim_ref *newref,
 293			      struct share_check *sc)
 294{
 295	struct rb_root_cached *root;
 296	struct rb_node *exist;
 297
 298	root = &preftree->root;
 299	exist = rb_find_add_cached(&newref->rbnode, root, prelim_ref_rb_add_cmp);
 300	if (exist) {
 301		struct prelim_ref *ref = rb_entry(exist, struct prelim_ref, rbnode);
 302		/* Identical refs, merge them and free @newref */
 303		struct extent_inode_elem *eie = ref->inode_list;
 304
 305		while (eie && eie->next)
 306			eie = eie->next;
 307
 308		if (!eie)
 309			ref->inode_list = newref->inode_list;
 310		else
 311			eie->next = newref->inode_list;
 312		trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 313							preftree->count);
 314		/*
 315		 * A delayed ref can have newref->count < 0.
 316		 * The ref->count is updated to follow any
 317		 * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 318		 */
 319		update_share_count(sc, ref->count,
 320					ref->count + newref->count, newref);
 321		ref->count += newref->count;
 322		free_pref(newref);
 323		return;
 324	}
 325
 326	update_share_count(sc, 0, newref->count, newref);
 327	preftree->count++;
 328	trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 329}
 330
 331/*
 332 * Release the entire tree.  We don't care about internal consistency so
 333 * just free everything and then reset the tree root.
 334 */
 335static void prelim_release(struct preftree *preftree)
 336{
 337	struct prelim_ref *ref, *next_ref;
 338
 339	rbtree_postorder_for_each_entry_safe(ref, next_ref,
 340					     &preftree->root.rb_root, rbnode) {
 341		free_inode_elem_list(ref->inode_list);
 342		free_pref(ref);
 343	}
 344
 345	preftree->root = RB_ROOT_CACHED;
 346	preftree->count = 0;
 347}
 348
 349/*
 350 * the rules for all callers of this function are:
 351 * - obtaining the parent is the goal
 352 * - if you add a key, you must know that it is a correct key
 353 * - if you cannot add the parent or a correct key, then we will look into the
 354 *   block later to set a correct key
 355 *
 356 * delayed refs
 357 * ============
 358 *        backref type | shared | indirect | shared | indirect
 359 * information         |   tree |     tree |   data |     data
 360 * --------------------+--------+----------+--------+----------
 361 *      parent logical |    y   |     -    |    -   |     -
 362 *      key to resolve |    -   |     y    |    y   |     y
 363 *  tree block logical |    -   |     -    |    -   |     -
 364 *  root for resolving |    y   |     y    |    y   |     y
 365 *
 366 * - column 1:       we've the parent -> done
 367 * - column 2, 3, 4: we use the key to find the parent
 368 *
 369 * on disk refs (inline or keyed)
 370 * ==============================
 371 *        backref type | shared | indirect | shared | indirect
 372 * information         |   tree |     tree |   data |     data
 373 * --------------------+--------+----------+--------+----------
 374 *      parent logical |    y   |     -    |    y   |     -
 375 *      key to resolve |    -   |     -    |    -   |     y
 376 *  tree block logical |    y   |     y    |    y   |     y
 377 *  root for resolving |    -   |     y    |    y   |     y
 378 *
 379 * - column 1, 3: we've the parent -> done
 380 * - column 2:    we take the first key from the block to find the parent
 381 *                (see add_missing_keys)
 382 * - column 4:    we use the key to find the parent
 383 *
 384 * additional information that's available but not required to find the parent
 385 * block might help in merging entries to gain some speed.
 386 */
 387static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 388			  struct preftree *preftree, u64 root_id,
 389			  const struct btrfs_key *key, int level, u64 parent,
 390			  u64 wanted_disk_byte, int count,
 391			  struct share_check *sc, gfp_t gfp_mask)
 392{
 393	struct prelim_ref *ref;
 394
 395	if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 396		return 0;
 397
 398	ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 399	if (!ref)
 400		return -ENOMEM;
 401
 402	ref->root_id = root_id;
 403	if (key)
 404		ref->key_for_search = *key;
 405	else
 406		memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 407
 408	ref->inode_list = NULL;
 409	ref->level = level;
 410	ref->count = count;
 411	ref->parent = parent;
 412	ref->wanted_disk_byte = wanted_disk_byte;
 413	prelim_ref_insert(fs_info, preftree, ref, sc);
 414	return extent_is_shared(sc);
 415}
 416
 417/* direct refs use root == 0, key == NULL */
 418static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 419			  struct preftrees *preftrees, int level, u64 parent,
 420			  u64 wanted_disk_byte, int count,
 421			  struct share_check *sc, gfp_t gfp_mask)
 422{
 423	return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 424			      parent, wanted_disk_byte, count, sc, gfp_mask);
 425}
 426
 427/* indirect refs use parent == 0 */
 428static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 429			    struct preftrees *preftrees, u64 root_id,
 430			    const struct btrfs_key *key, int level,
 431			    u64 wanted_disk_byte, int count,
 432			    struct share_check *sc, gfp_t gfp_mask)
 433{
 434	struct preftree *tree = &preftrees->indirect;
 435
 436	if (!key)
 437		tree = &preftrees->indirect_missing_keys;
 438	return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 439			      wanted_disk_byte, count, sc, gfp_mask);
 440}
 441
 442static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
 443{
 444	struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
 445	struct rb_node *parent = NULL;
 446	struct prelim_ref *ref = NULL;
 447	struct prelim_ref target = {};
 448	int result;
 449
 450	target.parent = bytenr;
 451
 452	while (*p) {
 453		parent = *p;
 454		ref = rb_entry(parent, struct prelim_ref, rbnode);
 455		result = prelim_ref_compare(ref, &target);
 456
 457		if (result < 0)
 458			p = &(*p)->rb_left;
 459		else if (result > 0)
 460			p = &(*p)->rb_right;
 461		else
 462			return 1;
 463	}
 464	return 0;
 465}
 466
 467static int add_all_parents(struct btrfs_backref_walk_ctx *ctx,
 468			   struct btrfs_root *root, struct btrfs_path *path,
 469			   struct ulist *parents,
 470			   struct preftrees *preftrees, struct prelim_ref *ref,
 471			   int level)
 472{
 473	int ret = 0;
 474	int slot;
 475	struct extent_buffer *eb;
 476	struct btrfs_key key;
 477	struct btrfs_key *key_for_search = &ref->key_for_search;
 478	struct btrfs_file_extent_item *fi;
 479	struct extent_inode_elem *eie = NULL, *old = NULL;
 480	u64 disk_byte;
 481	u64 wanted_disk_byte = ref->wanted_disk_byte;
 482	u64 count = 0;
 483	u64 data_offset;
 484	u8 type;
 485
 486	if (level != 0) {
 487		eb = path->nodes[level];
 488		ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 489		if (ret < 0)
 490			return ret;
 491		return 0;
 492	}
 493
 494	/*
 495	 * 1. We normally enter this function with the path already pointing to
 496	 *    the first item to check. But sometimes, we may enter it with
 497	 *    slot == nritems.
 498	 * 2. We are searching for normal backref but bytenr of this leaf
 499	 *    matches shared data backref
 500	 * 3. The leaf owner is not equal to the root we are searching
 501	 *
 502	 * For these cases, go to the next leaf before we continue.
 503	 */
 504	eb = path->nodes[0];
 505	if (path->slots[0] >= btrfs_header_nritems(eb) ||
 506	    is_shared_data_backref(preftrees, eb->start) ||
 507	    ref->root_id != btrfs_header_owner(eb)) {
 508		if (ctx->time_seq == BTRFS_SEQ_LAST)
 509			ret = btrfs_next_leaf(root, path);
 510		else
 511			ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
 512	}
 513
 514	while (!ret && count < ref->count) {
 515		eb = path->nodes[0];
 516		slot = path->slots[0];
 517
 518		btrfs_item_key_to_cpu(eb, &key, slot);
 519
 520		if (key.objectid != key_for_search->objectid ||
 521		    key.type != BTRFS_EXTENT_DATA_KEY)
 522			break;
 523
 524		/*
 525		 * We are searching for normal backref but bytenr of this leaf
 526		 * matches shared data backref, OR
 527		 * the leaf owner is not equal to the root we are searching for
 528		 */
 529		if (slot == 0 &&
 530		    (is_shared_data_backref(preftrees, eb->start) ||
 531		     ref->root_id != btrfs_header_owner(eb))) {
 532			if (ctx->time_seq == BTRFS_SEQ_LAST)
 533				ret = btrfs_next_leaf(root, path);
 534			else
 535				ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
 536			continue;
 537		}
 538		fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 539		type = btrfs_file_extent_type(eb, fi);
 540		if (type == BTRFS_FILE_EXTENT_INLINE)
 541			goto next;
 542		disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 543		data_offset = btrfs_file_extent_offset(eb, fi);
 544
 545		if (disk_byte == wanted_disk_byte) {
 546			eie = NULL;
 547			old = NULL;
 548			if (ref->key_for_search.offset == key.offset - data_offset)
 549				count++;
 550			else
 551				goto next;
 552			if (!ctx->skip_inode_ref_list) {
 553				ret = check_extent_in_eb(ctx, &key, eb, fi, &eie);
 554				if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
 555				    ret < 0)
 556					break;
 557			}
 558			if (ret > 0)
 559				goto next;
 560			ret = ulist_add_merge_ptr(parents, eb->start,
 561						  eie, (void **)&old, GFP_NOFS);
 562			if (ret < 0)
 563				break;
 564			if (!ret && !ctx->skip_inode_ref_list) {
 565				while (old->next)
 566					old = old->next;
 567				old->next = eie;
 568			}
 569			eie = NULL;
 570		}
 571next:
 572		if (ctx->time_seq == BTRFS_SEQ_LAST)
 573			ret = btrfs_next_item(root, path);
 574		else
 575			ret = btrfs_next_old_item(root, path, ctx->time_seq);
 576	}
 577
 578	if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
 579		free_inode_elem_list(eie);
 580	else if (ret > 0)
 581		ret = 0;
 582
 583	return ret;
 584}
 585
 586/*
 587 * resolve an indirect backref in the form (root_id, key, level)
 588 * to a logical address
 589 */
 590static int resolve_indirect_ref(struct btrfs_backref_walk_ctx *ctx,
 591				struct btrfs_path *path,
 592				struct preftrees *preftrees,
 593				struct prelim_ref *ref, struct ulist *parents)
 594{
 595	struct btrfs_root *root;
 596	struct extent_buffer *eb;
 597	int ret = 0;
 598	int root_level;
 599	int level = ref->level;
 600	struct btrfs_key search_key = ref->key_for_search;
 601
 602	/*
 603	 * If we're search_commit_root we could possibly be holding locks on
 604	 * other tree nodes.  This happens when qgroups does backref walks when
 605	 * adding new delayed refs.  To deal with this we need to look in cache
 606	 * for the root, and if we don't find it then we need to search the
 607	 * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
 608	 * here.
 609	 */
 610	if (path->search_commit_root)
 611		root = btrfs_get_fs_root_commit_root(ctx->fs_info, path, ref->root_id);
 612	else
 613		root = btrfs_get_fs_root(ctx->fs_info, ref->root_id, false);
 614	if (IS_ERR(root)) {
 615		ret = PTR_ERR(root);
 616		goto out_free;
 617	}
 618
 619	if (!path->search_commit_root &&
 620	    test_bit(BTRFS_ROOT_DELETING, &root->state)) {
 621		ret = -ENOENT;
 622		goto out;
 623	}
 624
 625	if (btrfs_is_testing(ctx->fs_info)) {
 626		ret = -ENOENT;
 627		goto out;
 628	}
 629
 630	if (path->search_commit_root)
 631		root_level = btrfs_header_level(root->commit_root);
 632	else if (ctx->time_seq == BTRFS_SEQ_LAST)
 633		root_level = btrfs_header_level(root->node);
 634	else
 635		root_level = btrfs_old_root_level(root, ctx->time_seq);
 636
 637	if (root_level + 1 == level)
 638		goto out;
 639
 640	/*
 641	 * We can often find data backrefs with an offset that is too large
 642	 * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
 643	 * subtracting a file's offset with the data offset of its
 644	 * corresponding extent data item. This can happen for example in the
 645	 * clone ioctl.
 646	 *
 647	 * So if we detect such case we set the search key's offset to zero to
 648	 * make sure we will find the matching file extent item at
 649	 * add_all_parents(), otherwise we will miss it because the offset
 650	 * taken form the backref is much larger then the offset of the file
 651	 * extent item. This can make us scan a very large number of file
 652	 * extent items, but at least it will not make us miss any.
 653	 *
 654	 * This is an ugly workaround for a behaviour that should have never
 655	 * existed, but it does and a fix for the clone ioctl would touch a lot
 656	 * of places, cause backwards incompatibility and would not fix the
 657	 * problem for extents cloned with older kernels.
 658	 */
 659	if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
 660	    search_key.offset >= LLONG_MAX)
 661		search_key.offset = 0;
 662	path->lowest_level = level;
 663	if (ctx->time_seq == BTRFS_SEQ_LAST)
 664		ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 665	else
 666		ret = btrfs_search_old_slot(root, &search_key, path, ctx->time_seq);
 667
 668	btrfs_debug(ctx->fs_info,
 669"search slot in root %llu (level %d, ref count %d) returned %d for key " BTRFS_KEY_FMT,
 670		    ref->root_id, level, ref->count, ret,
 671		    BTRFS_KEY_FMT_VALUE(&ref->key_for_search));
 672	if (ret < 0)
 673		goto out;
 674
 675	eb = path->nodes[level];
 676	while (!eb) {
 677		if (WARN_ON(!level)) {
 678			ret = 1;
 679			goto out;
 680		}
 681		level--;
 682		eb = path->nodes[level];
 683	}
 684
 685	ret = add_all_parents(ctx, root, path, parents, preftrees, ref, level);
 686out:
 687	btrfs_put_root(root);
 688out_free:
 689	path->lowest_level = 0;
 690	btrfs_release_path(path);
 691	return ret;
 692}
 693
 694static struct extent_inode_elem *
 695unode_aux_to_inode_list(struct ulist_node *node)
 696{
 697	if (!node)
 698		return NULL;
 699	return (struct extent_inode_elem *)(uintptr_t)node->aux;
 700}
 701
 702static void free_leaf_list(struct ulist *ulist)
 703{
 704	struct ulist_node *node;
 705	struct ulist_iterator uiter;
 706
 707	ULIST_ITER_INIT(&uiter);
 708	while ((node = ulist_next(ulist, &uiter)))
 709		free_inode_elem_list(unode_aux_to_inode_list(node));
 710
 711	ulist_free(ulist);
 712}
 713
 714/*
 715 * We maintain three separate rbtrees: one for direct refs, one for
 716 * indirect refs which have a key, and one for indirect refs which do not
 717 * have a key. Each tree does merge on insertion.
 718 *
 719 * Once all of the references are located, we iterate over the tree of
 720 * indirect refs with missing keys. An appropriate key is located and
 721 * the ref is moved onto the tree for indirect refs. After all missing
 722 * keys are thus located, we iterate over the indirect ref tree, resolve
 723 * each reference, and then insert the resolved reference onto the
 724 * direct tree (merging there too).
 725 *
 726 * New backrefs (i.e., for parent nodes) are added to the appropriate
 727 * rbtree as they are encountered. The new backrefs are subsequently
 728 * resolved as above.
 729 */
 730static int resolve_indirect_refs(struct btrfs_backref_walk_ctx *ctx,
 731				 struct btrfs_path *path,
 732				 struct preftrees *preftrees,
 733				 struct share_check *sc)
 734{
 735	int ret = 0;
 736	struct ulist *parents;
 737	struct ulist_node *node;
 738	struct ulist_iterator uiter;
 739	struct rb_node *rnode;
 740
 741	parents = ulist_alloc(GFP_NOFS);
 742	if (!parents)
 743		return -ENOMEM;
 744
 745	/*
 746	 * We could trade memory usage for performance here by iterating
 747	 * the tree, allocating new refs for each insertion, and then
 748	 * freeing the entire indirect tree when we're done.  In some test
 749	 * cases, the tree can grow quite large (~200k objects).
 750	 */
 751	while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
 752		struct prelim_ref *ref;
 753		int ret2;
 754
 755		ref = rb_entry(rnode, struct prelim_ref, rbnode);
 756		if (WARN(ref->parent,
 757			 "BUG: direct ref found in indirect tree")) {
 758			ret = -EINVAL;
 759			goto out;
 760		}
 761
 762		rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
 763		preftrees->indirect.count--;
 764
 765		if (ref->count == 0) {
 766			free_pref(ref);
 767			continue;
 768		}
 769
 770		if (sc && ref->root_id != btrfs_root_id(sc->root)) {
 771			free_pref(ref);
 772			ret = BACKREF_FOUND_SHARED;
 773			goto out;
 774		}
 775		ret2 = resolve_indirect_ref(ctx, path, preftrees, ref, parents);
 776		/*
 777		 * we can only tolerate ENOENT,otherwise,we should catch error
 778		 * and return directly.
 779		 */
 780		if (ret2 == -ENOENT) {
 781			prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref,
 782					  NULL);
 783			continue;
 784		} else if (ret2) {
 785			free_pref(ref);
 786			ret = ret2;
 787			goto out;
 788		}
 789
 790		/* we put the first parent into the ref at hand */
 791		ULIST_ITER_INIT(&uiter);
 792		node = ulist_next(parents, &uiter);
 793		ref->parent = node ? node->val : 0;
 794		ref->inode_list = unode_aux_to_inode_list(node);
 795
 796		/* Add a prelim_ref(s) for any other parent(s). */
 797		while ((node = ulist_next(parents, &uiter))) {
 798			struct prelim_ref *new_ref;
 799
 800			new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 801						   GFP_NOFS);
 802			if (!new_ref) {
 803				free_pref(ref);
 804				ret = -ENOMEM;
 805				goto out;
 806			}
 807			memcpy(new_ref, ref, sizeof(*ref));
 808			new_ref->parent = node->val;
 809			new_ref->inode_list = unode_aux_to_inode_list(node);
 810			prelim_ref_insert(ctx->fs_info, &preftrees->direct,
 811					  new_ref, NULL);
 812		}
 813
 814		/*
 815		 * Now it's a direct ref, put it in the direct tree. We must
 816		 * do this last because the ref could be merged/freed here.
 817		 */
 818		prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref, NULL);
 819
 820		ulist_reinit(parents);
 821		cond_resched();
 822	}
 823out:
 824	/*
 825	 * We may have inode lists attached to refs in the parents ulist, so we
 826	 * must free them before freeing the ulist and its refs.
 827	 */
 828	free_leaf_list(parents);
 829	return ret;
 830}
 831
 832/*
 833 * read tree blocks and add keys where required.
 834 */
 835static int add_missing_keys(struct btrfs_fs_info *fs_info,
 836			    struct preftrees *preftrees, bool lock)
 837{
 838	struct prelim_ref *ref;
 839	struct extent_buffer *eb;
 840	struct preftree *tree = &preftrees->indirect_missing_keys;
 841	struct rb_node *node;
 842
 843	while ((node = rb_first_cached(&tree->root))) {
 844		struct btrfs_tree_parent_check check = { 0 };
 845
 846		ref = rb_entry(node, struct prelim_ref, rbnode);
 847		rb_erase_cached(node, &tree->root);
 848
 849		BUG_ON(ref->parent);	/* should not be a direct ref */
 850		BUG_ON(ref->key_for_search.type);
 851		BUG_ON(!ref->wanted_disk_byte);
 852
 853		check.level = ref->level - 1;
 854		check.owner_root = ref->root_id;
 855
 856		eb = read_tree_block(fs_info, ref->wanted_disk_byte, &check);
 857		if (IS_ERR(eb)) {
 858			free_pref(ref);
 859			return PTR_ERR(eb);
 860		}
 861		if (unlikely(!extent_buffer_uptodate(eb))) {
 862			free_pref(ref);
 863			free_extent_buffer(eb);
 864			return -EIO;
 865		}
 866
 867		if (lock)
 868			btrfs_tree_read_lock(eb);
 869		if (btrfs_header_level(eb) == 0)
 870			btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 871		else
 872			btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 873		if (lock)
 874			btrfs_tree_read_unlock(eb);
 875		free_extent_buffer(eb);
 876		prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 877		cond_resched();
 878	}
 879	return 0;
 880}
 881
 882/*
 883 * add all currently queued delayed refs from this head whose seq nr is
 884 * smaller or equal that seq to the list
 885 */
 886static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 887			    struct btrfs_delayed_ref_head *head, u64 seq,
 888			    struct preftrees *preftrees, struct share_check *sc)
 889{
 890	struct btrfs_delayed_ref_node *node;
 891	struct btrfs_key key;
 892	struct rb_node *n;
 893	int count;
 894	int ret = 0;
 895
 896	spin_lock(&head->lock);
 897	for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 898		node = rb_entry(n, struct btrfs_delayed_ref_node,
 899				ref_node);
 900		if (node->seq > seq)
 901			continue;
 902
 903		switch (node->action) {
 904		case BTRFS_ADD_DELAYED_EXTENT:
 905		case BTRFS_UPDATE_DELAYED_HEAD:
 906			WARN_ON(1);
 907			continue;
 908		case BTRFS_ADD_DELAYED_REF:
 909			count = node->ref_mod;
 910			break;
 911		case BTRFS_DROP_DELAYED_REF:
 912			count = node->ref_mod * -1;
 913			break;
 914		default:
 915			BUG();
 916		}
 917		switch (node->type) {
 918		case BTRFS_TREE_BLOCK_REF_KEY: {
 919			/* NORMAL INDIRECT METADATA backref */
 920			struct btrfs_key *key_ptr = NULL;
 921			/* The owner of a tree block ref is the level. */
 922			int level = btrfs_delayed_ref_owner(node);
 923
 924			if (head->extent_op && head->extent_op->update_key) {
 925				btrfs_disk_key_to_cpu(&key, &head->extent_op->key);
 926				key_ptr = &key;
 927			}
 928
 929			ret = add_indirect_ref(fs_info, preftrees, node->ref_root,
 930					       key_ptr, level + 1, node->bytenr,
 931					       count, sc, GFP_ATOMIC);
 932			break;
 933		}
 934		case BTRFS_SHARED_BLOCK_REF_KEY: {
 935			/*
 936			 * SHARED DIRECT METADATA backref
 937			 *
 938			 * The owner of a tree block ref is the level.
 939			 */
 940			int level = btrfs_delayed_ref_owner(node);
 941
 942			ret = add_direct_ref(fs_info, preftrees, level + 1,
 943					     node->parent, node->bytenr, count,
 944					     sc, GFP_ATOMIC);
 945			break;
 946		}
 947		case BTRFS_EXTENT_DATA_REF_KEY: {
 948			/* NORMAL INDIRECT DATA backref */
 949			key.objectid = btrfs_delayed_ref_owner(node);
 950			key.type = BTRFS_EXTENT_DATA_KEY;
 951			key.offset = btrfs_delayed_ref_offset(node);
 952
 953			/*
 954			 * If we have a share check context and a reference for
 955			 * another inode, we can't exit immediately. This is
 956			 * because even if this is a BTRFS_ADD_DELAYED_REF
 957			 * reference we may find next a BTRFS_DROP_DELAYED_REF
 958			 * which cancels out this ADD reference.
 959			 *
 960			 * If this is a DROP reference and there was no previous
 961			 * ADD reference, then we need to signal that when we
 962			 * process references from the extent tree (through
 963			 * add_inline_refs() and add_keyed_refs()), we should
 964			 * not exit early if we find a reference for another
 965			 * inode, because one of the delayed DROP references
 966			 * may cancel that reference in the extent tree.
 967			 */
 968			if (sc && count < 0)
 969				sc->have_delayed_delete_refs = true;
 970
 971			ret = add_indirect_ref(fs_info, preftrees, node->ref_root,
 972					       &key, 0, node->bytenr, count, sc,
 973					       GFP_ATOMIC);
 974			break;
 975		}
 976		case BTRFS_SHARED_DATA_REF_KEY: {
 977			/* SHARED DIRECT FULL backref */
 978			ret = add_direct_ref(fs_info, preftrees, 0, node->parent,
 979					     node->bytenr, count, sc,
 980					     GFP_ATOMIC);
 981			break;
 982		}
 983		default:
 984			WARN_ON(1);
 985		}
 986		/*
 987		 * We must ignore BACKREF_FOUND_SHARED until all delayed
 988		 * refs have been checked.
 989		 */
 990		if (ret && (ret != BACKREF_FOUND_SHARED))
 991			break;
 992	}
 993	if (!ret)
 994		ret = extent_is_shared(sc);
 995
 996	spin_unlock(&head->lock);
 997	return ret;
 998}
 999
1000/*
1001 * add all inline backrefs for bytenr to the list
1002 *
1003 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1004 */
1005static int add_inline_refs(struct btrfs_backref_walk_ctx *ctx,
1006			   struct btrfs_path *path,
1007			   int *info_level, struct preftrees *preftrees,
1008			   struct share_check *sc)
1009{
1010	int ret = 0;
1011	int slot;
1012	struct extent_buffer *leaf;
1013	struct btrfs_key key;
1014	struct btrfs_key found_key;
1015	unsigned long ptr;
1016	unsigned long end;
1017	struct btrfs_extent_item *ei;
1018	u64 flags;
1019	u64 item_size;
1020
1021	/*
1022	 * enumerate all inline refs
1023	 */
1024	leaf = path->nodes[0];
1025	slot = path->slots[0];
1026
1027	item_size = btrfs_item_size(leaf, slot);
1028	ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
1029
1030	if (ctx->check_extent_item) {
1031		ret = ctx->check_extent_item(ctx->bytenr, ei, leaf, ctx->user_ctx);
1032		if (ret)
1033			return ret;
1034	}
1035
1036	flags = btrfs_extent_flags(leaf, ei);
1037	btrfs_item_key_to_cpu(leaf, &found_key, slot);
1038
1039	ptr = (unsigned long)(ei + 1);
1040	end = (unsigned long)ei + item_size;
1041
1042	if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
1043	    flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1044		struct btrfs_tree_block_info *info;
1045
1046		info = (struct btrfs_tree_block_info *)ptr;
1047		*info_level = btrfs_tree_block_level(leaf, info);
1048		ptr += sizeof(struct btrfs_tree_block_info);
1049		BUG_ON(ptr > end);
1050	} else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1051		*info_level = found_key.offset;
1052	} else {
1053		BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1054	}
1055
1056	while (ptr < end) {
1057		struct btrfs_extent_inline_ref *iref;
1058		u64 offset;
1059		int type;
1060
1061		iref = (struct btrfs_extent_inline_ref *)ptr;
1062		type = btrfs_get_extent_inline_ref_type(leaf, iref,
1063							BTRFS_REF_TYPE_ANY);
1064		if (unlikely(type == BTRFS_REF_TYPE_INVALID))
1065			return -EUCLEAN;
1066
1067		offset = btrfs_extent_inline_ref_offset(leaf, iref);
1068
1069		switch (type) {
1070		case BTRFS_SHARED_BLOCK_REF_KEY:
1071			ret = add_direct_ref(ctx->fs_info, preftrees,
1072					     *info_level + 1, offset,
1073					     ctx->bytenr, 1, NULL, GFP_NOFS);
1074			break;
1075		case BTRFS_SHARED_DATA_REF_KEY: {
1076			struct btrfs_shared_data_ref *sdref;
1077			int count;
1078
1079			sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1080			count = btrfs_shared_data_ref_count(leaf, sdref);
1081
1082			ret = add_direct_ref(ctx->fs_info, preftrees, 0, offset,
1083					     ctx->bytenr, count, sc, GFP_NOFS);
1084			break;
1085		}
1086		case BTRFS_TREE_BLOCK_REF_KEY:
1087			ret = add_indirect_ref(ctx->fs_info, preftrees, offset,
1088					       NULL, *info_level + 1,
1089					       ctx->bytenr, 1, NULL, GFP_NOFS);
1090			break;
1091		case BTRFS_EXTENT_DATA_REF_KEY: {
1092			struct btrfs_extent_data_ref *dref;
1093			int count;
1094			u64 root;
1095
1096			dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1097			count = btrfs_extent_data_ref_count(leaf, dref);
1098			key.objectid = btrfs_extent_data_ref_objectid(leaf,
1099								      dref);
1100			key.type = BTRFS_EXTENT_DATA_KEY;
1101			key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1102
1103			if (sc && key.objectid != sc->inum &&
1104			    !sc->have_delayed_delete_refs) {
1105				ret = BACKREF_FOUND_SHARED;
1106				break;
1107			}
1108
1109			root = btrfs_extent_data_ref_root(leaf, dref);
1110
1111			if (!ctx->skip_data_ref ||
1112			    !ctx->skip_data_ref(root, key.objectid, key.offset,
1113						ctx->user_ctx))
1114				ret = add_indirect_ref(ctx->fs_info, preftrees,
1115						       root, &key, 0, ctx->bytenr,
1116						       count, sc, GFP_NOFS);
1117			break;
1118		}
1119		case BTRFS_EXTENT_OWNER_REF_KEY:
1120			ASSERT(btrfs_fs_incompat(ctx->fs_info, SIMPLE_QUOTA));
1121			break;
1122		default:
1123			WARN_ON(1);
1124		}
1125		if (ret)
1126			return ret;
1127		ptr += btrfs_extent_inline_ref_size(type);
1128	}
1129
1130	return 0;
1131}
1132
1133/*
1134 * add all non-inline backrefs for bytenr to the list
1135 *
1136 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1137 */
1138static int add_keyed_refs(struct btrfs_backref_walk_ctx *ctx,
1139			  struct btrfs_root *extent_root,
1140			  struct btrfs_path *path,
1141			  int info_level, struct preftrees *preftrees,
1142			  struct share_check *sc)
1143{
1144	struct btrfs_fs_info *fs_info = extent_root->fs_info;
1145	int ret;
1146	int slot;
1147	struct extent_buffer *leaf;
1148	struct btrfs_key key;
1149
1150	while (1) {
1151		ret = btrfs_next_item(extent_root, path);
1152		if (ret < 0)
1153			break;
1154		if (ret) {
1155			ret = 0;
1156			break;
1157		}
1158
1159		slot = path->slots[0];
1160		leaf = path->nodes[0];
1161		btrfs_item_key_to_cpu(leaf, &key, slot);
1162
1163		if (key.objectid != ctx->bytenr)
1164			break;
1165		if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1166			continue;
1167		if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1168			break;
1169
1170		switch (key.type) {
1171		case BTRFS_SHARED_BLOCK_REF_KEY:
1172			/* SHARED DIRECT METADATA backref */
1173			ret = add_direct_ref(fs_info, preftrees,
1174					     info_level + 1, key.offset,
1175					     ctx->bytenr, 1, NULL, GFP_NOFS);
1176			break;
1177		case BTRFS_SHARED_DATA_REF_KEY: {
1178			/* SHARED DIRECT FULL backref */
1179			struct btrfs_shared_data_ref *sdref;
1180			int count;
1181
1182			sdref = btrfs_item_ptr(leaf, slot,
1183					      struct btrfs_shared_data_ref);
1184			count = btrfs_shared_data_ref_count(leaf, sdref);
1185			ret = add_direct_ref(fs_info, preftrees, 0,
1186					     key.offset, ctx->bytenr, count,
1187					     sc, GFP_NOFS);
1188			break;
1189		}
1190		case BTRFS_TREE_BLOCK_REF_KEY:
1191			/* NORMAL INDIRECT METADATA backref */
1192			ret = add_indirect_ref(fs_info, preftrees, key.offset,
1193					       NULL, info_level + 1, ctx->bytenr,
1194					       1, NULL, GFP_NOFS);
1195			break;
1196		case BTRFS_EXTENT_DATA_REF_KEY: {
1197			/* NORMAL INDIRECT DATA backref */
1198			struct btrfs_extent_data_ref *dref;
1199			int count;
1200			u64 root;
1201
1202			dref = btrfs_item_ptr(leaf, slot,
1203					      struct btrfs_extent_data_ref);
1204			count = btrfs_extent_data_ref_count(leaf, dref);
1205			key.objectid = btrfs_extent_data_ref_objectid(leaf,
1206								      dref);
1207			key.type = BTRFS_EXTENT_DATA_KEY;
1208			key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1209
1210			if (sc && key.objectid != sc->inum &&
1211			    !sc->have_delayed_delete_refs) {
1212				ret = BACKREF_FOUND_SHARED;
1213				break;
1214			}
1215
1216			root = btrfs_extent_data_ref_root(leaf, dref);
1217
1218			if (!ctx->skip_data_ref ||
1219			    !ctx->skip_data_ref(root, key.objectid, key.offset,
1220						ctx->user_ctx))
1221				ret = add_indirect_ref(fs_info, preftrees, root,
1222						       &key, 0, ctx->bytenr,
1223						       count, sc, GFP_NOFS);
1224			break;
1225		}
1226		default:
1227			WARN_ON(1);
1228		}
1229		if (ret)
1230			return ret;
1231
1232	}
1233
1234	return ret;
1235}
1236
1237/*
1238 * The caller has joined a transaction or is holding a read lock on the
1239 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1240 * snapshot field changing while updating or checking the cache.
1241 */
1242static bool lookup_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1243					struct btrfs_root *root,
1244					u64 bytenr, int level, bool *is_shared)
1245{
1246	const struct btrfs_fs_info *fs_info = root->fs_info;
1247	struct btrfs_backref_shared_cache_entry *entry;
1248
1249	if (!current->journal_info)
1250		lockdep_assert_held(&fs_info->commit_root_sem);
1251
1252	if (!ctx->use_path_cache)
1253		return false;
1254
1255	if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1256		return false;
1257
1258	/*
1259	 * Level -1 is used for the data extent, which is not reliable to cache
1260	 * because its reference count can increase or decrease without us
1261	 * realizing. We cache results only for extent buffers that lead from
1262	 * the root node down to the leaf with the file extent item.
1263	 */
1264	ASSERT(level >= 0);
1265
1266	entry = &ctx->path_cache_entries[level];
1267
1268	/* Unused cache entry or being used for some other extent buffer. */
1269	if (entry->bytenr != bytenr)
1270		return false;
1271
1272	/*
1273	 * We cached a false result, but the last snapshot generation of the
1274	 * root changed, so we now have a snapshot. Don't trust the result.
1275	 */
1276	if (!entry->is_shared &&
1277	    entry->gen != btrfs_root_last_snapshot(&root->root_item))
1278		return false;
1279
1280	/*
1281	 * If we cached a true result and the last generation used for dropping
1282	 * a root changed, we can not trust the result, because the dropped root
1283	 * could be a snapshot sharing this extent buffer.
1284	 */
1285	if (entry->is_shared &&
1286	    entry->gen != btrfs_get_last_root_drop_gen(fs_info))
1287		return false;
1288
1289	*is_shared = entry->is_shared;
1290	/*
1291	 * If the node at this level is shared, than all nodes below are also
1292	 * shared. Currently some of the nodes below may be marked as not shared
1293	 * because we have just switched from one leaf to another, and switched
1294	 * also other nodes above the leaf and below the current level, so mark
1295	 * them as shared.
1296	 */
1297	if (*is_shared) {
1298		for (int i = 0; i < level; i++) {
1299			ctx->path_cache_entries[i].is_shared = true;
1300			ctx->path_cache_entries[i].gen = entry->gen;
1301		}
1302	}
1303
1304	return true;
1305}
1306
1307/*
1308 * The caller has joined a transaction or is holding a read lock on the
1309 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1310 * snapshot field changing while updating or checking the cache.
1311 */
1312static void store_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1313				       struct btrfs_root *root,
1314				       u64 bytenr, int level, bool is_shared)
1315{
1316	const struct btrfs_fs_info *fs_info = root->fs_info;
1317	struct btrfs_backref_shared_cache_entry *entry;
1318	u64 gen;
1319
1320	if (!current->journal_info)
1321		lockdep_assert_held(&fs_info->commit_root_sem);
1322
1323	if (!ctx->use_path_cache)
1324		return;
1325
1326	if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1327		return;
1328
1329	/*
1330	 * Level -1 is used for the data extent, which is not reliable to cache
1331	 * because its reference count can increase or decrease without us
1332	 * realizing. We cache results only for extent buffers that lead from
1333	 * the root node down to the leaf with the file extent item.
1334	 */
1335	ASSERT(level >= 0);
1336
1337	if (is_shared)
1338		gen = btrfs_get_last_root_drop_gen(fs_info);
1339	else
1340		gen = btrfs_root_last_snapshot(&root->root_item);
1341
1342	entry = &ctx->path_cache_entries[level];
1343	entry->bytenr = bytenr;
1344	entry->is_shared = is_shared;
1345	entry->gen = gen;
1346
1347	/*
1348	 * If we found an extent buffer is shared, set the cache result for all
1349	 * extent buffers below it to true. As nodes in the path are COWed,
1350	 * their sharedness is moved to their children, and if a leaf is COWed,
1351	 * then the sharedness of a data extent becomes direct, the refcount of
1352	 * data extent is increased in the extent item at the extent tree.
1353	 */
1354	if (is_shared) {
1355		for (int i = 0; i < level; i++) {
1356			entry = &ctx->path_cache_entries[i];
1357			entry->is_shared = is_shared;
1358			entry->gen = gen;
1359		}
1360	}
1361}
1362
1363/*
1364 * this adds all existing backrefs (inline backrefs, backrefs and delayed
1365 * refs) for the given bytenr to the refs list, merges duplicates and resolves
1366 * indirect refs to their parent bytenr.
1367 * When roots are found, they're added to the roots list
1368 *
1369 * @ctx:     Backref walking context object, must be not NULL.
1370 * @sc:      If !NULL, then immediately return BACKREF_FOUND_SHARED when a
1371 *           shared extent is detected.
1372 *
1373 * Otherwise this returns 0 for success and <0 for an error.
1374 *
1375 * FIXME some caching might speed things up
1376 */
1377static int find_parent_nodes(struct btrfs_backref_walk_ctx *ctx,
1378			     struct share_check *sc)
1379{
1380	struct btrfs_root *root = btrfs_extent_root(ctx->fs_info, ctx->bytenr);
1381	struct btrfs_key key;
1382	struct btrfs_path *path;
1383	struct btrfs_delayed_ref_root *delayed_refs = NULL;
1384	struct btrfs_delayed_ref_head *head;
1385	int info_level = 0;
1386	int ret;
1387	struct prelim_ref *ref;
1388	struct rb_node *node;
1389	struct extent_inode_elem *eie = NULL;
1390	struct preftrees preftrees = {
1391		.direct = PREFTREE_INIT,
1392		.indirect = PREFTREE_INIT,
1393		.indirect_missing_keys = PREFTREE_INIT
1394	};
1395
1396	/* Roots ulist is not needed when using a sharedness check context. */
1397	if (sc)
1398		ASSERT(ctx->roots == NULL);
1399
1400	key.objectid = ctx->bytenr;
1401	if (btrfs_fs_incompat(ctx->fs_info, SKINNY_METADATA))
1402		key.type = BTRFS_METADATA_ITEM_KEY;
1403	else
1404		key.type = BTRFS_EXTENT_ITEM_KEY;
1405	key.offset = (u64)-1;
1406
1407	path = btrfs_alloc_path();
1408	if (!path)
1409		return -ENOMEM;
1410	if (!ctx->trans) {
1411		path->search_commit_root = true;
1412		path->skip_locking = true;
1413	}
1414
1415	if (ctx->time_seq == BTRFS_SEQ_LAST)
1416		path->skip_locking = true;
1417
1418again:
1419	head = NULL;
1420
1421	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1422	if (ret < 0)
1423		goto out;
1424	if (unlikely(ret == 0)) {
1425		/*
1426		 * Key with offset -1 found, there would have to exist an extent
1427		 * item with such offset, but this is out of the valid range.
1428		 */
1429		ret = -EUCLEAN;
1430		goto out;
1431	}
1432
1433	if (ctx->trans && likely(ctx->trans->type != __TRANS_DUMMY) &&
1434	    ctx->time_seq != BTRFS_SEQ_LAST) {
1435		/*
1436		 * We have a specific time_seq we care about and trans which
1437		 * means we have the path lock, we need to grab the ref head and
1438		 * lock it so we have a consistent view of the refs at the given
1439		 * time.
1440		 */
1441		delayed_refs = &ctx->trans->transaction->delayed_refs;
1442		spin_lock(&delayed_refs->lock);
1443		head = btrfs_find_delayed_ref_head(ctx->fs_info, delayed_refs,
1444						   ctx->bytenr);
1445		if (head) {
1446			if (!mutex_trylock(&head->mutex)) {
1447				refcount_inc(&head->refs);
1448				spin_unlock(&delayed_refs->lock);
1449
1450				btrfs_release_path(path);
1451
1452				/*
1453				 * Mutex was contended, block until it's
1454				 * released and try again
1455				 */
1456				mutex_lock(&head->mutex);
1457				mutex_unlock(&head->mutex);
1458				btrfs_put_delayed_ref_head(head);
1459				goto again;
1460			}
1461			spin_unlock(&delayed_refs->lock);
1462			ret = add_delayed_refs(ctx->fs_info, head, ctx->time_seq,
1463					       &preftrees, sc);
1464			mutex_unlock(&head->mutex);
1465			if (ret)
1466				goto out;
1467		} else {
1468			spin_unlock(&delayed_refs->lock);
1469		}
1470	}
1471
1472	if (path->slots[0]) {
1473		struct extent_buffer *leaf;
1474		int slot;
1475
1476		path->slots[0]--;
1477		leaf = path->nodes[0];
1478		slot = path->slots[0];
1479		btrfs_item_key_to_cpu(leaf, &key, slot);
1480		if (key.objectid == ctx->bytenr &&
1481		    (key.type == BTRFS_EXTENT_ITEM_KEY ||
1482		     key.type == BTRFS_METADATA_ITEM_KEY)) {
1483			ret = add_inline_refs(ctx, path, &info_level,
1484					      &preftrees, sc);
1485			if (ret)
1486				goto out;
1487			ret = add_keyed_refs(ctx, root, path, info_level,
1488					     &preftrees, sc);
1489			if (ret)
1490				goto out;
1491		}
1492	}
1493
1494	/*
1495	 * If we have a share context and we reached here, it means the extent
1496	 * is not directly shared (no multiple reference items for it),
1497	 * otherwise we would have exited earlier with a return value of
1498	 * BACKREF_FOUND_SHARED after processing delayed references or while
1499	 * processing inline or keyed references from the extent tree.
1500	 * The extent may however be indirectly shared through shared subtrees
1501	 * as a result from creating snapshots, so we determine below what is
1502	 * its parent node, in case we are dealing with a metadata extent, or
1503	 * what's the leaf (or leaves), from a fs tree, that has a file extent
1504	 * item pointing to it in case we are dealing with a data extent.
1505	 */
1506	ASSERT(extent_is_shared(sc) == 0);
1507
1508	/*
1509	 * If we are here for a data extent and we have a share_check structure
1510	 * it means the data extent is not directly shared (does not have
1511	 * multiple reference items), so we have to check if a path in the fs
1512	 * tree (going from the root node down to the leaf that has the file
1513	 * extent item pointing to the data extent) is shared, that is, if any
1514	 * of the extent buffers in the path is referenced by other trees.
1515	 */
1516	if (sc && ctx->bytenr == sc->data_bytenr) {
1517		/*
1518		 * If our data extent is from a generation more recent than the
1519		 * last generation used to snapshot the root, then we know that
1520		 * it can not be shared through subtrees, so we can skip
1521		 * resolving indirect references, there's no point in
1522		 * determining the extent buffers for the path from the fs tree
1523		 * root node down to the leaf that has the file extent item that
1524		 * points to the data extent.
1525		 */
1526		if (sc->data_extent_gen >
1527		    btrfs_root_last_snapshot(&sc->root->root_item)) {
1528			ret = BACKREF_FOUND_NOT_SHARED;
1529			goto out;
1530		}
1531
1532		/*
1533		 * If we are only determining if a data extent is shared or not
1534		 * and the corresponding file extent item is located in the same
1535		 * leaf as the previous file extent item, we can skip resolving
1536		 * indirect references for a data extent, since the fs tree path
1537		 * is the same (same leaf, so same path). We skip as long as the
1538		 * cached result for the leaf is valid and only if there's only
1539		 * one file extent item pointing to the data extent, because in
1540		 * the case of multiple file extent items, they may be located
1541		 * in different leaves and therefore we have multiple paths.
1542		 */
1543		if (sc->ctx->curr_leaf_bytenr == sc->ctx->prev_leaf_bytenr &&
1544		    sc->self_ref_count == 1) {
1545			bool cached;
1546			bool is_shared;
1547
1548			cached = lookup_backref_shared_cache(sc->ctx, sc->root,
1549						     sc->ctx->curr_leaf_bytenr,
1550						     0, &is_shared);
1551			if (cached) {
1552				if (is_shared)
1553					ret = BACKREF_FOUND_SHARED;
1554				else
1555					ret = BACKREF_FOUND_NOT_SHARED;
1556				goto out;
1557			}
1558		}
1559	}
1560
1561	btrfs_release_path(path);
1562
1563	ret = add_missing_keys(ctx->fs_info, &preftrees, !path->skip_locking);
1564	if (ret)
1565		goto out;
1566
1567	WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1568
1569	ret = resolve_indirect_refs(ctx, path, &preftrees, sc);
1570	if (ret)
1571		goto out;
1572
1573	WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1574
1575	/*
1576	 * This walks the tree of merged and resolved refs. Tree blocks are
1577	 * read in as needed. Unique entries are added to the ulist, and
1578	 * the list of found roots is updated.
1579	 *
1580	 * We release the entire tree in one go before returning.
1581	 */
1582	node = rb_first_cached(&preftrees.direct.root);
1583	while (node) {
1584		ref = rb_entry(node, struct prelim_ref, rbnode);
1585		node = rb_next(&ref->rbnode);
1586		/*
1587		 * ref->count < 0 can happen here if there are delayed
1588		 * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1589		 * prelim_ref_insert() relies on this when merging
1590		 * identical refs to keep the overall count correct.
1591		 * prelim_ref_insert() will merge only those refs
1592		 * which compare identically.  Any refs having
1593		 * e.g. different offsets would not be merged,
1594		 * and would retain their original ref->count < 0.
1595		 */
1596		if (ctx->roots && ref->count && ref->root_id && ref->parent == 0) {
1597			/* no parent == root of tree */
1598			ret = ulist_add(ctx->roots, ref->root_id, 0, GFP_NOFS);
1599			if (ret < 0)
1600				goto out;
1601		}
1602		if (ref->count && ref->parent) {
1603			if (!ctx->skip_inode_ref_list && !ref->inode_list &&
1604			    ref->level == 0) {
1605				struct btrfs_tree_parent_check check = { 0 };
1606				struct extent_buffer *eb;
1607
1608				check.level = ref->level;
1609
1610				eb = read_tree_block(ctx->fs_info, ref->parent,
1611						     &check);
1612				if (IS_ERR(eb)) {
1613					ret = PTR_ERR(eb);
1614					goto out;
1615				}
1616				if (unlikely(!extent_buffer_uptodate(eb))) {
1617					free_extent_buffer(eb);
1618					ret = -EIO;
1619					goto out;
1620				}
1621
1622				if (!path->skip_locking)
1623					btrfs_tree_read_lock(eb);
1624				ret = find_extent_in_eb(ctx, eb, &eie);
1625				if (!path->skip_locking)
1626					btrfs_tree_read_unlock(eb);
1627				free_extent_buffer(eb);
1628				if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1629				    ret < 0)
1630					goto out;
1631				ref->inode_list = eie;
1632				/*
1633				 * We transferred the list ownership to the ref,
1634				 * so set to NULL to avoid a double free in case
1635				 * an error happens after this.
1636				 */
1637				eie = NULL;
1638			}
1639			ret = ulist_add_merge_ptr(ctx->refs, ref->parent,
1640						  ref->inode_list,
1641						  (void **)&eie, GFP_NOFS);
1642			if (ret < 0)
1643				goto out;
1644			if (!ret && !ctx->skip_inode_ref_list) {
1645				/*
1646				 * We've recorded that parent, so we must extend
1647				 * its inode list here.
1648				 *
1649				 * However if there was corruption we may not
1650				 * have found an eie, return an error in this
1651				 * case.
1652				 */
1653				ASSERT(eie);
1654				if (unlikely(!eie)) {
1655					ret = -EUCLEAN;
1656					goto out;
1657				}
1658				while (eie->next)
1659					eie = eie->next;
1660				eie->next = ref->inode_list;
1661			}
1662			eie = NULL;
1663			/*
1664			 * We have transferred the inode list ownership from
1665			 * this ref to the ref we added to the 'refs' ulist.
1666			 * So set this ref's inode list to NULL to avoid
1667			 * use-after-free when our caller uses it or double
1668			 * frees in case an error happens before we return.
1669			 */
1670			ref->inode_list = NULL;
1671		}
1672		cond_resched();
1673	}
1674
1675out:
1676	btrfs_free_path(path);
1677
1678	prelim_release(&preftrees.direct);
1679	prelim_release(&preftrees.indirect);
1680	prelim_release(&preftrees.indirect_missing_keys);
1681
1682	if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
1683		free_inode_elem_list(eie);
1684	return ret;
1685}
1686
1687/*
1688 * Finds all leaves with a reference to the specified combination of
1689 * @ctx->bytenr and @ctx->extent_item_pos. The bytenr of the found leaves are
1690 * added to the ulist at @ctx->refs, and that ulist is allocated by this
1691 * function. The caller should free the ulist with free_leaf_list() if
1692 * @ctx->ignore_extent_item_pos is false, otherwise a simple ulist_free() is
1693 * enough.
1694 *
1695 * Returns 0 on success and < 0 on error. On error @ctx->refs is not allocated.
1696 */
1697int btrfs_find_all_leafs(struct btrfs_backref_walk_ctx *ctx)
1698{
1699	int ret;
1700
1701	ASSERT(ctx->refs == NULL);
1702
1703	ctx->refs = ulist_alloc(GFP_NOFS);
1704	if (!ctx->refs)
1705		return -ENOMEM;
1706
1707	ret = find_parent_nodes(ctx, NULL);
1708	if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1709	    (ret < 0 && ret != -ENOENT)) {
1710		free_leaf_list(ctx->refs);
1711		ctx->refs = NULL;
1712		return ret;
1713	}
1714
1715	return 0;
1716}
1717
1718/*
1719 * Walk all backrefs for a given extent to find all roots that reference this
1720 * extent. Walking a backref means finding all extents that reference this
1721 * extent and in turn walk the backrefs of those, too. Naturally this is a
1722 * recursive process, but here it is implemented in an iterative fashion: We
1723 * find all referencing extents for the extent in question and put them on a
1724 * list. In turn, we find all referencing extents for those, further appending
1725 * to the list. The way we iterate the list allows adding more elements after
1726 * the current while iterating. The process stops when we reach the end of the
1727 * list.
1728 *
1729 * Found roots are added to @ctx->roots, which is allocated by this function if
1730 * it points to NULL, in which case the caller is responsible for freeing it
1731 * after it's not needed anymore.
1732 * This function requires @ctx->refs to be NULL, as it uses it for allocating a
1733 * ulist to do temporary work, and frees it before returning.
1734 *
1735 * Returns 0 on success, < 0 on error.
1736 */
1737static int btrfs_find_all_roots_safe(struct btrfs_backref_walk_ctx *ctx)
1738{
1739	const u64 orig_bytenr = ctx->bytenr;
1740	const bool orig_skip_inode_ref_list = ctx->skip_inode_ref_list;
1741	bool roots_ulist_allocated = false;
1742	struct ulist_iterator uiter;
1743	int ret = 0;
1744
1745	ASSERT(ctx->refs == NULL);
1746
1747	ctx->refs = ulist_alloc(GFP_NOFS);
1748	if (!ctx->refs)
1749		return -ENOMEM;
1750
1751	if (!ctx->roots) {
1752		ctx->roots = ulist_alloc(GFP_NOFS);
1753		if (!ctx->roots) {
1754			ulist_free(ctx->refs);
1755			ctx->refs = NULL;
1756			return -ENOMEM;
1757		}
1758		roots_ulist_allocated = true;
1759	}
1760
1761	ctx->skip_inode_ref_list = true;
1762
1763	ULIST_ITER_INIT(&uiter);
1764	while (1) {
1765		struct ulist_node *node;
1766
1767		ret = find_parent_nodes(ctx, NULL);
1768		if (ret < 0 && ret != -ENOENT) {
1769			if (roots_ulist_allocated) {
1770				ulist_free(ctx->roots);
1771				ctx->roots = NULL;
1772			}
1773			break;
1774		}
1775		ret = 0;
1776		node = ulist_next(ctx->refs, &uiter);
1777		if (!node)
1778			break;
1779		ctx->bytenr = node->val;
1780		cond_resched();
1781	}
1782
1783	ulist_free(ctx->refs);
1784	ctx->refs = NULL;
1785	ctx->bytenr = orig_bytenr;
1786	ctx->skip_inode_ref_list = orig_skip_inode_ref_list;
1787
1788	return ret;
1789}
1790
1791int btrfs_find_all_roots(struct btrfs_backref_walk_ctx *ctx,
1792			 bool skip_commit_root_sem)
1793{
1794	int ret;
1795
1796	if (!ctx->trans && !skip_commit_root_sem)
1797		down_read(&ctx->fs_info->commit_root_sem);
1798	ret = btrfs_find_all_roots_safe(ctx);
1799	if (!ctx->trans && !skip_commit_root_sem)
1800		up_read(&ctx->fs_info->commit_root_sem);
1801	return ret;
1802}
1803
1804struct btrfs_backref_share_check_ctx *btrfs_alloc_backref_share_check_ctx(void)
1805{
1806	struct btrfs_backref_share_check_ctx *ctx;
1807
1808	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1809	if (!ctx)
1810		return NULL;
1811
1812	ulist_init(&ctx->refs);
1813
1814	return ctx;
1815}
1816
1817void btrfs_free_backref_share_ctx(struct btrfs_backref_share_check_ctx *ctx)
1818{
1819	if (!ctx)
1820		return;
1821
1822	ulist_release(&ctx->refs);
1823	kfree(ctx);
1824}
1825
1826/*
1827 * Check if a data extent is shared or not.
1828 *
1829 * @inode:       The inode whose extent we are checking.
1830 * @bytenr:      Logical bytenr of the extent we are checking.
1831 * @extent_gen:  Generation of the extent (file extent item) or 0 if it is
1832 *               not known.
1833 * @ctx:         A backref sharedness check context.
1834 *
1835 * btrfs_is_data_extent_shared uses the backref walking code but will short
1836 * circuit as soon as it finds a root or inode that doesn't match the
1837 * one passed in. This provides a significant performance benefit for
1838 * callers (such as fiemap) which want to know whether the extent is
1839 * shared but do not need a ref count.
1840 *
1841 * This attempts to attach to the running transaction in order to account for
1842 * delayed refs, but continues on even when no running transaction exists.
1843 *
1844 * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1845 */
1846int btrfs_is_data_extent_shared(struct btrfs_inode *inode, u64 bytenr,
1847				u64 extent_gen,
1848				struct btrfs_backref_share_check_ctx *ctx)
1849{
1850	struct btrfs_backref_walk_ctx walk_ctx = { 0 };
1851	struct btrfs_root *root = inode->root;
1852	struct btrfs_fs_info *fs_info = root->fs_info;
1853	struct btrfs_trans_handle *trans;
1854	struct ulist_iterator uiter;
1855	struct ulist_node *node;
1856	struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1857	int ret = 0;
1858	struct share_check shared = {
1859		.ctx = ctx,
1860		.root = root,
1861		.inum = btrfs_ino(inode),
1862		.data_bytenr = bytenr,
1863		.data_extent_gen = extent_gen,
1864		.share_count = 0,
1865		.self_ref_count = 0,
1866		.have_delayed_delete_refs = false,
1867	};
1868	int level;
1869	bool leaf_cached;
1870	bool leaf_is_shared;
1871
1872	for (int i = 0; i < BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE; i++) {
1873		if (ctx->prev_extents_cache[i].bytenr == bytenr)
1874			return ctx->prev_extents_cache[i].is_shared;
1875	}
1876
1877	ulist_init(&ctx->refs);
1878
1879	trans = btrfs_join_transaction_nostart(root);
1880	if (IS_ERR(trans)) {
1881		if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1882			ret = PTR_ERR(trans);
1883			goto out;
1884		}
1885		trans = NULL;
1886		down_read(&fs_info->commit_root_sem);
1887	} else {
1888		btrfs_get_tree_mod_seq(fs_info, &elem);
1889		walk_ctx.time_seq = elem.seq;
1890	}
1891
1892	ctx->use_path_cache = true;
1893
1894	/*
1895	 * We may have previously determined that the current leaf is shared.
1896	 * If it is, then we have a data extent that is shared due to a shared
1897	 * subtree (caused by snapshotting) and we don't need to check for data
1898	 * backrefs. If the leaf is not shared, then we must do backref walking
1899	 * to determine if the data extent is shared through reflinks.
1900	 */
1901	leaf_cached = lookup_backref_shared_cache(ctx, root,
1902						  ctx->curr_leaf_bytenr, 0,
1903						  &leaf_is_shared);
1904	if (leaf_cached && leaf_is_shared) {
1905		ret = 1;
1906		goto out_trans;
1907	}
1908
1909	walk_ctx.skip_inode_ref_list = true;
1910	walk_ctx.trans = trans;
1911	walk_ctx.fs_info = fs_info;
1912	walk_ctx.refs = &ctx->refs;
1913
1914	/* -1 means we are in the bytenr of the data extent. */
1915	level = -1;
1916	ULIST_ITER_INIT(&uiter);
1917	while (1) {
1918		const unsigned long prev_ref_count = ctx->refs.nnodes;
1919
1920		walk_ctx.bytenr = bytenr;
1921		ret = find_parent_nodes(&walk_ctx, &shared);
1922		if (ret == BACKREF_FOUND_SHARED ||
1923		    ret == BACKREF_FOUND_NOT_SHARED) {
1924			/* If shared must return 1, otherwise return 0. */
1925			ret = (ret == BACKREF_FOUND_SHARED) ? 1 : 0;
1926			if (level >= 0)
1927				store_backref_shared_cache(ctx, root, bytenr,
1928							   level, ret == 1);
1929			break;
1930		}
1931		if (ret < 0 && ret != -ENOENT)
1932			break;
1933		ret = 0;
1934
1935		/*
1936		 * More than one extent buffer (bytenr) may have been added to
1937		 * the ctx->refs ulist, in which case we have to check multiple
1938		 * tree paths in case the first one is not shared, so we can not
1939		 * use the path cache which is made for a single path. Multiple
1940		 * extent buffers at the current level happen when:
1941		 *
1942		 * 1) level -1, the data extent: If our data extent was not
1943		 *    directly shared (without multiple reference items), then
1944		 *    it might have a single reference item with a count > 1 for
1945		 *    the same offset, which means there are 2 (or more) file
1946		 *    extent items that point to the data extent - this happens
1947		 *    when a file extent item needs to be split and then one
1948		 *    item gets moved to another leaf due to a b+tree leaf split
1949		 *    when inserting some item. In this case the file extent
1950		 *    items may be located in different leaves and therefore
1951		 *    some of the leaves may be referenced through shared
1952		 *    subtrees while others are not. Since our extent buffer
1953		 *    cache only works for a single path (by far the most common
1954		 *    case and simpler to deal with), we can not use it if we
1955		 *    have multiple leaves (which implies multiple paths).
1956		 *
1957		 * 2) level >= 0, a tree node/leaf: We can have a mix of direct
1958		 *    and indirect references on a b+tree node/leaf, so we have
1959		 *    to check multiple paths, and the extent buffer (the
1960		 *    current bytenr) may be shared or not. One example is
1961		 *    during relocation as we may get a shared tree block ref
1962		 *    (direct ref) and a non-shared tree block ref (indirect
1963		 *    ref) for the same node/leaf.
1964		 */
1965		if ((ctx->refs.nnodes - prev_ref_count) > 1)
1966			ctx->use_path_cache = false;
1967
1968		if (level >= 0)
1969			store_backref_shared_cache(ctx, root, bytenr,
1970						   level, false);
1971		node = ulist_next(&ctx->refs, &uiter);
1972		if (!node)
1973			break;
1974		bytenr = node->val;
1975		if (ctx->use_path_cache) {
1976			bool is_shared;
1977			bool cached;
1978
1979			level++;
1980			cached = lookup_backref_shared_cache(ctx, root, bytenr,
1981							     level, &is_shared);
1982			if (cached) {
1983				ret = (is_shared ? 1 : 0);
1984				break;
1985			}
1986		}
1987		shared.share_count = 0;
1988		shared.have_delayed_delete_refs = false;
1989		cond_resched();
1990	}
1991
1992	/*
1993	 * If the path cache is disabled, then it means at some tree level we
1994	 * got multiple parents due to a mix of direct and indirect backrefs or
1995	 * multiple leaves with file extent items pointing to the same data
1996	 * extent. We have to invalidate the cache and cache only the sharedness
1997	 * result for the levels where we got only one node/reference.
1998	 */
1999	if (!ctx->use_path_cache) {
2000		int i = 0;
2001
2002		level--;
2003		if (ret >= 0 && level >= 0) {
2004			bytenr = ctx->path_cache_entries[level].bytenr;
2005			ctx->use_path_cache = true;
2006			store_backref_shared_cache(ctx, root, bytenr, level, ret);
2007			i = level + 1;
2008		}
2009
2010		for ( ; i < BTRFS_MAX_LEVEL; i++)
2011			ctx->path_cache_entries[i].bytenr = 0;
2012	}
2013
2014	/*
2015	 * Cache the sharedness result for the data extent if we know our inode
2016	 * has more than 1 file extent item that refers to the data extent.
2017	 */
2018	if (ret >= 0 && shared.self_ref_count > 1) {
2019		int slot = ctx->prev_extents_cache_slot;
2020
2021		ctx->prev_extents_cache[slot].bytenr = shared.data_bytenr;
2022		ctx->prev_extents_cache[slot].is_shared = (ret == 1);
2023
2024		slot = (slot + 1) % BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE;
2025		ctx->prev_extents_cache_slot = slot;
2026	}
2027
2028out_trans:
2029	if (trans) {
2030		btrfs_put_tree_mod_seq(fs_info, &elem);
2031		btrfs_end_transaction(trans);
2032	} else {
2033		up_read(&fs_info->commit_root_sem);
2034	}
2035out:
2036	ulist_release(&ctx->refs);
2037	ctx->prev_leaf_bytenr = ctx->curr_leaf_bytenr;
2038
2039	return ret;
2040}
2041
2042int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
2043			  u64 start_off, struct btrfs_path *path,
2044			  struct btrfs_inode_extref **ret_extref,
2045			  u64 *found_off)
2046{
2047	int ret, slot;
2048	struct btrfs_key key;
2049	struct btrfs_key found_key;
2050	struct btrfs_inode_extref *extref;
2051	const struct extent_buffer *leaf;
2052	unsigned long ptr;
2053
2054	key.objectid = inode_objectid;
2055	key.type = BTRFS_INODE_EXTREF_KEY;
2056	key.offset = start_off;
2057
2058	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2059	if (ret < 0)
2060		return ret;
2061
2062	while (1) {
2063		leaf = path->nodes[0];
2064		slot = path->slots[0];
2065		if (slot >= btrfs_header_nritems(leaf)) {
2066			/*
2067			 * If the item at offset is not found,
2068			 * btrfs_search_slot will point us to the slot
2069			 * where it should be inserted. In our case
2070			 * that will be the slot directly before the
2071			 * next INODE_REF_KEY_V2 item. In the case
2072			 * that we're pointing to the last slot in a
2073			 * leaf, we must move one leaf over.
2074			 */
2075			ret = btrfs_next_leaf(root, path);
2076			if (ret) {
2077				if (ret >= 1)
2078					ret = -ENOENT;
2079				break;
2080			}
2081			continue;
2082		}
2083
2084		btrfs_item_key_to_cpu(leaf, &found_key, slot);
2085
2086		/*
2087		 * Check that we're still looking at an extended ref key for
2088		 * this particular objectid. If we have different
2089		 * objectid or type then there are no more to be found
2090		 * in the tree and we can exit.
2091		 */
2092		ret = -ENOENT;
2093		if (found_key.objectid != inode_objectid)
2094			break;
2095		if (found_key.type != BTRFS_INODE_EXTREF_KEY)
2096			break;
2097
2098		ret = 0;
2099		ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2100		extref = (struct btrfs_inode_extref *)ptr;
2101		*ret_extref = extref;
2102		if (found_off)
2103			*found_off = found_key.offset;
2104		break;
2105	}
2106
2107	return ret;
2108}
2109
2110/*
2111 * this iterates to turn a name (from iref/extref) into a full filesystem path.
2112 * Elements of the path are separated by '/' and the path is guaranteed to be
2113 * 0-terminated. the path is only given within the current file system.
2114 * Therefore, it never starts with a '/'. the caller is responsible to provide
2115 * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
2116 * the start point of the resulting string is returned. this pointer is within
2117 * dest, normally.
2118 * in case the path buffer would overflow, the pointer is decremented further
2119 * as if output was written to the buffer, though no more output is actually
2120 * generated. that way, the caller can determine how much space would be
2121 * required for the path to fit into the buffer. in that case, the returned
2122 * value will be smaller than dest. callers must check this!
2123 */
2124char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
2125			u32 name_len, unsigned long name_off,
2126			struct extent_buffer *eb_in, u64 parent,
2127			char *dest, u32 size)
2128{
2129	int slot;
2130	u64 next_inum;
2131	int ret;
2132	s64 bytes_left = ((s64)size) - 1;
2133	struct extent_buffer *eb = eb_in;
2134	struct btrfs_key found_key;
2135	struct btrfs_inode_ref *iref;
2136
2137	if (bytes_left >= 0)
2138		dest[bytes_left] = '\0';
2139
2140	while (1) {
2141		bytes_left -= name_len;
2142		if (bytes_left >= 0)
2143			read_extent_buffer(eb, dest + bytes_left,
2144					   name_off, name_len);
2145		if (eb != eb_in) {
2146			if (!path->skip_locking)
2147				btrfs_tree_read_unlock(eb);
2148			free_extent_buffer(eb);
2149		}
2150		ret = btrfs_find_item(fs_root, path, parent, 0,
2151				BTRFS_INODE_REF_KEY, &found_key);
2152		if (ret > 0)
2153			ret = -ENOENT;
2154		if (ret)
2155			break;
2156
2157		next_inum = found_key.offset;
2158
2159		/* regular exit ahead */
2160		if (parent == next_inum)
2161			break;
2162
2163		slot = path->slots[0];
2164		eb = path->nodes[0];
2165		/* make sure we can use eb after releasing the path */
2166		if (eb != eb_in) {
2167			path->nodes[0] = NULL;
2168			path->locks[0] = 0;
2169		}
2170		btrfs_release_path(path);
2171		iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2172
2173		name_len = btrfs_inode_ref_name_len(eb, iref);
2174		name_off = (unsigned long)(iref + 1);
2175
2176		parent = next_inum;
2177		--bytes_left;
2178		if (bytes_left >= 0)
2179			dest[bytes_left] = '/';
2180	}
2181
2182	btrfs_release_path(path);
2183
2184	if (ret)
2185		return ERR_PTR(ret);
2186
2187	return dest + bytes_left;
2188}
2189
2190/*
2191 * this makes the path point to (logical EXTENT_ITEM *)
2192 * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
2193 * tree blocks and <0 on error.
2194 */
2195int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
2196			struct btrfs_path *path, struct btrfs_key *found_key,
2197			u64 *flags_ret)
2198{
2199	struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
2200	int ret;
2201	u64 flags;
2202	u64 size = 0;
2203	const struct extent_buffer *eb;
2204	struct btrfs_extent_item *ei;
2205	struct btrfs_key key;
2206
2207	key.objectid = logical;
2208	if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2209		key.type = BTRFS_METADATA_ITEM_KEY;
2210	else
2211		key.type = BTRFS_EXTENT_ITEM_KEY;
2212	key.offset = (u64)-1;
2213
2214	ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2215	if (ret < 0)
2216		return ret;
2217	if (unlikely(ret == 0)) {
2218		/*
2219		 * Key with offset -1 found, there would have to exist an extent
2220		 * item with such offset, but this is out of the valid range.
2221		 */
2222		return -EUCLEAN;
2223	}
2224
2225	ret = btrfs_previous_extent_item(extent_root, path, 0);
2226	if (ret) {
2227		if (ret > 0)
2228			ret = -ENOENT;
2229		return ret;
2230	}
2231	btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
2232	if (found_key->type == BTRFS_METADATA_ITEM_KEY)
2233		size = fs_info->nodesize;
2234	else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
2235		size = found_key->offset;
2236
2237	if (found_key->objectid > logical ||
2238	    found_key->objectid + size <= logical) {
2239		btrfs_debug(fs_info,
2240			"logical %llu is not within any extent", logical);
2241		return -ENOENT;
2242	}
2243
2244	eb = path->nodes[0];
2245
2246	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
2247	flags = btrfs_extent_flags(eb, ei);
2248
2249	btrfs_debug(fs_info,
2250		"logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
2251		 logical, logical - found_key->objectid, found_key->objectid,
2252		 found_key->offset, flags, btrfs_item_size(eb, path->slots[0]));
2253
2254	WARN_ON(!flags_ret);
2255	if (flags_ret) {
2256		if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2257			*flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
2258		else if (flags & BTRFS_EXTENT_FLAG_DATA)
2259			*flags_ret = BTRFS_EXTENT_FLAG_DATA;
2260		else
2261			BUG();
2262		return 0;
2263	}
2264
2265	return -EIO;
2266}
2267
2268/*
2269 * helper function to iterate extent inline refs. ptr must point to a 0 value
2270 * for the first call and may be modified. it is used to track state.
2271 * if more refs exist, 0 is returned and the next call to
2272 * get_extent_inline_ref must pass the modified ptr parameter to get the
2273 * next ref. after the last ref was processed, 1 is returned.
2274 * returns <0 on error
2275 */
2276static int get_extent_inline_ref(unsigned long *ptr,
2277				 const struct extent_buffer *eb,
2278				 const struct btrfs_key *key,
2279				 const struct btrfs_extent_item *ei,
2280				 u32 item_size,
2281				 struct btrfs_extent_inline_ref **out_eiref,
2282				 int *out_type)
2283{
2284	unsigned long end;
2285	u64 flags;
2286	struct btrfs_tree_block_info *info;
2287
2288	if (!*ptr) {
2289		/* first call */
2290		flags = btrfs_extent_flags(eb, ei);
2291		if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2292			if (key->type == BTRFS_METADATA_ITEM_KEY) {
2293				/* a skinny metadata extent */
2294				*out_eiref =
2295				     (struct btrfs_extent_inline_ref *)(ei + 1);
2296			} else {
2297				WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
2298				info = (struct btrfs_tree_block_info *)(ei + 1);
2299				*out_eiref =
2300				   (struct btrfs_extent_inline_ref *)(info + 1);
2301			}
2302		} else {
2303			*out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
2304		}
2305		*ptr = (unsigned long)*out_eiref;
2306		if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
2307			return -ENOENT;
2308	}
2309
2310	end = (unsigned long)ei + item_size;
2311	*out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
2312	*out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
2313						     BTRFS_REF_TYPE_ANY);
2314	if (unlikely(*out_type == BTRFS_REF_TYPE_INVALID))
2315		return -EUCLEAN;
2316
2317	*ptr += btrfs_extent_inline_ref_size(*out_type);
2318	WARN_ON(*ptr > end);
2319	if (*ptr == end)
2320		return 1; /* last */
2321
2322	return 0;
2323}
2324
2325/*
2326 * reads the tree block backref for an extent. tree level and root are returned
2327 * through out_level and out_root. ptr must point to a 0 value for the first
2328 * call and may be modified (see get_extent_inline_ref comment).
2329 * returns 0 if data was provided, 1 if there was no more data to provide or
2330 * <0 on error.
2331 */
2332int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
2333			    struct btrfs_key *key, struct btrfs_extent_item *ei,
2334			    u32 item_size, u64 *out_root, u8 *out_level)
2335{
2336	int ret;
2337	int type;
2338	struct btrfs_extent_inline_ref *eiref;
2339
2340	if (*ptr == (unsigned long)-1)
2341		return 1;
2342
2343	while (1) {
2344		ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
2345					      &eiref, &type);
2346		if (ret < 0)
2347			return ret;
2348
2349		if (type == BTRFS_TREE_BLOCK_REF_KEY ||
2350		    type == BTRFS_SHARED_BLOCK_REF_KEY)
2351			break;
2352
2353		if (ret == 1)
2354			return 1;
2355	}
2356
2357	/* we can treat both ref types equally here */
2358	*out_root = btrfs_extent_inline_ref_offset(eb, eiref);
2359
2360	if (key->type == BTRFS_EXTENT_ITEM_KEY) {
2361		struct btrfs_tree_block_info *info;
2362
2363		info = (struct btrfs_tree_block_info *)(ei + 1);
2364		*out_level = btrfs_tree_block_level(eb, info);
2365	} else {
2366		ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
2367		*out_level = (u8)key->offset;
2368	}
2369
2370	if (ret == 1)
2371		*ptr = (unsigned long)-1;
2372
2373	return 0;
2374}
2375
2376static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
2377			     struct extent_inode_elem *inode_list,
2378			     u64 root, u64 extent_item_objectid,
2379			     iterate_extent_inodes_t *iterate, void *ctx)
2380{
2381	struct extent_inode_elem *eie;
2382	int ret = 0;
2383
2384	for (eie = inode_list; eie; eie = eie->next) {
2385		btrfs_debug(fs_info,
2386			    "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
2387			    extent_item_objectid, eie->inum,
2388			    eie->offset, root);
2389		ret = iterate(eie->inum, eie->offset, eie->num_bytes, root, ctx);
2390		if (ret) {
2391			btrfs_debug(fs_info,
2392				    "stopping iteration for %llu due to ret=%d",
2393				    extent_item_objectid, ret);
2394			break;
2395		}
2396	}
2397
2398	return ret;
2399}
2400
2401/*
2402 * calls iterate() for every inode that references the extent identified by
2403 * the given parameters.
2404 * when the iterator function returns a non-zero value, iteration stops.
2405 */
2406int iterate_extent_inodes(struct btrfs_backref_walk_ctx *ctx,
2407			  bool search_commit_root,
2408			  iterate_extent_inodes_t *iterate, void *user_ctx)
2409{
2410	int ret;
2411	struct ulist *refs;
2412	struct ulist_node *ref_node;
2413	struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
2414	struct ulist_iterator ref_uiter;
2415
2416	btrfs_debug(ctx->fs_info, "resolving all inodes for extent %llu",
2417		    ctx->bytenr);
2418
2419	ASSERT(ctx->trans == NULL);
2420	ASSERT(ctx->roots == NULL);
2421
2422	if (!search_commit_root) {
2423		struct btrfs_trans_handle *trans;
2424
2425		trans = btrfs_attach_transaction(ctx->fs_info->tree_root);
2426		if (IS_ERR(trans)) {
2427			if (PTR_ERR(trans) != -ENOENT &&
2428			    PTR_ERR(trans) != -EROFS)
2429				return PTR_ERR(trans);
2430			trans = NULL;
2431		}
2432		ctx->trans = trans;
2433	}
2434
2435	if (ctx->trans) {
2436		btrfs_get_tree_mod_seq(ctx->fs_info, &seq_elem);
2437		ctx->time_seq = seq_elem.seq;
2438	} else {
2439		down_read(&ctx->fs_info->commit_root_sem);
2440	}
2441
2442	ret = btrfs_find_all_leafs(ctx);
2443	if (ret)
2444		goto out;
2445	refs = ctx->refs;
2446	ctx->refs = NULL;
2447
2448	ULIST_ITER_INIT(&ref_uiter);
2449	while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2450		const u64 leaf_bytenr = ref_node->val;
2451		struct ulist_node *root_node;
2452		struct ulist_iterator root_uiter;
2453		struct extent_inode_elem *inode_list;
2454
2455		inode_list = (struct extent_inode_elem *)(uintptr_t)ref_node->aux;
2456
2457		if (ctx->cache_lookup) {
2458			const u64 *root_ids;
2459			int root_count;
2460			bool cached;
2461
2462			cached = ctx->cache_lookup(leaf_bytenr, ctx->user_ctx,
2463						   &root_ids, &root_count);
2464			if (cached) {
2465				for (int i = 0; i < root_count; i++) {
2466					ret = iterate_leaf_refs(ctx->fs_info,
2467								inode_list,
2468								root_ids[i],
2469								leaf_bytenr,
2470								iterate,
2471								user_ctx);
2472					if (ret)
2473						break;
2474				}
2475				continue;
2476			}
2477		}
2478
2479		if (!ctx->roots) {
2480			ctx->roots = ulist_alloc(GFP_NOFS);
2481			if (!ctx->roots) {
2482				ret = -ENOMEM;
2483				break;
2484			}
2485		}
2486
2487		ctx->bytenr = leaf_bytenr;
2488		ret = btrfs_find_all_roots_safe(ctx);
2489		if (ret)
2490			break;
2491
2492		if (ctx->cache_store)
2493			ctx->cache_store(leaf_bytenr, ctx->roots, ctx->user_ctx);
2494
2495		ULIST_ITER_INIT(&root_uiter);
2496		while (!ret && (root_node = ulist_next(ctx->roots, &root_uiter))) {
2497			btrfs_debug(ctx->fs_info,
2498				    "root %llu references leaf %llu, data list %#llx",
2499				    root_node->val, ref_node->val,
2500				    ref_node->aux);
2501			ret = iterate_leaf_refs(ctx->fs_info, inode_list,
2502						root_node->val, ctx->bytenr,
2503						iterate, user_ctx);
2504		}
2505		ulist_reinit(ctx->roots);
2506	}
2507
2508	free_leaf_list(refs);
2509out:
2510	if (ctx->trans) {
2511		btrfs_put_tree_mod_seq(ctx->fs_info, &seq_elem);
2512		btrfs_end_transaction(ctx->trans);
2513		ctx->trans = NULL;
2514	} else {
2515		up_read(&ctx->fs_info->commit_root_sem);
2516	}
2517
2518	ulist_free(ctx->roots);
2519	ctx->roots = NULL;
2520
2521	if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP)
2522		ret = 0;
2523
2524	return ret;
2525}
2526
2527static int build_ino_list(u64 inum, u64 offset, u64 num_bytes, u64 root, void *ctx)
2528{
2529	struct btrfs_data_container *inodes = ctx;
2530	const size_t c = 3 * sizeof(u64);
2531
2532	if (inodes->bytes_left >= c) {
2533		inodes->bytes_left -= c;
2534		inodes->val[inodes->elem_cnt] = inum;
2535		inodes->val[inodes->elem_cnt + 1] = offset;
2536		inodes->val[inodes->elem_cnt + 2] = root;
2537		inodes->elem_cnt += 3;
2538	} else {
2539		inodes->bytes_missing += c - inodes->bytes_left;
2540		inodes->bytes_left = 0;
2541		inodes->elem_missed += 3;
2542	}
2543
2544	return 0;
2545}
2546
2547int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2548				void *ctx, bool ignore_offset)
2549{
2550	struct btrfs_backref_walk_ctx walk_ctx = { 0 };
2551	int ret;
2552	u64 flags = 0;
2553	struct btrfs_key found_key;
2554	struct btrfs_path *path;
2555
2556	path = btrfs_alloc_path();
2557	if (!path)
2558		return -ENOMEM;
2559
2560	ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2561	btrfs_free_path(path);
2562	if (ret < 0)
2563		return ret;
2564	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2565		return -EINVAL;
2566
2567	walk_ctx.bytenr = found_key.objectid;
2568	if (ignore_offset)
2569		walk_ctx.ignore_extent_item_pos = true;
2570	else
2571		walk_ctx.extent_item_pos = logical - found_key.objectid;
2572	walk_ctx.fs_info = fs_info;
2573
2574	return iterate_extent_inodes(&walk_ctx, false, build_ino_list, ctx);
2575}
2576
2577static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2578			 struct extent_buffer *eb, struct inode_fs_paths *ipath);
2579
2580static int iterate_inode_refs(u64 inum, struct inode_fs_paths *ipath)
2581{
2582	int ret = 0;
2583	int slot;
2584	u32 cur;
2585	u32 len;
2586	u32 name_len;
2587	u64 parent = 0;
2588	int found = 0;
2589	struct btrfs_root *fs_root = ipath->fs_root;
2590	struct btrfs_path *path = ipath->btrfs_path;
2591	struct extent_buffer *eb;
2592	struct btrfs_inode_ref *iref;
2593	struct btrfs_key found_key;
2594
2595	while (!ret) {
2596		ret = btrfs_find_item(fs_root, path, inum,
2597				parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2598				&found_key);
2599
2600		if (ret < 0)
2601			break;
2602		if (ret) {
2603			ret = found ? 0 : -ENOENT;
2604			break;
2605		}
2606		++found;
2607
2608		parent = found_key.offset;
2609		slot = path->slots[0];
2610		eb = btrfs_clone_extent_buffer(path->nodes[0]);
2611		if (!eb) {
2612			ret = -ENOMEM;
2613			break;
2614		}
2615		btrfs_release_path(path);
2616
2617		iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2618
2619		for (cur = 0; cur < btrfs_item_size(eb, slot); cur += len) {
2620			name_len = btrfs_inode_ref_name_len(eb, iref);
2621			/* path must be released before calling iterate()! */
2622			btrfs_debug(fs_root->fs_info,
2623				"following ref at offset %u for inode %llu in tree %llu",
2624				cur, found_key.objectid,
2625				btrfs_root_id(fs_root));
2626			ret = inode_to_path(parent, name_len,
2627				      (unsigned long)(iref + 1), eb, ipath);
2628			if (ret)
2629				break;
2630			len = sizeof(*iref) + name_len;
2631			iref = (struct btrfs_inode_ref *)((char *)iref + len);
2632		}
2633		free_extent_buffer(eb);
2634	}
2635
2636	btrfs_release_path(path);
2637
2638	return ret;
2639}
2640
2641static int iterate_inode_extrefs(u64 inum, struct inode_fs_paths *ipath)
2642{
2643	int ret;
2644	int slot;
2645	u64 offset = 0;
2646	u64 parent;
2647	int found = 0;
2648	struct btrfs_root *fs_root = ipath->fs_root;
2649	struct btrfs_path *path = ipath->btrfs_path;
2650	struct extent_buffer *eb;
2651	struct btrfs_inode_extref *extref;
2652	u32 item_size;
2653	u32 cur_offset;
2654	unsigned long ptr;
2655
2656	while (1) {
2657		ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2658					    &offset);
2659		if (ret < 0)
2660			break;
2661		if (ret) {
2662			ret = found ? 0 : -ENOENT;
2663			break;
2664		}
2665		++found;
2666
2667		slot = path->slots[0];
2668		eb = btrfs_clone_extent_buffer(path->nodes[0]);
2669		if (!eb) {
2670			ret = -ENOMEM;
2671			break;
2672		}
2673		btrfs_release_path(path);
2674
2675		item_size = btrfs_item_size(eb, slot);
2676		ptr = btrfs_item_ptr_offset(eb, slot);
2677		cur_offset = 0;
2678
2679		while (cur_offset < item_size) {
2680			u32 name_len;
2681
2682			extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2683			parent = btrfs_inode_extref_parent(eb, extref);
2684			name_len = btrfs_inode_extref_name_len(eb, extref);
2685			ret = inode_to_path(parent, name_len,
2686				      (unsigned long)&extref->name, eb, ipath);
2687			if (ret)
2688				break;
2689
2690			cur_offset += btrfs_inode_extref_name_len(eb, extref);
2691			cur_offset += sizeof(*extref);
2692		}
2693		free_extent_buffer(eb);
2694
2695		offset++;
2696	}
2697
2698	btrfs_release_path(path);
2699
2700	return ret;
2701}
2702
2703/*
2704 * returns 0 if the path could be dumped (probably truncated)
2705 * returns <0 in case of an error
2706 */
2707static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2708			 struct extent_buffer *eb, struct inode_fs_paths *ipath)
2709{
2710	char *fspath;
2711	char *fspath_min;
2712	int i = ipath->fspath->elem_cnt;
2713	const int s_ptr = sizeof(char *);
2714	u32 bytes_left;
2715
2716	bytes_left = ipath->fspath->bytes_left > s_ptr ?
2717					ipath->fspath->bytes_left - s_ptr : 0;
2718
2719	fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2720	fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2721				   name_off, eb, inum, fspath_min, bytes_left);
2722	if (IS_ERR(fspath))
2723		return PTR_ERR(fspath);
2724
2725	if (fspath > fspath_min) {
2726		ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2727		++ipath->fspath->elem_cnt;
2728		ipath->fspath->bytes_left = fspath - fspath_min;
2729	} else {
2730		++ipath->fspath->elem_missed;
2731		ipath->fspath->bytes_missing += fspath_min - fspath;
2732		ipath->fspath->bytes_left = 0;
2733	}
2734
2735	return 0;
2736}
2737
2738/*
2739 * this dumps all file system paths to the inode into the ipath struct, provided
2740 * is has been created large enough. each path is zero-terminated and accessed
2741 * from ipath->fspath->val[i].
2742 * when it returns, there are ipath->fspath->elem_cnt number of paths available
2743 * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2744 * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2745 * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2746 * have been needed to return all paths.
2747 */
2748int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2749{
2750	int ret;
2751	int found_refs = 0;
2752
2753	ret = iterate_inode_refs(inum, ipath);
2754	if (!ret)
2755		++found_refs;
2756	else if (ret != -ENOENT)
2757		return ret;
2758
2759	ret = iterate_inode_extrefs(inum, ipath);
2760	if (ret == -ENOENT && found_refs)
2761		return 0;
2762
2763	return ret;
2764}
2765
2766struct btrfs_data_container *init_data_container(u32 total_bytes)
2767{
2768	struct btrfs_data_container *data;
2769	size_t alloc_bytes;
2770
2771	alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2772	data = kvzalloc(alloc_bytes, GFP_KERNEL);
2773	if (!data)
2774		return ERR_PTR(-ENOMEM);
2775
2776	if (total_bytes >= sizeof(*data))
2777		data->bytes_left = total_bytes - sizeof(*data);
2778	else
2779		data->bytes_missing = sizeof(*data) - total_bytes;
2780
2781	return data;
2782}
2783
2784/*
2785 * allocates space to return multiple file system paths for an inode.
2786 * total_bytes to allocate are passed, note that space usable for actual path
2787 * information will be total_bytes - sizeof(struct inode_fs_paths).
2788 * the returned pointer must be freed with __free_inode_fs_paths() in the end.
2789 */
2790struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2791					struct btrfs_path *path)
2792{
2793	struct inode_fs_paths *ifp;
2794	struct btrfs_data_container *fspath;
2795
2796	fspath = init_data_container(total_bytes);
2797	if (IS_ERR(fspath))
2798		return ERR_CAST(fspath);
2799
2800	ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2801	if (!ifp) {
2802		kvfree(fspath);
2803		return ERR_PTR(-ENOMEM);
2804	}
2805
2806	ifp->btrfs_path = path;
2807	ifp->fspath = fspath;
2808	ifp->fs_root = fs_root;
2809
2810	return ifp;
2811}
2812
2813struct btrfs_backref_iter *btrfs_backref_iter_alloc(struct btrfs_fs_info *fs_info)
2814{
2815	struct btrfs_backref_iter *ret;
2816
2817	ret = kzalloc(sizeof(*ret), GFP_NOFS);
2818	if (!ret)
2819		return NULL;
2820
2821	ret->path = btrfs_alloc_path();
2822	if (!ret->path) {
2823		kfree(ret);
2824		return NULL;
2825	}
2826
2827	/* Current backref iterator only supports iteration in commit root */
2828	ret->path->search_commit_root = true;
2829	ret->path->skip_locking = true;
2830	ret->fs_info = fs_info;
2831
2832	return ret;
2833}
2834
2835static void btrfs_backref_iter_release(struct btrfs_backref_iter *iter)
2836{
2837	iter->bytenr = 0;
2838	iter->item_ptr = 0;
2839	iter->cur_ptr = 0;
2840	iter->end_ptr = 0;
2841	btrfs_release_path(iter->path);
2842	memset(&iter->cur_key, 0, sizeof(iter->cur_key));
2843}
2844
2845int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2846{
2847	struct btrfs_fs_info *fs_info = iter->fs_info;
2848	struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2849	struct btrfs_path *path = iter->path;
2850	struct btrfs_extent_item *ei;
2851	struct btrfs_key key;
2852	int ret;
2853
2854	key.objectid = bytenr;
2855	key.type = BTRFS_METADATA_ITEM_KEY;
2856	key.offset = (u64)-1;
2857	iter->bytenr = bytenr;
2858
2859	ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2860	if (ret < 0)
2861		return ret;
2862	if (unlikely(ret == 0)) {
2863		/*
2864		 * Key with offset -1 found, there would have to exist an extent
2865		 * item with such offset, but this is out of the valid range.
2866		 */
2867		ret = -EUCLEAN;
2868		goto release;
2869	}
2870	if (unlikely(path->slots[0] == 0)) {
2871		DEBUG_WARN();
2872		ret = -EUCLEAN;
2873		goto release;
2874	}
2875	path->slots[0]--;
2876
2877	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2878	if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2879	     key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2880		ret = -ENOENT;
2881		goto release;
2882	}
2883	memcpy(&iter->cur_key, &key, sizeof(key));
2884	iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2885						    path->slots[0]);
2886	iter->end_ptr = (u32)(iter->item_ptr +
2887			btrfs_item_size(path->nodes[0], path->slots[0]));
2888	ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2889			    struct btrfs_extent_item);
2890
2891	/*
2892	 * Only support iteration on tree backref yet.
2893	 *
2894	 * This is an extra precaution for non skinny-metadata, where
2895	 * EXTENT_ITEM is also used for tree blocks, that we can only use
2896	 * extent flags to determine if it's a tree block.
2897	 */
2898	if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2899		ret = -ENOTSUPP;
2900		goto release;
2901	}
2902	iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2903
2904	/* If there is no inline backref, go search for keyed backref */
2905	if (iter->cur_ptr >= iter->end_ptr) {
2906		ret = btrfs_next_item(extent_root, path);
2907
2908		/* No inline nor keyed ref */
2909		if (ret > 0) {
2910			ret = -ENOENT;
2911			goto release;
2912		}
2913		if (ret < 0)
2914			goto release;
2915
2916		btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2917				path->slots[0]);
2918		if (iter->cur_key.objectid != bytenr ||
2919		    (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2920		     iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2921			ret = -ENOENT;
2922			goto release;
2923		}
2924		iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2925							   path->slots[0]);
2926		iter->item_ptr = iter->cur_ptr;
2927		iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size(
2928				      path->nodes[0], path->slots[0]));
2929	}
2930
2931	return 0;
2932release:
2933	btrfs_backref_iter_release(iter);
2934	return ret;
2935}
2936
2937static bool btrfs_backref_iter_is_inline_ref(struct btrfs_backref_iter *iter)
2938{
2939	if (iter->cur_key.type == BTRFS_EXTENT_ITEM_KEY ||
2940	    iter->cur_key.type == BTRFS_METADATA_ITEM_KEY)
2941		return true;
2942	return false;
2943}
2944
2945/*
2946 * Go to the next backref item of current bytenr, can be either inlined or
2947 * keyed.
2948 *
2949 * Caller needs to check whether it's inline ref or not by iter->cur_key.
2950 *
2951 * Return 0 if we get next backref without problem.
2952 * Return >0 if there is no extra backref for this bytenr.
2953 * Return <0 if there is something wrong happened.
2954 */
2955int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2956{
2957	struct extent_buffer *eb = iter->path->nodes[0];
2958	struct btrfs_root *extent_root;
2959	struct btrfs_path *path = iter->path;
2960	struct btrfs_extent_inline_ref *iref;
2961	int ret;
2962	u32 size;
2963
2964	if (btrfs_backref_iter_is_inline_ref(iter)) {
2965		/* We're still inside the inline refs */
2966		ASSERT(iter->cur_ptr < iter->end_ptr);
2967
2968		if (btrfs_backref_has_tree_block_info(iter)) {
2969			/* First tree block info */
2970			size = sizeof(struct btrfs_tree_block_info);
2971		} else {
2972			/* Use inline ref type to determine the size */
2973			int type;
2974
2975			iref = (struct btrfs_extent_inline_ref *)
2976				((unsigned long)iter->cur_ptr);
2977			type = btrfs_extent_inline_ref_type(eb, iref);
2978
2979			size = btrfs_extent_inline_ref_size(type);
2980		}
2981		iter->cur_ptr += size;
2982		if (iter->cur_ptr < iter->end_ptr)
2983			return 0;
2984
2985		/* All inline items iterated, fall through */
2986	}
2987
2988	/* We're at keyed items, there is no inline item, go to the next one */
2989	extent_root = btrfs_extent_root(iter->fs_info, iter->bytenr);
2990	ret = btrfs_next_item(extent_root, iter->path);
2991	if (ret)
2992		return ret;
2993
2994	btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2995	if (iter->cur_key.objectid != iter->bytenr ||
2996	    (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2997	     iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2998		return 1;
2999	iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
3000					path->slots[0]);
3001	iter->cur_ptr = iter->item_ptr;
3002	iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size(path->nodes[0],
3003						path->slots[0]);
3004	return 0;
3005}
3006
3007void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
3008			      struct btrfs_backref_cache *cache, bool is_reloc)
3009{
3010	int i;
3011
3012	cache->rb_root = RB_ROOT;
3013	for (i = 0; i < BTRFS_MAX_LEVEL; i++)
3014		INIT_LIST_HEAD(&cache->pending[i]);
3015	INIT_LIST_HEAD(&cache->pending_edge);
3016	INIT_LIST_HEAD(&cache->useless_node);
3017	cache->fs_info = fs_info;
3018	cache->is_reloc = is_reloc;
3019}
3020
3021struct btrfs_backref_node *btrfs_backref_alloc_node(
3022		struct btrfs_backref_cache *cache, u64 bytenr, int level)
3023{
3024	struct btrfs_backref_node *node;
3025
3026	ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
3027	node = kzalloc(sizeof(*node), GFP_NOFS);
3028	if (!node)
3029		return node;
3030
3031	INIT_LIST_HEAD(&node->list);
3032	INIT_LIST_HEAD(&node->upper);
3033	INIT_LIST_HEAD(&node->lower);
3034	RB_CLEAR_NODE(&node->rb_node);
3035	cache->nr_nodes++;
3036	node->level = level;
3037	node->bytenr = bytenr;
3038
3039	return node;
3040}
3041
3042void btrfs_backref_free_node(struct btrfs_backref_cache *cache,
3043			     struct btrfs_backref_node *node)
3044{
3045	if (node) {
3046		ASSERT(list_empty(&node->list));
3047		ASSERT(list_empty(&node->lower));
3048		ASSERT(node->eb == NULL);
3049		cache->nr_nodes--;
3050		btrfs_put_root(node->root);
3051		kfree(node);
3052	}
3053}
3054
3055struct btrfs_backref_edge *btrfs_backref_alloc_edge(
3056		struct btrfs_backref_cache *cache)
3057{
3058	struct btrfs_backref_edge *edge;
3059
3060	edge = kzalloc(sizeof(*edge), GFP_NOFS);
3061	if (edge)
3062		cache->nr_edges++;
3063	return edge;
3064}
3065
3066void btrfs_backref_free_edge(struct btrfs_backref_cache *cache,
3067			     struct btrfs_backref_edge *edge)
3068{
3069	if (edge) {
3070		cache->nr_edges--;
3071		kfree(edge);
3072	}
3073}
3074
3075void btrfs_backref_unlock_node_buffer(struct btrfs_backref_node *node)
3076{
3077	if (node->locked) {
3078		btrfs_tree_unlock(node->eb);
3079		node->locked = 0;
3080	}
3081}
3082
3083void btrfs_backref_drop_node_buffer(struct btrfs_backref_node *node)
3084{
3085	if (node->eb) {
3086		btrfs_backref_unlock_node_buffer(node);
3087		free_extent_buffer(node->eb);
3088		node->eb = NULL;
3089	}
3090}
3091
3092/*
3093 * Drop the backref node from cache without cleaning up its children
3094 * edges.
3095 *
3096 * This can only be called on node without parent edges.
3097 * The children edges are still kept as is.
3098 */
3099void btrfs_backref_drop_node(struct btrfs_backref_cache *tree,
3100			     struct btrfs_backref_node *node)
3101{
3102	ASSERT(list_empty(&node->upper));
3103
3104	btrfs_backref_drop_node_buffer(node);
3105	list_del_init(&node->list);
3106	list_del_init(&node->lower);
3107	if (!RB_EMPTY_NODE(&node->rb_node))
3108		rb_erase(&node->rb_node, &tree->rb_root);
3109	btrfs_backref_free_node(tree, node);
3110}
3111
3112/*
3113 * Drop the backref node from cache, also cleaning up all its
3114 * upper edges and any uncached nodes in the path.
3115 *
3116 * This cleanup happens bottom up, thus the node should either
3117 * be the lowest node in the cache or a detached node.
3118 */
3119void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
3120				struct btrfs_backref_node *node)
3121{
3122	struct btrfs_backref_edge *edge;
3123
3124	if (!node)
3125		return;
3126
3127	while (!list_empty(&node->upper)) {
3128		edge = list_first_entry(&node->upper, struct btrfs_backref_edge,
3129					list[LOWER]);
3130		list_del(&edge->list[LOWER]);
3131		list_del(&edge->list[UPPER]);
3132		btrfs_backref_free_edge(cache, edge);
3133	}
3134
3135	btrfs_backref_drop_node(cache, node);
3136}
3137
3138/*
3139 * Release all nodes/edges from current cache
3140 */
3141void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
3142{
3143	struct btrfs_backref_node *node;
3144
3145	while ((node = rb_entry_safe(rb_first(&cache->rb_root),
3146				     struct btrfs_backref_node, rb_node)))
3147		btrfs_backref_cleanup_node(cache, node);
3148
3149	ASSERT(list_empty(&cache->pending_edge));
3150	ASSERT(list_empty(&cache->useless_node));
3151	ASSERT(!cache->nr_nodes);
3152	ASSERT(!cache->nr_edges);
3153}
3154
3155static void btrfs_backref_link_edge(struct btrfs_backref_edge *edge,
3156				    struct btrfs_backref_node *lower,
3157				    struct btrfs_backref_node *upper)
3158{
3159	ASSERT(upper && lower && upper->level == lower->level + 1);
3160	edge->node[LOWER] = lower;
3161	edge->node[UPPER] = upper;
3162	list_add_tail(&edge->list[LOWER], &lower->upper);
3163}
3164/*
3165 * Handle direct tree backref
3166 *
3167 * Direct tree backref means, the backref item shows its parent bytenr
3168 * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
3169 *
3170 * @ref_key:	The converted backref key.
3171 *		For keyed backref, it's the item key.
3172 *		For inlined backref, objectid is the bytenr,
3173 *		type is btrfs_inline_ref_type, offset is
3174 *		btrfs_inline_ref_offset.
3175 */
3176static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
3177				      struct btrfs_key *ref_key,
3178				      struct btrfs_backref_node *cur)
3179{
3180	struct btrfs_backref_edge *edge;
3181	struct btrfs_backref_node *upper;
3182	struct rb_node *rb_node;
3183
3184	ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
3185
3186	/* Only reloc root uses backref pointing to itself */
3187	if (ref_key->objectid == ref_key->offset) {
3188		struct btrfs_root *root;
3189
3190		cur->is_reloc_root = 1;
3191		/* Only reloc backref cache cares about a specific root */
3192		if (cache->is_reloc) {
3193			root = find_reloc_root(cache->fs_info, cur->bytenr);
3194			if (!root)
3195				return -ENOENT;
3196			cur->root = root;
3197		} else {
3198			/*
3199			 * For generic purpose backref cache, reloc root node
3200			 * is useless.
3201			 */
3202			list_add(&cur->list, &cache->useless_node);
3203		}
3204		return 0;
3205	}
3206
3207	edge = btrfs_backref_alloc_edge(cache);
3208	if (!edge)
3209		return -ENOMEM;
3210
3211	rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
3212	if (!rb_node) {
3213		/* Parent node not yet cached */
3214		upper = btrfs_backref_alloc_node(cache, ref_key->offset,
3215					   cur->level + 1);
3216		if (!upper) {
3217			btrfs_backref_free_edge(cache, edge);
3218			return -ENOMEM;
3219		}
3220
3221		/*
3222		 *  Backrefs for the upper level block isn't cached, add the
3223		 *  block to pending list
3224		 */
3225		list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3226	} else {
3227		/* Parent node already cached */
3228		upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
3229		ASSERT(upper->checked);
3230		INIT_LIST_HEAD(&edge->list[UPPER]);
3231	}
3232	btrfs_backref_link_edge(edge, cur, upper);
3233	return 0;
3234}
3235
3236/*
3237 * Handle indirect tree backref
3238 *
3239 * Indirect tree backref means, we only know which tree the node belongs to.
3240 * We still need to do a tree search to find out the parents. This is for
3241 * TREE_BLOCK_REF backref (keyed or inlined).
3242 *
3243 * @trans:	Transaction handle.
3244 * @ref_key:	The same as @ref_key in  handle_direct_tree_backref()
3245 * @tree_key:	The first key of this tree block.
3246 * @path:	A clean (released) path, to avoid allocating path every time
3247 *		the function get called.
3248 */
3249static int handle_indirect_tree_backref(struct btrfs_trans_handle *trans,
3250					struct btrfs_backref_cache *cache,
3251					struct btrfs_path *path,
3252					struct btrfs_key *ref_key,
3253					struct btrfs_key *tree_key,
3254					struct btrfs_backref_node *cur)
3255{
3256	struct btrfs_fs_info *fs_info = cache->fs_info;
3257	struct btrfs_backref_node *upper;
3258	struct btrfs_backref_node *lower;
3259	struct btrfs_backref_edge *edge;
3260	struct extent_buffer *eb;
3261	struct btrfs_root *root;
3262	struct rb_node *rb_node;
3263	int level;
3264	bool need_check = true;
3265	int ret;
3266
3267	root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
3268	if (IS_ERR(root))
3269		return PTR_ERR(root);
3270
3271	/* We shouldn't be using backref cache for non-shareable roots. */
3272	if (unlikely(!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))) {
3273		btrfs_put_root(root);
3274		return -EUCLEAN;
3275	}
3276
3277	if (btrfs_root_level(&root->root_item) == cur->level) {
3278		/* Tree root */
3279		ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
3280		/*
3281		 * For reloc backref cache, we may ignore reloc root.  But for
3282		 * general purpose backref cache, we can't rely on
3283		 * btrfs_should_ignore_reloc_root() as it may conflict with
3284		 * current running relocation and lead to missing root.
3285		 *
3286		 * For general purpose backref cache, reloc root detection is
3287		 * completely relying on direct backref (key->offset is parent
3288		 * bytenr), thus only do such check for reloc cache.
3289		 */
3290		if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
3291			btrfs_put_root(root);
3292			list_add(&cur->list, &cache->useless_node);
3293		} else {
3294			cur->root = root;
3295		}
3296		return 0;
3297	}
3298
3299	level = cur->level + 1;
3300
3301	/* Search the tree to find parent blocks referring to the block */
3302	path->search_commit_root = true;
3303	path->skip_locking = true;
3304	path->lowest_level = level;
3305	ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
3306	path->lowest_level = 0;
3307	if (ret < 0) {
3308		btrfs_put_root(root);
3309		return ret;
3310	}
3311	if (ret > 0 && path->slots[level] > 0)
3312		path->slots[level]--;
3313
3314	eb = path->nodes[level];
3315	if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
3316		btrfs_err(fs_info,
3317"couldn't find block (%llu) (level %d) in tree (%llu) with key " BTRFS_KEY_FMT,
3318			  cur->bytenr, level - 1, btrfs_root_id(root),
3319			  BTRFS_KEY_FMT_VALUE(tree_key));
3320		btrfs_put_root(root);
3321		ret = -ENOENT;
3322		goto out;
3323	}
3324	lower = cur;
3325
3326	/* Add all nodes and edges in the path */
3327	for (; level < BTRFS_MAX_LEVEL; level++) {
3328		if (!path->nodes[level]) {
3329			ASSERT(btrfs_root_bytenr(&root->root_item) ==
3330			       lower->bytenr);
3331			/* Same as previous should_ignore_reloc_root() call */
3332			if (btrfs_should_ignore_reloc_root(root) &&
3333			    cache->is_reloc) {
3334				btrfs_put_root(root);
3335				list_add(&lower->list, &cache->useless_node);
3336			} else {
3337				lower->root = root;
3338			}
3339			break;
3340		}
3341
3342		edge = btrfs_backref_alloc_edge(cache);
3343		if (!edge) {
3344			btrfs_put_root(root);
3345			ret = -ENOMEM;
3346			goto out;
3347		}
3348
3349		eb = path->nodes[level];
3350		rb_node = rb_simple_search(&cache->rb_root, eb->start);
3351		if (!rb_node) {
3352			upper = btrfs_backref_alloc_node(cache, eb->start,
3353							 lower->level + 1);
3354			if (!upper) {
3355				btrfs_put_root(root);
3356				btrfs_backref_free_edge(cache, edge);
3357				ret = -ENOMEM;
3358				goto out;
3359			}
3360			upper->owner = btrfs_header_owner(eb);
3361
3362			/* We shouldn't be using backref cache for non shareable roots. */
3363			if (unlikely(!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))) {
3364				btrfs_put_root(root);
3365				btrfs_backref_free_edge(cache, edge);
3366				btrfs_backref_free_node(cache, upper);
3367				ret = -EUCLEAN;
3368				goto out;
3369			}
3370
3371			/*
3372			 * If we know the block isn't shared we can avoid
3373			 * checking its backrefs.
3374			 */
3375			if (btrfs_block_can_be_shared(trans, root, eb))
3376				upper->checked = 0;
3377			else
3378				upper->checked = 1;
3379
3380			/*
3381			 * Add the block to pending list if we need to check its
3382			 * backrefs, we only do this once while walking up a
3383			 * tree as we will catch anything else later on.
3384			 */
3385			if (!upper->checked && need_check) {
3386				need_check = false;
3387				list_add_tail(&edge->list[UPPER],
3388					      &cache->pending_edge);
3389			} else {
3390				if (upper->checked)
3391					need_check = true;
3392				INIT_LIST_HEAD(&edge->list[UPPER]);
3393			}
3394		} else {
3395			upper = rb_entry(rb_node, struct btrfs_backref_node,
3396					 rb_node);
3397			ASSERT(upper->checked);
3398			INIT_LIST_HEAD(&edge->list[UPPER]);
3399			if (!upper->owner)
3400				upper->owner = btrfs_header_owner(eb);
3401		}
3402		btrfs_backref_link_edge(edge, lower, upper);
3403
3404		if (rb_node) {
3405			btrfs_put_root(root);
3406			break;
3407		}
3408		lower = upper;
3409		upper = NULL;
3410	}
3411out:
3412	btrfs_release_path(path);
3413	return ret;
3414}
3415
3416/*
3417 * Add backref node @cur into @cache.
3418 *
3419 * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
3420 *	 links aren't yet bi-directional. Needs to finish such links.
3421 *	 Use btrfs_backref_finish_upper_links() to finish such linkage.
3422 *
3423 * @trans:	Transaction handle.
3424 * @path:	Released path for indirect tree backref lookup
3425 * @iter:	Released backref iter for extent tree search
3426 * @node_key:	The first key of the tree block
3427 */
3428int btrfs_backref_add_tree_node(struct btrfs_trans_handle *trans,
3429				struct btrfs_backref_cache *cache,
3430				struct btrfs_path *path,
3431				struct btrfs_backref_iter *iter,
3432				struct btrfs_key *node_key,
3433				struct btrfs_backref_node *cur)
3434{
3435	struct btrfs_backref_edge *edge;
3436	struct btrfs_backref_node *exist;
3437	int ret;
3438
3439	ret = btrfs_backref_iter_start(iter, cur->bytenr);
3440	if (ret < 0)
3441		return ret;
3442	/*
3443	 * We skip the first btrfs_tree_block_info, as we don't use the key
3444	 * stored in it, but fetch it from the tree block
3445	 */
3446	if (btrfs_backref_has_tree_block_info(iter)) {
3447		ret = btrfs_backref_iter_next(iter);
3448		if (ret < 0)
3449			goto out;
3450		/* No extra backref? This means the tree block is corrupted */
3451		if (unlikely(ret > 0)) {
3452			ret = -EUCLEAN;
3453			goto out;
3454		}
3455	}
3456	WARN_ON(cur->checked);
3457	if (!list_empty(&cur->upper)) {
3458		/*
3459		 * The backref was added previously when processing backref of
3460		 * type BTRFS_TREE_BLOCK_REF_KEY
3461		 */
3462		ASSERT(list_is_singular(&cur->upper));
3463		edge = list_first_entry(&cur->upper, struct btrfs_backref_edge,
3464					list[LOWER]);
3465		ASSERT(list_empty(&edge->list[UPPER]));
3466		exist = edge->node[UPPER];
3467		/*
3468		 * Add the upper level block to pending list if we need check
3469		 * its backrefs
3470		 */
3471		if (!exist->checked)
3472			list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3473	} else {
3474		exist = NULL;
3475	}
3476
3477	for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
3478		struct extent_buffer *eb;
3479		struct btrfs_key key;
3480		int type;
3481
3482		cond_resched();
3483		eb = iter->path->nodes[0];
3484
3485		key.objectid = iter->bytenr;
3486		if (btrfs_backref_iter_is_inline_ref(iter)) {
3487			struct btrfs_extent_inline_ref *iref;
3488
3489			/* Update key for inline backref */
3490			iref = (struct btrfs_extent_inline_ref *)
3491				((unsigned long)iter->cur_ptr);
3492			type = btrfs_get_extent_inline_ref_type(eb, iref,
3493							BTRFS_REF_TYPE_BLOCK);
3494			if (unlikely(type == BTRFS_REF_TYPE_INVALID)) {
3495				ret = -EUCLEAN;
3496				goto out;
3497			}
3498			key.type = type;
3499			key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3500		} else {
3501			key.type = iter->cur_key.type;
3502			key.offset = iter->cur_key.offset;
3503		}
3504
3505		/*
3506		 * Parent node found and matches current inline ref, no need to
3507		 * rebuild this node for this inline ref
3508		 */
3509		if (exist &&
3510		    ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
3511		      exist->owner == key.offset) ||
3512		     (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
3513		      exist->bytenr == key.offset))) {
3514			exist = NULL;
3515			continue;
3516		}
3517
3518		/* SHARED_BLOCK_REF means key.offset is the parent bytenr */
3519		if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
3520			ret = handle_direct_tree_backref(cache, &key, cur);
3521			if (ret < 0)
3522				goto out;
3523		} else if (key.type == BTRFS_TREE_BLOCK_REF_KEY) {
3524			/*
3525			 * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref
3526			 * offset means the root objectid. We need to search
3527			 * the tree to get its parent bytenr.
3528			 */
3529			ret = handle_indirect_tree_backref(trans, cache, path,
3530							   &key, node_key, cur);
3531			if (ret < 0)
3532				goto out;
3533		}
3534		/*
3535		 * Unrecognized tree backref items (if it can pass tree-checker)
3536		 * would be ignored.
3537		 */
3538	}
3539	ret = 0;
3540	cur->checked = 1;
3541	WARN_ON(exist);
3542out:
3543	btrfs_backref_iter_release(iter);
3544	return ret;
3545}
3546
3547/*
3548 * Finish the upwards linkage created by btrfs_backref_add_tree_node()
3549 */
3550int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
3551				     struct btrfs_backref_node *start)
3552{
3553	struct list_head *useless_node = &cache->useless_node;
3554	struct btrfs_backref_edge *edge;
3555	struct rb_node *rb_node;
3556	LIST_HEAD(pending_edge);
3557
3558	ASSERT(start->checked);
3559
3560	rb_node = rb_simple_insert(&cache->rb_root, &start->simple_node);
3561	if (rb_node)
3562		btrfs_backref_panic(cache->fs_info, start->bytenr, -EEXIST);
3563
3564	/*
3565	 * Use breadth first search to iterate all related edges.
3566	 *
3567	 * The starting points are all the edges of this node
3568	 */
3569	list_for_each_entry(edge, &start->upper, list[LOWER])
3570		list_add_tail(&edge->list[UPPER], &pending_edge);
3571
3572	while (!list_empty(&pending_edge)) {
3573		struct btrfs_backref_node *upper;
3574		struct btrfs_backref_node *lower;
3575
3576		edge = list_first_entry(&pending_edge,
3577				struct btrfs_backref_edge, list[UPPER]);
3578		list_del_init(&edge->list[UPPER]);
3579		upper = edge->node[UPPER];
3580		lower = edge->node[LOWER];
3581
3582		/* Parent is detached, no need to keep any edges */
3583		if (upper->detached) {
3584			list_del(&edge->list[LOWER]);
3585			btrfs_backref_free_edge(cache, edge);
3586
3587			/* Lower node is orphan, queue for cleanup */
3588			if (list_empty(&lower->upper))
3589				list_add(&lower->list, useless_node);
3590			continue;
3591		}
3592
3593		/*
3594		 * All new nodes added in current build_backref_tree() haven't
3595		 * been linked to the cache rb tree.
3596		 * So if we have upper->rb_node populated, this means a cache
3597		 * hit. We only need to link the edge, as @upper and all its
3598		 * parents have already been linked.
3599		 */
3600		if (!RB_EMPTY_NODE(&upper->rb_node)) {
3601			list_add_tail(&edge->list[UPPER], &upper->lower);
3602			continue;
3603		}
3604
3605		/* Sanity check, we shouldn't have any unchecked nodes */
3606		if (unlikely(!upper->checked)) {
3607			DEBUG_WARN("we should not have any unchecked nodes");
3608			return -EUCLEAN;
3609		}
3610
3611		rb_node = rb_simple_insert(&cache->rb_root, &upper->simple_node);
3612		if (unlikely(rb_node)) {
3613			btrfs_backref_panic(cache->fs_info, upper->bytenr, -EEXIST);
3614			return -EUCLEAN;
3615		}
3616
3617		list_add_tail(&edge->list[UPPER], &upper->lower);
3618
3619		/*
3620		 * Also queue all the parent edges of this uncached node
3621		 * to finish the upper linkage
3622		 */
3623		list_for_each_entry(edge, &upper->upper, list[LOWER])
3624			list_add_tail(&edge->list[UPPER], &pending_edge);
3625	}
3626	return 0;
3627}
3628
3629void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3630				 struct btrfs_backref_node *node)
3631{
3632	struct btrfs_backref_node *lower;
3633	struct btrfs_backref_node *upper;
3634	struct btrfs_backref_edge *edge;
3635
3636	while (!list_empty(&cache->useless_node)) {
3637		lower = list_first_entry(&cache->useless_node,
3638				   struct btrfs_backref_node, list);
3639		list_del_init(&lower->list);
3640	}
3641	while (!list_empty(&cache->pending_edge)) {
3642		edge = list_first_entry(&cache->pending_edge,
3643				struct btrfs_backref_edge, list[UPPER]);
3644		list_del(&edge->list[UPPER]);
3645		list_del(&edge->list[LOWER]);
3646		lower = edge->node[LOWER];
3647		upper = edge->node[UPPER];
3648		btrfs_backref_free_edge(cache, edge);
3649
3650		/*
3651		 * Lower is no longer linked to any upper backref nodes and
3652		 * isn't in the cache, we can free it ourselves.
3653		 */
3654		if (list_empty(&lower->upper) &&
3655		    RB_EMPTY_NODE(&lower->rb_node))
3656			list_add(&lower->list, &cache->useless_node);
3657
3658		if (!RB_EMPTY_NODE(&upper->rb_node))
3659			continue;
3660
3661		/* Add this guy's upper edges to the list to process */
3662		list_for_each_entry(edge, &upper->upper, list[LOWER])
3663			list_add_tail(&edge->list[UPPER],
3664				      &cache->pending_edge);
3665		if (list_empty(&upper->upper))
3666			list_add(&upper->list, &cache->useless_node);
3667	}
3668
3669	while (!list_empty(&cache->useless_node)) {
3670		lower = list_first_entry(&cache->useless_node,
3671				   struct btrfs_backref_node, list);
3672		list_del_init(&lower->list);
3673		if (lower == node)
3674			node = NULL;
3675		btrfs_backref_drop_node(cache, lower);
3676	}
3677
3678	btrfs_backref_cleanup_node(cache, node);
3679	ASSERT(list_empty(&cache->useless_node) &&
3680	       list_empty(&cache->pending_edge));
3681}