1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2008 Oracle.  All rights reserved.
   4 */
   5
   6#include <linux/sched.h>
   7#include <linux/slab.h>
   8#include <linux/blkdev.h>
   9#include <linux/list_sort.h>
  10#include <linux/iversion.h>
  11#include "misc.h"
  12#include "ctree.h"
  13#include "tree-log.h"
  14#include "disk-io.h"
  15#include "locking.h"
  16#include "backref.h"
  17#include "compression.h"
  18#include "qgroup.h"
  19#include "block-group.h"
  20#include "space-info.h"
  21#include "inode-item.h"
  22#include "fs.h"
  23#include "accessors.h"
  24#include "extent-tree.h"
  25#include "root-tree.h"
  26#include "dir-item.h"
  27#include "file-item.h"
  28#include "file.h"
  29#include "orphan.h"
  30#include "tree-checker.h"
  31
  32#define MAX_CONFLICT_INODES 10
  33
  34/* magic values for the inode_only field in btrfs_log_inode:
  35 *
  36 * LOG_INODE_ALL means to log everything
  37 * LOG_INODE_EXISTS means to log just enough to recreate the inode
  38 * during log replay
  39 */
  40enum {
  41	LOG_INODE_ALL,
  42	LOG_INODE_EXISTS,
  43};
  44
  45/*
  46 * directory trouble cases
  47 *
  48 * 1) on rename or unlink, if the inode being unlinked isn't in the fsync
  49 * log, we must force a full commit before doing an fsync of the directory
  50 * where the unlink was done.
  51 * ---> record transid of last unlink/rename per directory
  52 *
  53 * mkdir foo/some_dir
  54 * normal commit
  55 * rename foo/some_dir foo2/some_dir
  56 * mkdir foo/some_dir
  57 * fsync foo/some_dir/some_file
  58 *
  59 * The fsync above will unlink the original some_dir without recording
  60 * it in its new location (foo2).  After a crash, some_dir will be gone
  61 * unless the fsync of some_file forces a full commit
  62 *
  63 * 2) we must log any new names for any file or dir that is in the fsync
  64 * log. ---> check inode while renaming/linking.
  65 *
  66 * 2a) we must log any new names for any file or dir during rename
  67 * when the directory they are being removed from was logged.
  68 * ---> check inode and old parent dir during rename
  69 *
  70 *  2a is actually the more important variant.  With the extra logging
  71 *  a crash might unlink the old name without recreating the new one
  72 *
  73 * 3) after a crash, we must go through any directories with a link count
  74 * of zero and redo the rm -rf
  75 *
  76 * mkdir f1/foo
  77 * normal commit
  78 * rm -rf f1/foo
  79 * fsync(f1)
  80 *
  81 * The directory f1 was fully removed from the FS, but fsync was never
  82 * called on f1, only its parent dir.  After a crash the rm -rf must
  83 * be replayed.  This must be able to recurse down the entire
  84 * directory tree.  The inode link count fixup code takes care of the
  85 * ugly details.
  86 */
  87
  88/*
  89 * stages for the tree walking.  The first
  90 * stage (0) is to only pin down the blocks we find
  91 * the second stage (1) is to make sure that all the inodes
  92 * we find in the log are created in the subvolume.
  93 *
  94 * The last stage is to deal with directories and links and extents
  95 * and all the other fun semantics
  96 */
  97enum {
  98	LOG_WALK_PIN_ONLY,
  99	LOG_WALK_REPLAY_INODES,
 100	LOG_WALK_REPLAY_DIR_INDEX,
 101	LOG_WALK_REPLAY_ALL,
 102};
 103
 104static int btrfs_log_inode(struct btrfs_trans_handle *trans,
 105			   struct btrfs_inode *inode,
 106			   int inode_only,
 107			   struct btrfs_log_ctx *ctx);
 108static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
 109			     struct btrfs_root *root,
 110			     struct btrfs_path *path, u64 objectid);
 111static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
 112				       struct btrfs_root *root,
 113				       struct btrfs_root *log,
 114				       struct btrfs_path *path,
 115				       u64 dirid, int del_all);
 116static void wait_log_commit(struct btrfs_root *root, int transid);
 117
 118/*
 119 * tree logging is a special write ahead log used to make sure that
 120 * fsyncs and O_SYNCs can happen without doing full tree commits.
 121 *
 122 * Full tree commits are expensive because they require commonly
 123 * modified blocks to be recowed, creating many dirty pages in the
 124 * extent tree an 4x-6x higher write load than ext3.
 125 *
 126 * Instead of doing a tree commit on every fsync, we use the
 127 * key ranges and transaction ids to find items for a given file or directory
 128 * that have changed in this transaction.  Those items are copied into
 129 * a special tree (one per subvolume root), that tree is written to disk
 130 * and then the fsync is considered complete.
 131 *
 132 * After a crash, items are copied out of the log-tree back into the
 133 * subvolume tree.  Any file data extents found are recorded in the extent
 134 * allocation tree, and the log-tree freed.
 135 *
 136 * The log tree is read three times, once to pin down all the extents it is
 137 * using in ram and once, once to create all the inodes logged in the tree
 138 * and once to do all the other items.
 139 */
 140
 141static struct inode *btrfs_iget_logging(u64 objectid, struct btrfs_root *root)
 142{
 143	unsigned int nofs_flag;
 144	struct inode *inode;
 145
 146	/*
 147	 * We're holding a transaction handle whether we are logging or
 148	 * replaying a log tree, so we must make sure NOFS semantics apply
 149	 * because btrfs_alloc_inode() may be triggered and it uses GFP_KERNEL
 150	 * to allocate an inode, which can recurse back into the filesystem and
 151	 * attempt a transaction commit, resulting in a deadlock.
 152	 */
 153	nofs_flag = memalloc_nofs_save();
 154	inode = btrfs_iget(objectid, root);
 155	memalloc_nofs_restore(nofs_flag);
 156
 157	return inode;
 158}
 159
 160/*
 161 * start a sub transaction and setup the log tree
 162 * this increments the log tree writer count to make the people
 163 * syncing the tree wait for us to finish
 164 */
 165static int start_log_trans(struct btrfs_trans_handle *trans,
 166			   struct btrfs_root *root,
 167			   struct btrfs_log_ctx *ctx)
 168{
 169	struct btrfs_fs_info *fs_info = root->fs_info;
 170	struct btrfs_root *tree_root = fs_info->tree_root;
 171	const bool zoned = btrfs_is_zoned(fs_info);
 172	int ret = 0;
 173	bool created = false;
 174
 175	/*
 176	 * First check if the log root tree was already created. If not, create
 177	 * it before locking the root's log_mutex, just to keep lockdep happy.
 178	 */
 179	if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &tree_root->state)) {
 180		mutex_lock(&tree_root->log_mutex);
 181		if (!fs_info->log_root_tree) {
 182			ret = btrfs_init_log_root_tree(trans, fs_info);
 183			if (!ret) {
 184				set_bit(BTRFS_ROOT_HAS_LOG_TREE, &tree_root->state);
 185				created = true;
 186			}
 187		}
 188		mutex_unlock(&tree_root->log_mutex);
 189		if (ret)
 190			return ret;
 191	}
 192
 193	mutex_lock(&root->log_mutex);
 194
 195again:
 196	if (root->log_root) {
 197		int index = (root->log_transid + 1) % 2;
 198
 199		if (btrfs_need_log_full_commit(trans)) {
 200			ret = BTRFS_LOG_FORCE_COMMIT;
 201			goto out;
 202		}
 203
 204		if (zoned && atomic_read(&root->log_commit[index])) {
 205			wait_log_commit(root, root->log_transid - 1);
 206			goto again;
 207		}
 208
 209		if (!root->log_start_pid) {
 210			clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
 211			root->log_start_pid = current->pid;
 212		} else if (root->log_start_pid != current->pid) {
 213			set_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
 214		}
 215	} else {
 216		/*
 217		 * This means fs_info->log_root_tree was already created
 218		 * for some other FS trees. Do the full commit not to mix
 219		 * nodes from multiple log transactions to do sequential
 220		 * writing.
 221		 */
 222		if (zoned && !created) {
 223			ret = BTRFS_LOG_FORCE_COMMIT;
 224			goto out;
 225		}
 226
 227		ret = btrfs_add_log_tree(trans, root);
 228		if (ret)
 229			goto out;
 230
 231		set_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state);
 232		clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
 233		root->log_start_pid = current->pid;
 234	}
 235
 236	atomic_inc(&root->log_writers);
 237	if (!ctx->logging_new_name) {
 238		int index = root->log_transid % 2;
 239		list_add_tail(&ctx->list, &root->log_ctxs[index]);
 240		ctx->log_transid = root->log_transid;
 241	}
 242
 243out:
 244	mutex_unlock(&root->log_mutex);
 245	return ret;
 246}
 247
 248/*
 249 * returns 0 if there was a log transaction running and we were able
 250 * to join, or returns -ENOENT if there were not transactions
 251 * in progress
 252 */
 253static int join_running_log_trans(struct btrfs_root *root)
 254{
 255	const bool zoned = btrfs_is_zoned(root->fs_info);
 256	int ret = -ENOENT;
 257
 258	if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state))
 259		return ret;
 260
 261	mutex_lock(&root->log_mutex);
 262again:
 263	if (root->log_root) {
 264		int index = (root->log_transid + 1) % 2;
 265
 266		ret = 0;
 267		if (zoned && atomic_read(&root->log_commit[index])) {
 268			wait_log_commit(root, root->log_transid - 1);
 269			goto again;
 270		}
 271		atomic_inc(&root->log_writers);
 272	}
 273	mutex_unlock(&root->log_mutex);
 274	return ret;
 275}
 276
 277/*
 278 * This either makes the current running log transaction wait
 279 * until you call btrfs_end_log_trans() or it makes any future
 280 * log transactions wait until you call btrfs_end_log_trans()
 281 */
 282void btrfs_pin_log_trans(struct btrfs_root *root)
 283{
 284	atomic_inc(&root->log_writers);
 285}
 286
 287/*
 288 * indicate we're done making changes to the log tree
 289 * and wake up anyone waiting to do a sync
 290 */
 291void btrfs_end_log_trans(struct btrfs_root *root)
 292{
 293	if (atomic_dec_and_test(&root->log_writers)) {
 294		/* atomic_dec_and_test implies a barrier */
 295		cond_wake_up_nomb(&root->log_writer_wait);
 296	}
 297}
 298
 299/*
 300 * the walk control struct is used to pass state down the chain when
 301 * processing the log tree.  The stage field tells us which part
 302 * of the log tree processing we are currently doing.  The others
 303 * are state fields used for that specific part
 304 */
 305struct walk_control {
 306	/* should we free the extent on disk when done?  This is used
 307	 * at transaction commit time while freeing a log tree
 308	 */
 309	int free;
 310
 311	/* pin only walk, we record which extents on disk belong to the
 312	 * log trees
 313	 */
 314	int pin;
 315
 316	/* what stage of the replay code we're currently in */
 317	int stage;
 318
 319	/*
 320	 * Ignore any items from the inode currently being processed. Needs
 321	 * to be set every time we find a BTRFS_INODE_ITEM_KEY and we are in
 322	 * the LOG_WALK_REPLAY_INODES stage.
 323	 */
 324	bool ignore_cur_inode;
 325
 326	/* the root we are currently replaying */
 327	struct btrfs_root *replay_dest;
 328
 329	/* the trans handle for the current replay */
 330	struct btrfs_trans_handle *trans;
 331
 332	/* the function that gets used to process blocks we find in the
 333	 * tree.  Note the extent_buffer might not be up to date when it is
 334	 * passed in, and it must be checked or read if you need the data
 335	 * inside it
 336	 */
 337	int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
 338			    struct walk_control *wc, u64 gen, int level);
 339};
 340
 341/*
 342 * process_func used to pin down extents, write them or wait on them
 343 */
 344static int process_one_buffer(struct btrfs_root *log,
 345			      struct extent_buffer *eb,
 346			      struct walk_control *wc, u64 gen, int level)
 347{
 348	struct btrfs_fs_info *fs_info = log->fs_info;
 349	int ret = 0;
 350
 351	/*
 352	 * If this fs is mixed then we need to be able to process the leaves to
 353	 * pin down any logged extents, so we have to read the block.
 354	 */
 355	if (btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
 356		struct btrfs_tree_parent_check check = {
 357			.level = level,
 358			.transid = gen
 359		};
 360
 361		ret = btrfs_read_extent_buffer(eb, &check);
 362		if (ret)
 363			return ret;
 364	}
 365
 366	if (wc->pin) {
 367		ret = btrfs_pin_extent_for_log_replay(wc->trans, eb);
 368		if (ret)
 369			return ret;
 370
 371		if (btrfs_buffer_uptodate(eb, gen, 0) &&
 372		    btrfs_header_level(eb) == 0)
 373			ret = btrfs_exclude_logged_extents(eb);
 374	}
 375	return ret;
 376}
 377
 378/*
 379 * Item overwrite used by replay and tree logging.  eb, slot and key all refer
 380 * to the src data we are copying out.
 381 *
 382 * root is the tree we are copying into, and path is a scratch
 383 * path for use in this function (it should be released on entry and
 384 * will be released on exit).
 385 *
 386 * If the key is already in the destination tree the existing item is
 387 * overwritten.  If the existing item isn't big enough, it is extended.
 388 * If it is too large, it is truncated.
 389 *
 390 * If the key isn't in the destination yet, a new item is inserted.
 391 */
 392static int overwrite_item(struct btrfs_trans_handle *trans,
 393			  struct btrfs_root *root,
 394			  struct btrfs_path *path,
 395			  struct extent_buffer *eb, int slot,
 396			  struct btrfs_key *key)
 397{
 398	int ret;
 399	u32 item_size;
 400	u64 saved_i_size = 0;
 401	int save_old_i_size = 0;
 402	unsigned long src_ptr;
 403	unsigned long dst_ptr;
 404	bool inode_item = key->type == BTRFS_INODE_ITEM_KEY;
 405
 406	/*
 407	 * This is only used during log replay, so the root is always from a
 408	 * fs/subvolume tree. In case we ever need to support a log root, then
 409	 * we'll have to clone the leaf in the path, release the path and use
 410	 * the leaf before writing into the log tree. See the comments at
 411	 * copy_items() for more details.
 412	 */
 413	ASSERT(btrfs_root_id(root) != BTRFS_TREE_LOG_OBJECTID);
 414
 415	item_size = btrfs_item_size(eb, slot);
 416	src_ptr = btrfs_item_ptr_offset(eb, slot);
 417
 418	/* Look for the key in the destination tree. */
 419	ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
 420	if (ret < 0)
 421		return ret;
 422
 423	if (ret == 0) {
 424		char *src_copy;
 425		char *dst_copy;
 426		u32 dst_size = btrfs_item_size(path->nodes[0],
 427						  path->slots[0]);
 428		if (dst_size != item_size)
 429			goto insert;
 430
 431		if (item_size == 0) {
 432			btrfs_release_path(path);
 433			return 0;
 434		}
 435		dst_copy = kmalloc(item_size, GFP_NOFS);
 436		src_copy = kmalloc(item_size, GFP_NOFS);
 437		if (!dst_copy || !src_copy) {
 438			btrfs_release_path(path);
 439			kfree(dst_copy);
 440			kfree(src_copy);
 441			return -ENOMEM;
 442		}
 443
 444		read_extent_buffer(eb, src_copy, src_ptr, item_size);
 445
 446		dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
 447		read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
 448				   item_size);
 449		ret = memcmp(dst_copy, src_copy, item_size);
 450
 451		kfree(dst_copy);
 452		kfree(src_copy);
 453		/*
 454		 * they have the same contents, just return, this saves
 455		 * us from cowing blocks in the destination tree and doing
 456		 * extra writes that may not have been done by a previous
 457		 * sync
 458		 */
 459		if (ret == 0) {
 460			btrfs_release_path(path);
 461			return 0;
 462		}
 463
 464		/*
 465		 * We need to load the old nbytes into the inode so when we
 466		 * replay the extents we've logged we get the right nbytes.
 467		 */
 468		if (inode_item) {
 469			struct btrfs_inode_item *item;
 470			u64 nbytes;
 471			u32 mode;
 472
 473			item = btrfs_item_ptr(path->nodes[0], path->slots[0],
 474					      struct btrfs_inode_item);
 475			nbytes = btrfs_inode_nbytes(path->nodes[0], item);
 476			item = btrfs_item_ptr(eb, slot,
 477					      struct btrfs_inode_item);
 478			btrfs_set_inode_nbytes(eb, item, nbytes);
 479
 480			/*
 481			 * If this is a directory we need to reset the i_size to
 482			 * 0 so that we can set it up properly when replaying
 483			 * the rest of the items in this log.
 484			 */
 485			mode = btrfs_inode_mode(eb, item);
 486			if (S_ISDIR(mode))
 487				btrfs_set_inode_size(eb, item, 0);
 488		}
 489	} else if (inode_item) {
 490		struct btrfs_inode_item *item;
 491		u32 mode;
 492
 493		/*
 494		 * New inode, set nbytes to 0 so that the nbytes comes out
 495		 * properly when we replay the extents.
 496		 */
 497		item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
 498		btrfs_set_inode_nbytes(eb, item, 0);
 499
 500		/*
 501		 * If this is a directory we need to reset the i_size to 0 so
 502		 * that we can set it up properly when replaying the rest of
 503		 * the items in this log.
 504		 */
 505		mode = btrfs_inode_mode(eb, item);
 506		if (S_ISDIR(mode))
 507			btrfs_set_inode_size(eb, item, 0);
 508	}
 509insert:
 510	btrfs_release_path(path);
 511	/* try to insert the key into the destination tree */
 512	path->skip_release_on_error = 1;
 513	ret = btrfs_insert_empty_item(trans, root, path,
 514				      key, item_size);
 515	path->skip_release_on_error = 0;
 516
 517	/* make sure any existing item is the correct size */
 518	if (ret == -EEXIST || ret == -EOVERFLOW) {
 519		u32 found_size;
 520		found_size = btrfs_item_size(path->nodes[0],
 521						path->slots[0]);
 522		if (found_size > item_size)
 523			btrfs_truncate_item(trans, path, item_size, 1);
 524		else if (found_size < item_size)
 525			btrfs_extend_item(trans, path, item_size - found_size);
 526	} else if (ret) {
 527		return ret;
 528	}
 529	dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
 530					path->slots[0]);
 531
 532	/* don't overwrite an existing inode if the generation number
 533	 * was logged as zero.  This is done when the tree logging code
 534	 * is just logging an inode to make sure it exists after recovery.
 535	 *
 536	 * Also, don't overwrite i_size on directories during replay.
 537	 * log replay inserts and removes directory items based on the
 538	 * state of the tree found in the subvolume, and i_size is modified
 539	 * as it goes
 540	 */
 541	if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
 542		struct btrfs_inode_item *src_item;
 543		struct btrfs_inode_item *dst_item;
 544
 545		src_item = (struct btrfs_inode_item *)src_ptr;
 546		dst_item = (struct btrfs_inode_item *)dst_ptr;
 547
 548		if (btrfs_inode_generation(eb, src_item) == 0) {
 549			struct extent_buffer *dst_eb = path->nodes[0];
 550			const u64 ino_size = btrfs_inode_size(eb, src_item);
 551
 552			/*
 553			 * For regular files an ino_size == 0 is used only when
 554			 * logging that an inode exists, as part of a directory
 555			 * fsync, and the inode wasn't fsynced before. In this
 556			 * case don't set the size of the inode in the fs/subvol
 557			 * tree, otherwise we would be throwing valid data away.
 558			 */
 559			if (S_ISREG(btrfs_inode_mode(eb, src_item)) &&
 560			    S_ISREG(btrfs_inode_mode(dst_eb, dst_item)) &&
 561			    ino_size != 0)
 562				btrfs_set_inode_size(dst_eb, dst_item, ino_size);
 563			goto no_copy;
 564		}
 565
 566		if (S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
 567		    S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
 568			save_old_i_size = 1;
 569			saved_i_size = btrfs_inode_size(path->nodes[0],
 570							dst_item);
 571		}
 572	}
 573
 574	copy_extent_buffer(path->nodes[0], eb, dst_ptr,
 575			   src_ptr, item_size);
 576
 577	if (save_old_i_size) {
 578		struct btrfs_inode_item *dst_item;
 579		dst_item = (struct btrfs_inode_item *)dst_ptr;
 580		btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
 581	}
 582
 583	/* make sure the generation is filled in */
 584	if (key->type == BTRFS_INODE_ITEM_KEY) {
 585		struct btrfs_inode_item *dst_item;
 586		dst_item = (struct btrfs_inode_item *)dst_ptr;
 587		if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
 588			btrfs_set_inode_generation(path->nodes[0], dst_item,
 589						   trans->transid);
 590		}
 591	}
 592no_copy:
 593	btrfs_mark_buffer_dirty(trans, path->nodes[0]);
 594	btrfs_release_path(path);
 595	return 0;
 596}
 597
 598static int read_alloc_one_name(struct extent_buffer *eb, void *start, int len,
 599			       struct fscrypt_str *name)
 600{
 601	char *buf;
 602
 603	buf = kmalloc(len, GFP_NOFS);
 604	if (!buf)
 605		return -ENOMEM;
 606
 607	read_extent_buffer(eb, buf, (unsigned long)start, len);
 608	name->name = buf;
 609	name->len = len;
 610	return 0;
 611}
 612
 613/*
 614 * simple helper to read an inode off the disk from a given root
 615 * This can only be called for subvolume roots and not for the log
 616 */
 617static noinline struct inode *read_one_inode(struct btrfs_root *root,
 618					     u64 objectid)
 619{
 620	struct inode *inode;
 621
 622	inode = btrfs_iget_logging(objectid, root);
 623	if (IS_ERR(inode))
 624		inode = NULL;
 625	return inode;
 626}
 627
 628/* replays a single extent in 'eb' at 'slot' with 'key' into the
 629 * subvolume 'root'.  path is released on entry and should be released
 630 * on exit.
 631 *
 632 * extents in the log tree have not been allocated out of the extent
 633 * tree yet.  So, this completes the allocation, taking a reference
 634 * as required if the extent already exists or creating a new extent
 635 * if it isn't in the extent allocation tree yet.
 636 *
 637 * The extent is inserted into the file, dropping any existing extents
 638 * from the file that overlap the new one.
 639 */
 640static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
 641				      struct btrfs_root *root,
 642				      struct btrfs_path *path,
 643				      struct extent_buffer *eb, int slot,
 644				      struct btrfs_key *key)
 645{
 646	struct btrfs_drop_extents_args drop_args = { 0 };
 647	struct btrfs_fs_info *fs_info = root->fs_info;
 648	int found_type;
 649	u64 extent_end;
 650	u64 start = key->offset;
 651	u64 nbytes = 0;
 652	struct btrfs_file_extent_item *item;
 653	struct inode *inode = NULL;
 654	unsigned long size;
 655	int ret = 0;
 656
 657	item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 658	found_type = btrfs_file_extent_type(eb, item);
 659
 660	if (found_type == BTRFS_FILE_EXTENT_REG ||
 661	    found_type == BTRFS_FILE_EXTENT_PREALLOC) {
 662		nbytes = btrfs_file_extent_num_bytes(eb, item);
 663		extent_end = start + nbytes;
 664
 665		/*
 666		 * We don't add to the inodes nbytes if we are prealloc or a
 667		 * hole.
 668		 */
 669		if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
 670			nbytes = 0;
 671	} else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 672		size = btrfs_file_extent_ram_bytes(eb, item);
 673		nbytes = btrfs_file_extent_ram_bytes(eb, item);
 674		extent_end = ALIGN(start + size,
 675				   fs_info->sectorsize);
 676	} else {
 677		ret = 0;
 678		goto out;
 679	}
 680
 681	inode = read_one_inode(root, key->objectid);
 682	if (!inode) {
 683		ret = -EIO;
 684		goto out;
 685	}
 686
 687	/*
 688	 * first check to see if we already have this extent in the
 689	 * file.  This must be done before the btrfs_drop_extents run
 690	 * so we don't try to drop this extent.
 691	 */
 692	ret = btrfs_lookup_file_extent(trans, root, path,
 693			btrfs_ino(BTRFS_I(inode)), start, 0);
 694
 695	if (ret == 0 &&
 696	    (found_type == BTRFS_FILE_EXTENT_REG ||
 697	     found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
 698		struct btrfs_file_extent_item cmp1;
 699		struct btrfs_file_extent_item cmp2;
 700		struct btrfs_file_extent_item *existing;
 701		struct extent_buffer *leaf;
 702
 703		leaf = path->nodes[0];
 704		existing = btrfs_item_ptr(leaf, path->slots[0],
 705					  struct btrfs_file_extent_item);
 706
 707		read_extent_buffer(eb, &cmp1, (unsigned long)item,
 708				   sizeof(cmp1));
 709		read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
 710				   sizeof(cmp2));
 711
 712		/*
 713		 * we already have a pointer to this exact extent,
 714		 * we don't have to do anything
 715		 */
 716		if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
 717			btrfs_release_path(path);
 718			goto out;
 719		}
 720	}
 721	btrfs_release_path(path);
 722
 723	/* drop any overlapping extents */
 724	drop_args.start = start;
 725	drop_args.end = extent_end;
 726	drop_args.drop_cache = true;
 727	ret = btrfs_drop_extents(trans, root, BTRFS_I(inode), &drop_args);
 728	if (ret)
 729		goto out;
 730
 731	if (found_type == BTRFS_FILE_EXTENT_REG ||
 732	    found_type == BTRFS_FILE_EXTENT_PREALLOC) {
 733		u64 offset;
 734		unsigned long dest_offset;
 735		struct btrfs_key ins;
 736
 737		if (btrfs_file_extent_disk_bytenr(eb, item) == 0 &&
 738		    btrfs_fs_incompat(fs_info, NO_HOLES))
 739			goto update_inode;
 740
 741		ret = btrfs_insert_empty_item(trans, root, path, key,
 742					      sizeof(*item));
 743		if (ret)
 744			goto out;
 745		dest_offset = btrfs_item_ptr_offset(path->nodes[0],
 746						    path->slots[0]);
 747		copy_extent_buffer(path->nodes[0], eb, dest_offset,
 748				(unsigned long)item,  sizeof(*item));
 749
 750		ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
 751		ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
 752		ins.type = BTRFS_EXTENT_ITEM_KEY;
 753		offset = key->offset - btrfs_file_extent_offset(eb, item);
 754
 755		/*
 756		 * Manually record dirty extent, as here we did a shallow
 757		 * file extent item copy and skip normal backref update,
 758		 * but modifying extent tree all by ourselves.
 759		 * So need to manually record dirty extent for qgroup,
 760		 * as the owner of the file extent changed from log tree
 761		 * (doesn't affect qgroup) to fs/file tree(affects qgroup)
 762		 */
 763		ret = btrfs_qgroup_trace_extent(trans,
 764				btrfs_file_extent_disk_bytenr(eb, item),
 765				btrfs_file_extent_disk_num_bytes(eb, item));
 766		if (ret < 0)
 767			goto out;
 768
 769		if (ins.objectid > 0) {
 
 770			u64 csum_start;
 771			u64 csum_end;
 772			LIST_HEAD(ordered_sums);
 773
 774			/*
 775			 * is this extent already allocated in the extent
 776			 * allocation tree?  If so, just add a reference
 777			 */
 778			ret = btrfs_lookup_data_extent(fs_info, ins.objectid,
 779						ins.offset);
 780			if (ret < 0) {
 781				goto out;
 782			} else if (ret == 0) {
 783				struct btrfs_ref ref = {
 784					.action = BTRFS_ADD_DELAYED_REF,
 785					.bytenr = ins.objectid,
 786					.num_bytes = ins.offset,
 787					.owning_root = btrfs_root_id(root),
 788					.ref_root = btrfs_root_id(root),
 789				};
 790				btrfs_init_data_ref(&ref, key->objectid, offset,
 791						    0, false);
 792				ret = btrfs_inc_extent_ref(trans, &ref);
 793				if (ret)
 794					goto out;
 795			} else {
 796				/*
 797				 * insert the extent pointer in the extent
 798				 * allocation tree
 799				 */
 800				ret = btrfs_alloc_logged_file_extent(trans,
 801						btrfs_root_id(root),
 802						key->objectid, offset, &ins);
 803				if (ret)
 804					goto out;
 805			}
 806			btrfs_release_path(path);
 807
 808			if (btrfs_file_extent_compression(eb, item)) {
 809				csum_start = ins.objectid;
 810				csum_end = csum_start + ins.offset;
 811			} else {
 812				csum_start = ins.objectid +
 813					btrfs_file_extent_offset(eb, item);
 814				csum_end = csum_start +
 815					btrfs_file_extent_num_bytes(eb, item);
 816			}
 817
 818			ret = btrfs_lookup_csums_list(root->log_root,
 819						csum_start, csum_end - 1,
 820						&ordered_sums, false);
 821			if (ret < 0)
 822				goto out;
 823			ret = 0;
 824			/*
 825			 * Now delete all existing cums in the csum root that
 826			 * cover our range. We do this because we can have an
 827			 * extent that is completely referenced by one file
 828			 * extent item and partially referenced by another
 829			 * file extent item (like after using the clone or
 830			 * extent_same ioctls). In this case if we end up doing
 831			 * the replay of the one that partially references the
 832			 * extent first, and we do not do the csum deletion
 833			 * below, we can get 2 csum items in the csum tree that
 834			 * overlap each other. For example, imagine our log has
 835			 * the two following file extent items:
 836			 *
 837			 * key (257 EXTENT_DATA 409600)
 838			 *     extent data disk byte 12845056 nr 102400
 839			 *     extent data offset 20480 nr 20480 ram 102400
 840			 *
 841			 * key (257 EXTENT_DATA 819200)
 842			 *     extent data disk byte 12845056 nr 102400
 843			 *     extent data offset 0 nr 102400 ram 102400
 844			 *
 845			 * Where the second one fully references the 100K extent
 846			 * that starts at disk byte 12845056, and the log tree
 847			 * has a single csum item that covers the entire range
 848			 * of the extent:
 849			 *
 850			 * key (EXTENT_CSUM EXTENT_CSUM 12845056) itemsize 100
 851			 *
 852			 * After the first file extent item is replayed, the
 853			 * csum tree gets the following csum item:
 854			 *
 855			 * key (EXTENT_CSUM EXTENT_CSUM 12865536) itemsize 20
 856			 *
 857			 * Which covers the 20K sub-range starting at offset 20K
 858			 * of our extent. Now when we replay the second file
 859			 * extent item, if we do not delete existing csum items
 860			 * that cover any of its blocks, we end up getting two
 861			 * csum items in our csum tree that overlap each other:
 862			 *
 863			 * key (EXTENT_CSUM EXTENT_CSUM 12845056) itemsize 100
 864			 * key (EXTENT_CSUM EXTENT_CSUM 12865536) itemsize 20
 865			 *
 866			 * Which is a problem, because after this anyone trying
 867			 * to lookup up for the checksum of any block of our
 868			 * extent starting at an offset of 40K or higher, will
 869			 * end up looking at the second csum item only, which
 870			 * does not contain the checksum for any block starting
 871			 * at offset 40K or higher of our extent.
 872			 */
 873			while (!list_empty(&ordered_sums)) {
 874				struct btrfs_ordered_sum *sums;
 875				struct btrfs_root *csum_root;
 876
 877				sums = list_entry(ordered_sums.next,
 878						struct btrfs_ordered_sum,
 879						list);
 880				csum_root = btrfs_csum_root(fs_info,
 881							    sums->logical);
 882				if (!ret)
 883					ret = btrfs_del_csums(trans, csum_root,
 884							      sums->logical,
 885							      sums->len);
 886				if (!ret)
 887					ret = btrfs_csum_file_blocks(trans,
 888								     csum_root,
 889								     sums);
 890				list_del(&sums->list);
 891				kfree(sums);
 892			}
 893			if (ret)
 894				goto out;
 895		} else {
 896			btrfs_release_path(path);
 897		}
 898	} else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 899		/* inline extents are easy, we just overwrite them */
 900		ret = overwrite_item(trans, root, path, eb, slot, key);
 901		if (ret)
 902			goto out;
 903	}
 904
 905	ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), start,
 906						extent_end - start);
 907	if (ret)
 908		goto out;
 909
 910update_inode:
 911	btrfs_update_inode_bytes(BTRFS_I(inode), nbytes, drop_args.bytes_found);
 912	ret = btrfs_update_inode(trans, BTRFS_I(inode));
 913out:
 914	iput(inode);
 915	return ret;
 916}
 917
 918static int unlink_inode_for_log_replay(struct btrfs_trans_handle *trans,
 919				       struct btrfs_inode *dir,
 920				       struct btrfs_inode *inode,
 921				       const struct fscrypt_str *name)
 922{
 923	int ret;
 924
 925	ret = btrfs_unlink_inode(trans, dir, inode, name);
 926	if (ret)
 927		return ret;
 928	/*
 929	 * Whenever we need to check if a name exists or not, we check the
 930	 * fs/subvolume tree. So after an unlink we must run delayed items, so
 931	 * that future checks for a name during log replay see that the name
 932	 * does not exists anymore.
 933	 */
 934	return btrfs_run_delayed_items(trans);
 935}
 936
 937/*
 938 * when cleaning up conflicts between the directory names in the
 939 * subvolume, directory names in the log and directory names in the
 940 * inode back references, we may have to unlink inodes from directories.
 941 *
 942 * This is a helper function to do the unlink of a specific directory
 943 * item
 944 */
 945static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
 946				      struct btrfs_path *path,
 947				      struct btrfs_inode *dir,
 948				      struct btrfs_dir_item *di)
 949{
 950	struct btrfs_root *root = dir->root;
 951	struct inode *inode;
 952	struct fscrypt_str name;
 953	struct extent_buffer *leaf;
 954	struct btrfs_key location;
 955	int ret;
 956
 957	leaf = path->nodes[0];
 958
 959	btrfs_dir_item_key_to_cpu(leaf, di, &location);
 960	ret = read_alloc_one_name(leaf, di + 1, btrfs_dir_name_len(leaf, di), &name);
 961	if (ret)
 962		return -ENOMEM;
 963
 964	btrfs_release_path(path);
 965
 966	inode = read_one_inode(root, location.objectid);
 967	if (!inode) {
 968		ret = -EIO;
 969		goto out;
 970	}
 971
 972	ret = link_to_fixup_dir(trans, root, path, location.objectid);
 973	if (ret)
 974		goto out;
 975
 976	ret = unlink_inode_for_log_replay(trans, dir, BTRFS_I(inode), &name);
 977out:
 978	kfree(name.name);
 979	iput(inode);
 980	return ret;
 981}
 982
 983/*
 984 * See if a given name and sequence number found in an inode back reference are
 985 * already in a directory and correctly point to this inode.
 986 *
 987 * Returns: < 0 on error, 0 if the directory entry does not exists and 1 if it
 988 * exists.
 989 */
 990static noinline int inode_in_dir(struct btrfs_root *root,
 991				 struct btrfs_path *path,
 992				 u64 dirid, u64 objectid, u64 index,
 993				 struct fscrypt_str *name)
 994{
 995	struct btrfs_dir_item *di;
 996	struct btrfs_key location;
 997	int ret = 0;
 998
 999	di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
1000					 index, name, 0);
1001	if (IS_ERR(di)) {
1002		ret = PTR_ERR(di);
1003		goto out;
1004	} else if (di) {
1005		btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
1006		if (location.objectid != objectid)
1007			goto out;
1008	} else {
1009		goto out;
1010	}
1011
1012	btrfs_release_path(path);
1013	di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, 0);
1014	if (IS_ERR(di)) {
1015		ret = PTR_ERR(di);
1016		goto out;
1017	} else if (di) {
1018		btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
1019		if (location.objectid == objectid)
1020			ret = 1;
1021	}
1022out:
1023	btrfs_release_path(path);
1024	return ret;
1025}
1026
1027/*
1028 * helper function to check a log tree for a named back reference in
1029 * an inode.  This is used to decide if a back reference that is
1030 * found in the subvolume conflicts with what we find in the log.
1031 *
1032 * inode backreferences may have multiple refs in a single item,
1033 * during replay we process one reference at a time, and we don't
1034 * want to delete valid links to a file from the subvolume if that
1035 * link is also in the log.
1036 */
1037static noinline int backref_in_log(struct btrfs_root *log,
1038				   struct btrfs_key *key,
1039				   u64 ref_objectid,
1040				   const struct fscrypt_str *name)
1041{
1042	struct btrfs_path *path;
1043	int ret;
1044
1045	path = btrfs_alloc_path();
1046	if (!path)
1047		return -ENOMEM;
1048
1049	ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
1050	if (ret < 0) {
1051		goto out;
1052	} else if (ret == 1) {
1053		ret = 0;
1054		goto out;
1055	}
1056
1057	if (key->type == BTRFS_INODE_EXTREF_KEY)
1058		ret = !!btrfs_find_name_in_ext_backref(path->nodes[0],
1059						       path->slots[0],
1060						       ref_objectid, name);
1061	else
1062		ret = !!btrfs_find_name_in_backref(path->nodes[0],
1063						   path->slots[0], name);
1064out:
1065	btrfs_free_path(path);
1066	return ret;
1067}
1068
1069static inline int __add_inode_ref(struct btrfs_trans_handle *trans,
1070				  struct btrfs_root *root,
1071				  struct btrfs_path *path,
1072				  struct btrfs_root *log_root,
1073				  struct btrfs_inode *dir,
1074				  struct btrfs_inode *inode,
1075				  u64 inode_objectid, u64 parent_objectid,
1076				  u64 ref_index, struct fscrypt_str *name)
1077{
1078	int ret;
1079	struct extent_buffer *leaf;
1080	struct btrfs_dir_item *di;
1081	struct btrfs_key search_key;
1082	struct btrfs_inode_extref *extref;
1083
1084again:
1085	/* Search old style refs */
1086	search_key.objectid = inode_objectid;
1087	search_key.type = BTRFS_INODE_REF_KEY;
1088	search_key.offset = parent_objectid;
1089	ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
1090	if (ret == 0) {
1091		struct btrfs_inode_ref *victim_ref;
1092		unsigned long ptr;
1093		unsigned long ptr_end;
1094
1095		leaf = path->nodes[0];
1096
1097		/* are we trying to overwrite a back ref for the root directory
1098		 * if so, just jump out, we're done
1099		 */
1100		if (search_key.objectid == search_key.offset)
1101			return 1;
1102
1103		/* check all the names in this back reference to see
1104		 * if they are in the log.  if so, we allow them to stay
1105		 * otherwise they must be unlinked as a conflict
1106		 */
1107		ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1108		ptr_end = ptr + btrfs_item_size(leaf, path->slots[0]);
1109		while (ptr < ptr_end) {
1110			struct fscrypt_str victim_name;
1111
1112			victim_ref = (struct btrfs_inode_ref *)ptr;
1113			ret = read_alloc_one_name(leaf, (victim_ref + 1),
1114				 btrfs_inode_ref_name_len(leaf, victim_ref),
1115				 &victim_name);
1116			if (ret)
1117				return ret;
1118
1119			ret = backref_in_log(log_root, &search_key,
1120					     parent_objectid, &victim_name);
1121			if (ret < 0) {
1122				kfree(victim_name.name);
1123				return ret;
1124			} else if (!ret) {
1125				inc_nlink(&inode->vfs_inode);
1126				btrfs_release_path(path);
1127
1128				ret = unlink_inode_for_log_replay(trans, dir, inode,
1129						&victim_name);
1130				kfree(victim_name.name);
1131				if (ret)
1132					return ret;
1133				goto again;
1134			}
1135			kfree(victim_name.name);
1136
1137			ptr = (unsigned long)(victim_ref + 1) + victim_name.len;
1138		}
1139	}
1140	btrfs_release_path(path);
1141
1142	/* Same search but for extended refs */
1143	extref = btrfs_lookup_inode_extref(NULL, root, path, name,
1144					   inode_objectid, parent_objectid, 0,
1145					   0);
1146	if (IS_ERR(extref)) {
1147		return PTR_ERR(extref);
1148	} else if (extref) {
1149		u32 item_size;
1150		u32 cur_offset = 0;
1151		unsigned long base;
1152		struct inode *victim_parent;
1153
1154		leaf = path->nodes[0];
1155
1156		item_size = btrfs_item_size(leaf, path->slots[0]);
1157		base = btrfs_item_ptr_offset(leaf, path->slots[0]);
1158
1159		while (cur_offset < item_size) {
1160			struct fscrypt_str victim_name;
1161
1162			extref = (struct btrfs_inode_extref *)(base + cur_offset);
1163
1164			if (btrfs_inode_extref_parent(leaf, extref) != parent_objectid)
1165				goto next;
1166
1167			ret = read_alloc_one_name(leaf, &extref->name,
1168				 btrfs_inode_extref_name_len(leaf, extref),
1169				 &victim_name);
1170			if (ret)
1171				return ret;
1172
1173			search_key.objectid = inode_objectid;
1174			search_key.type = BTRFS_INODE_EXTREF_KEY;
1175			search_key.offset = btrfs_extref_hash(parent_objectid,
1176							      victim_name.name,
1177							      victim_name.len);
1178			ret = backref_in_log(log_root, &search_key,
1179					     parent_objectid, &victim_name);
1180			if (ret < 0) {
1181				kfree(victim_name.name);
1182				return ret;
1183			} else if (!ret) {
1184				ret = -ENOENT;
1185				victim_parent = read_one_inode(root,
1186						parent_objectid);
1187				if (victim_parent) {
1188					inc_nlink(&inode->vfs_inode);
1189					btrfs_release_path(path);
1190
1191					ret = unlink_inode_for_log_replay(trans,
1192							BTRFS_I(victim_parent),
1193							inode, &victim_name);
1194				}
1195				iput(victim_parent);
1196				kfree(victim_name.name);
1197				if (ret)
1198					return ret;
1199				goto again;
1200			}
1201			kfree(victim_name.name);
1202next:
1203			cur_offset += victim_name.len + sizeof(*extref);
1204		}
1205	}
1206	btrfs_release_path(path);
1207
1208	/* look for a conflicting sequence number */
1209	di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
1210					 ref_index, name, 0);
1211	if (IS_ERR(di)) {
1212		return PTR_ERR(di);
1213	} else if (di) {
1214		ret = drop_one_dir_item(trans, path, dir, di);
1215		if (ret)
1216			return ret;
1217	}
1218	btrfs_release_path(path);
1219
1220	/* look for a conflicting name */
1221	di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir), name, 0);
1222	if (IS_ERR(di)) {
1223		return PTR_ERR(di);
1224	} else if (di) {
1225		ret = drop_one_dir_item(trans, path, dir, di);
1226		if (ret)
1227			return ret;
1228	}
1229	btrfs_release_path(path);
1230
1231	return 0;
1232}
1233
1234static int extref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1235			     struct fscrypt_str *name, u64 *index,
1236			     u64 *parent_objectid)
1237{
1238	struct btrfs_inode_extref *extref;
1239	int ret;
1240
1241	extref = (struct btrfs_inode_extref *)ref_ptr;
1242
1243	ret = read_alloc_one_name(eb, &extref->name,
1244				  btrfs_inode_extref_name_len(eb, extref), name);
1245	if (ret)
1246		return ret;
1247
1248	if (index)
1249		*index = btrfs_inode_extref_index(eb, extref);
1250	if (parent_objectid)
1251		*parent_objectid = btrfs_inode_extref_parent(eb, extref);
1252
1253	return 0;
1254}
1255
1256static int ref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1257			  struct fscrypt_str *name, u64 *index)
1258{
1259	struct btrfs_inode_ref *ref;
1260	int ret;
1261
1262	ref = (struct btrfs_inode_ref *)ref_ptr;
1263
1264	ret = read_alloc_one_name(eb, ref + 1, btrfs_inode_ref_name_len(eb, ref),
1265				  name);
1266	if (ret)
1267		return ret;
1268
1269	if (index)
1270		*index = btrfs_inode_ref_index(eb, ref);
1271
1272	return 0;
1273}
1274
1275/*
1276 * Take an inode reference item from the log tree and iterate all names from the
1277 * inode reference item in the subvolume tree with the same key (if it exists).
1278 * For any name that is not in the inode reference item from the log tree, do a
1279 * proper unlink of that name (that is, remove its entry from the inode
1280 * reference item and both dir index keys).
1281 */
1282static int unlink_old_inode_refs(struct btrfs_trans_handle *trans,
1283				 struct btrfs_root *root,
1284				 struct btrfs_path *path,
1285				 struct btrfs_inode *inode,
1286				 struct extent_buffer *log_eb,
1287				 int log_slot,
1288				 struct btrfs_key *key)
1289{
1290	int ret;
1291	unsigned long ref_ptr;
1292	unsigned long ref_end;
1293	struct extent_buffer *eb;
1294
1295again:
1296	btrfs_release_path(path);
1297	ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
1298	if (ret > 0) {
1299		ret = 0;
1300		goto out;
1301	}
1302	if (ret < 0)
1303		goto out;
1304
1305	eb = path->nodes[0];
1306	ref_ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
1307	ref_end = ref_ptr + btrfs_item_size(eb, path->slots[0]);
1308	while (ref_ptr < ref_end) {
1309		struct fscrypt_str name;
1310		u64 parent_id;
1311
1312		if (key->type == BTRFS_INODE_EXTREF_KEY) {
1313			ret = extref_get_fields(eb, ref_ptr, &name,
1314						NULL, &parent_id);
1315		} else {
1316			parent_id = key->offset;
1317			ret = ref_get_fields(eb, ref_ptr, &name, NULL);
1318		}
1319		if (ret)
1320			goto out;
1321
1322		if (key->type == BTRFS_INODE_EXTREF_KEY)
1323			ret = !!btrfs_find_name_in_ext_backref(log_eb, log_slot,
1324							       parent_id, &name);
1325		else
1326			ret = !!btrfs_find_name_in_backref(log_eb, log_slot, &name);
1327
1328		if (!ret) {
1329			struct inode *dir;
1330
1331			btrfs_release_path(path);
1332			dir = read_one_inode(root, parent_id);
1333			if (!dir) {
1334				ret = -ENOENT;
1335				kfree(name.name);
1336				goto out;
1337			}
1338			ret = unlink_inode_for_log_replay(trans, BTRFS_I(dir),
1339						 inode, &name);
1340			kfree(name.name);
1341			iput(dir);
1342			if (ret)
1343				goto out;
1344			goto again;
1345		}
1346
1347		kfree(name.name);
1348		ref_ptr += name.len;
1349		if (key->type == BTRFS_INODE_EXTREF_KEY)
1350			ref_ptr += sizeof(struct btrfs_inode_extref);
1351		else
1352			ref_ptr += sizeof(struct btrfs_inode_ref);
1353	}
1354	ret = 0;
1355 out:
1356	btrfs_release_path(path);
1357	return ret;
1358}
1359
1360/*
1361 * replay one inode back reference item found in the log tree.
1362 * eb, slot and key refer to the buffer and key found in the log tree.
1363 * root is the destination we are replaying into, and path is for temp
1364 * use by this function.  (it should be released on return).
1365 */
1366static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
1367				  struct btrfs_root *root,
1368				  struct btrfs_root *log,
1369				  struct btrfs_path *path,
1370				  struct extent_buffer *eb, int slot,
1371				  struct btrfs_key *key)
1372{
1373	struct inode *dir = NULL;
1374	struct inode *inode = NULL;
1375	unsigned long ref_ptr;
1376	unsigned long ref_end;
1377	struct fscrypt_str name = { 0 };
1378	int ret;
1379	int log_ref_ver = 0;
1380	u64 parent_objectid;
1381	u64 inode_objectid;
1382	u64 ref_index = 0;
1383	int ref_struct_size;
1384
1385	ref_ptr = btrfs_item_ptr_offset(eb, slot);
1386	ref_end = ref_ptr + btrfs_item_size(eb, slot);
1387
1388	if (key->type == BTRFS_INODE_EXTREF_KEY) {
1389		struct btrfs_inode_extref *r;
1390
1391		ref_struct_size = sizeof(struct btrfs_inode_extref);
1392		log_ref_ver = 1;
1393		r = (struct btrfs_inode_extref *)ref_ptr;
1394		parent_objectid = btrfs_inode_extref_parent(eb, r);
1395	} else {
1396		ref_struct_size = sizeof(struct btrfs_inode_ref);
1397		parent_objectid = key->offset;
1398	}
1399	inode_objectid = key->objectid;
1400
1401	/*
1402	 * it is possible that we didn't log all the parent directories
1403	 * for a given inode.  If we don't find the dir, just don't
1404	 * copy the back ref in.  The link count fixup code will take
1405	 * care of the rest
1406	 */
1407	dir = read_one_inode(root, parent_objectid);
1408	if (!dir) {
1409		ret = -ENOENT;
1410		goto out;
1411	}
1412
1413	inode = read_one_inode(root, inode_objectid);
1414	if (!inode) {
1415		ret = -EIO;
1416		goto out;
1417	}
1418
1419	while (ref_ptr < ref_end) {
1420		if (log_ref_ver) {
1421			ret = extref_get_fields(eb, ref_ptr, &name,
1422						&ref_index, &parent_objectid);
1423			/*
1424			 * parent object can change from one array
1425			 * item to another.
1426			 */
1427			if (!dir)
1428				dir = read_one_inode(root, parent_objectid);
1429			if (!dir) {
1430				ret = -ENOENT;
1431				goto out;
1432			}
1433		} else {
1434			ret = ref_get_fields(eb, ref_ptr, &name, &ref_index);
1435		}
1436		if (ret)
1437			goto out;
1438
1439		ret = inode_in_dir(root, path, btrfs_ino(BTRFS_I(dir)),
1440				   btrfs_ino(BTRFS_I(inode)), ref_index, &name);
1441		if (ret < 0) {
1442			goto out;
1443		} else if (ret == 0) {
1444			/*
1445			 * look for a conflicting back reference in the
1446			 * metadata. if we find one we have to unlink that name
1447			 * of the file before we add our new link.  Later on, we
1448			 * overwrite any existing back reference, and we don't
1449			 * want to create dangling pointers in the directory.
1450			 */
1451			ret = __add_inode_ref(trans, root, path, log,
1452					      BTRFS_I(dir), BTRFS_I(inode),
1453					      inode_objectid, parent_objectid,
1454					      ref_index, &name);
1455			if (ret) {
1456				if (ret == 1)
1457					ret = 0;
1458				goto out;
1459			}
1460
1461			/* insert our name */
1462			ret = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
1463					     &name, 0, ref_index);
1464			if (ret)
1465				goto out;
1466
1467			ret = btrfs_update_inode(trans, BTRFS_I(inode));
1468			if (ret)
1469				goto out;
1470		}
1471		/* Else, ret == 1, we already have a perfect match, we're done. */
1472
1473		ref_ptr = (unsigned long)(ref_ptr + ref_struct_size) + name.len;
1474		kfree(name.name);
1475		name.name = NULL;
1476		if (log_ref_ver) {
1477			iput(dir);
1478			dir = NULL;
1479		}
1480	}
1481
1482	/*
1483	 * Before we overwrite the inode reference item in the subvolume tree
1484	 * with the item from the log tree, we must unlink all names from the
1485	 * parent directory that are in the subvolume's tree inode reference
1486	 * item, otherwise we end up with an inconsistent subvolume tree where
1487	 * dir index entries exist for a name but there is no inode reference
1488	 * item with the same name.
1489	 */
1490	ret = unlink_old_inode_refs(trans, root, path, BTRFS_I(inode), eb, slot,
1491				    key);
1492	if (ret)
1493		goto out;
1494
1495	/* finally write the back reference in the inode */
1496	ret = overwrite_item(trans, root, path, eb, slot, key);
1497out:
1498	btrfs_release_path(path);
1499	kfree(name.name);
1500	iput(dir);
1501	iput(inode);
1502	return ret;
1503}
1504
1505static int count_inode_extrefs(struct btrfs_inode *inode, struct btrfs_path *path)
1506{
1507	int ret = 0;
1508	int name_len;
1509	unsigned int nlink = 0;
1510	u32 item_size;
1511	u32 cur_offset = 0;
1512	u64 inode_objectid = btrfs_ino(inode);
1513	u64 offset = 0;
1514	unsigned long ptr;
1515	struct btrfs_inode_extref *extref;
1516	struct extent_buffer *leaf;
1517
1518	while (1) {
1519		ret = btrfs_find_one_extref(inode->root, inode_objectid, offset,
1520					    path, &extref, &offset);
1521		if (ret)
1522			break;
1523
1524		leaf = path->nodes[0];
1525		item_size = btrfs_item_size(leaf, path->slots[0]);
1526		ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1527		cur_offset = 0;
1528
1529		while (cur_offset < item_size) {
1530			extref = (struct btrfs_inode_extref *) (ptr + cur_offset);
1531			name_len = btrfs_inode_extref_name_len(leaf, extref);
1532
1533			nlink++;
1534
1535			cur_offset += name_len + sizeof(*extref);
1536		}
1537
1538		offset++;
1539		btrfs_release_path(path);
1540	}
1541	btrfs_release_path(path);
1542
1543	if (ret < 0 && ret != -ENOENT)
1544		return ret;
1545	return nlink;
1546}
1547
1548static int count_inode_refs(struct btrfs_inode *inode, struct btrfs_path *path)
1549{
1550	int ret;
1551	struct btrfs_key key;
1552	unsigned int nlink = 0;
1553	unsigned long ptr;
1554	unsigned long ptr_end;
1555	int name_len;
1556	u64 ino = btrfs_ino(inode);
1557
1558	key.objectid = ino;
1559	key.type = BTRFS_INODE_REF_KEY;
1560	key.offset = (u64)-1;
1561
1562	while (1) {
1563		ret = btrfs_search_slot(NULL, inode->root, &key, path, 0, 0);
1564		if (ret < 0)
1565			break;
1566		if (ret > 0) {
1567			if (path->slots[0] == 0)
1568				break;
1569			path->slots[0]--;
1570		}
1571process_slot:
1572		btrfs_item_key_to_cpu(path->nodes[0], &key,
1573				      path->slots[0]);
1574		if (key.objectid != ino ||
1575		    key.type != BTRFS_INODE_REF_KEY)
1576			break;
1577		ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1578		ptr_end = ptr + btrfs_item_size(path->nodes[0],
1579						   path->slots[0]);
1580		while (ptr < ptr_end) {
1581			struct btrfs_inode_ref *ref;
1582
1583			ref = (struct btrfs_inode_ref *)ptr;
1584			name_len = btrfs_inode_ref_name_len(path->nodes[0],
1585							    ref);
1586			ptr = (unsigned long)(ref + 1) + name_len;
1587			nlink++;
1588		}
1589
1590		if (key.offset == 0)
1591			break;
1592		if (path->slots[0] > 0) {
1593			path->slots[0]--;
1594			goto process_slot;
1595		}
1596		key.offset--;
1597		btrfs_release_path(path);
1598	}
1599	btrfs_release_path(path);
1600
1601	return nlink;
1602}
1603
1604/*
1605 * There are a few corners where the link count of the file can't
1606 * be properly maintained during replay.  So, instead of adding
1607 * lots of complexity to the log code, we just scan the backrefs
1608 * for any file that has been through replay.
1609 *
1610 * The scan will update the link count on the inode to reflect the
1611 * number of back refs found.  If it goes down to zero, the iput
1612 * will free the inode.
1613 */
1614static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
1615					   struct inode *inode)
1616{
1617	struct btrfs_root *root = BTRFS_I(inode)->root;
1618	struct btrfs_path *path;
1619	int ret;
1620	u64 nlink = 0;
1621	u64 ino = btrfs_ino(BTRFS_I(inode));
1622
1623	path = btrfs_alloc_path();
1624	if (!path)
1625		return -ENOMEM;
1626
1627	ret = count_inode_refs(BTRFS_I(inode), path);
1628	if (ret < 0)
1629		goto out;
1630
1631	nlink = ret;
1632
1633	ret = count_inode_extrefs(BTRFS_I(inode), path);
1634	if (ret < 0)
1635		goto out;
1636
1637	nlink += ret;
1638
1639	ret = 0;
1640
1641	if (nlink != inode->i_nlink) {
1642		set_nlink(inode, nlink);
1643		ret = btrfs_update_inode(trans, BTRFS_I(inode));
1644		if (ret)
1645			goto out;
1646	}
1647	if (S_ISDIR(inode->i_mode))
1648		BTRFS_I(inode)->index_cnt = (u64)-1;
1649
1650	if (inode->i_nlink == 0) {
1651		if (S_ISDIR(inode->i_mode)) {
1652			ret = replay_dir_deletes(trans, root, NULL, path,
1653						 ino, 1);
1654			if (ret)
1655				goto out;
1656		}
1657		ret = btrfs_insert_orphan_item(trans, root, ino);
1658		if (ret == -EEXIST)
1659			ret = 0;
1660	}
1661
1662out:
1663	btrfs_free_path(path);
1664	return ret;
1665}
1666
1667static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1668					    struct btrfs_root *root,
1669					    struct btrfs_path *path)
1670{
1671	int ret;
1672	struct btrfs_key key;
1673	struct inode *inode;
1674
1675	key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1676	key.type = BTRFS_ORPHAN_ITEM_KEY;
1677	key.offset = (u64)-1;
1678	while (1) {
1679		ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1680		if (ret < 0)
1681			break;
1682
1683		if (ret == 1) {
1684			ret = 0;
1685			if (path->slots[0] == 0)
1686				break;
1687			path->slots[0]--;
1688		}
1689
1690		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1691		if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1692		    key.type != BTRFS_ORPHAN_ITEM_KEY)
1693			break;
1694
1695		ret = btrfs_del_item(trans, root, path);
1696		if (ret)
1697			break;
1698
1699		btrfs_release_path(path);
1700		inode = read_one_inode(root, key.offset);
1701		if (!inode) {
1702			ret = -EIO;
1703			break;
1704		}
1705
1706		ret = fixup_inode_link_count(trans, inode);
1707		iput(inode);
1708		if (ret)
1709			break;
1710
1711		/*
1712		 * fixup on a directory may create new entries,
1713		 * make sure we always look for the highset possible
1714		 * offset
1715		 */
1716		key.offset = (u64)-1;
1717	}
1718	btrfs_release_path(path);
1719	return ret;
1720}
1721
1722
1723/*
1724 * record a given inode in the fixup dir so we can check its link
1725 * count when replay is done.  The link count is incremented here
1726 * so the inode won't go away until we check it
1727 */
1728static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1729				      struct btrfs_root *root,
1730				      struct btrfs_path *path,
1731				      u64 objectid)
1732{
1733	struct btrfs_key key;
1734	int ret = 0;
1735	struct inode *inode;
1736
1737	inode = read_one_inode(root, objectid);
1738	if (!inode)
1739		return -EIO;
1740
1741	key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1742	key.type = BTRFS_ORPHAN_ITEM_KEY;
1743	key.offset = objectid;
1744
1745	ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1746
1747	btrfs_release_path(path);
1748	if (ret == 0) {
1749		if (!inode->i_nlink)
1750			set_nlink(inode, 1);
1751		else
1752			inc_nlink(inode);
1753		ret = btrfs_update_inode(trans, BTRFS_I(inode));
1754	} else if (ret == -EEXIST) {
1755		ret = 0;
1756	}
1757	iput(inode);
1758
1759	return ret;
1760}
1761
1762/*
1763 * when replaying the log for a directory, we only insert names
1764 * for inodes that actually exist.  This means an fsync on a directory
1765 * does not implicitly fsync all the new files in it
1766 */
1767static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1768				    struct btrfs_root *root,
1769				    u64 dirid, u64 index,
1770				    const struct fscrypt_str *name,
1771				    struct btrfs_key *location)
1772{
1773	struct inode *inode;
1774	struct inode *dir;
1775	int ret;
1776
1777	inode = read_one_inode(root, location->objectid);
1778	if (!inode)
1779		return -ENOENT;
1780
1781	dir = read_one_inode(root, dirid);
1782	if (!dir) {
1783		iput(inode);
1784		return -EIO;
1785	}
1786
1787	ret = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode), name,
1788			     1, index);
1789
1790	/* FIXME, put inode into FIXUP list */
1791
1792	iput(inode);
1793	iput(dir);
1794	return ret;
1795}
1796
1797static int delete_conflicting_dir_entry(struct btrfs_trans_handle *trans,
1798					struct btrfs_inode *dir,
1799					struct btrfs_path *path,
1800					struct btrfs_dir_item *dst_di,
1801					const struct btrfs_key *log_key,
1802					u8 log_flags,
1803					bool exists)
1804{
1805	struct btrfs_key found_key;
1806
1807	btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1808	/* The existing dentry points to the same inode, don't delete it. */
1809	if (found_key.objectid == log_key->objectid &&
1810	    found_key.type == log_key->type &&
1811	    found_key.offset == log_key->offset &&
1812	    btrfs_dir_flags(path->nodes[0], dst_di) == log_flags)
1813		return 1;
1814
1815	/*
1816	 * Don't drop the conflicting directory entry if the inode for the new
1817	 * entry doesn't exist.
1818	 */
1819	if (!exists)
1820		return 0;
1821
1822	return drop_one_dir_item(trans, path, dir, dst_di);
1823}
1824
1825/*
1826 * take a single entry in a log directory item and replay it into
1827 * the subvolume.
1828 *
1829 * if a conflicting item exists in the subdirectory already,
1830 * the inode it points to is unlinked and put into the link count
1831 * fix up tree.
1832 *
1833 * If a name from the log points to a file or directory that does
1834 * not exist in the FS, it is skipped.  fsyncs on directories
1835 * do not force down inodes inside that directory, just changes to the
1836 * names or unlinks in a directory.
1837 *
1838 * Returns < 0 on error, 0 if the name wasn't replayed (dentry points to a
1839 * non-existing inode) and 1 if the name was replayed.
1840 */
1841static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1842				    struct btrfs_root *root,
1843				    struct btrfs_path *path,
1844				    struct extent_buffer *eb,
1845				    struct btrfs_dir_item *di,
1846				    struct btrfs_key *key)
1847{
1848	struct fscrypt_str name = { 0 };
1849	struct btrfs_dir_item *dir_dst_di;
1850	struct btrfs_dir_item *index_dst_di;
1851	bool dir_dst_matches = false;
1852	bool index_dst_matches = false;
1853	struct btrfs_key log_key;
1854	struct btrfs_key search_key;
1855	struct inode *dir;
1856	u8 log_flags;
1857	bool exists;
1858	int ret;
1859	bool update_size = true;
1860	bool name_added = false;
1861
1862	dir = read_one_inode(root, key->objectid);
1863	if (!dir)
1864		return -EIO;
1865
1866	ret = read_alloc_one_name(eb, di + 1, btrfs_dir_name_len(eb, di), &name);
1867	if (ret)
1868		goto out;
1869
1870	log_flags = btrfs_dir_flags(eb, di);
1871	btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1872	ret = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1873	btrfs_release_path(path);
1874	if (ret < 0)
1875		goto out;
1876	exists = (ret == 0);
1877	ret = 0;
1878
1879	dir_dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1880					   &name, 1);
1881	if (IS_ERR(dir_dst_di)) {
1882		ret = PTR_ERR(dir_dst_di);
1883		goto out;
1884	} else if (dir_dst_di) {
1885		ret = delete_conflicting_dir_entry(trans, BTRFS_I(dir), path,
1886						   dir_dst_di, &log_key,
1887						   log_flags, exists);
1888		if (ret < 0)
1889			goto out;
1890		dir_dst_matches = (ret == 1);
1891	}
1892
1893	btrfs_release_path(path);
1894
1895	index_dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1896						   key->objectid, key->offset,
1897						   &name, 1);
1898	if (IS_ERR(index_dst_di)) {
1899		ret = PTR_ERR(index_dst_di);
1900		goto out;
1901	} else if (index_dst_di) {
1902		ret = delete_conflicting_dir_entry(trans, BTRFS_I(dir), path,
1903						   index_dst_di, &log_key,
1904						   log_flags, exists);
1905		if (ret < 0)
1906			goto out;
1907		index_dst_matches = (ret == 1);
1908	}
1909
1910	btrfs_release_path(path);
1911
1912	if (dir_dst_matches && index_dst_matches) {
1913		ret = 0;
1914		update_size = false;
1915		goto out;
1916	}
1917
1918	/*
1919	 * Check if the inode reference exists in the log for the given name,
1920	 * inode and parent inode
1921	 */
1922	search_key.objectid = log_key.objectid;
1923	search_key.type = BTRFS_INODE_REF_KEY;
1924	search_key.offset = key->objectid;
1925	ret = backref_in_log(root->log_root, &search_key, 0, &name);
1926	if (ret < 0) {
1927	        goto out;
1928	} else if (ret) {
1929	        /* The dentry will be added later. */
1930	        ret = 0;
1931	        update_size = false;
1932	        goto out;
1933	}
1934
1935	search_key.objectid = log_key.objectid;
1936	search_key.type = BTRFS_INODE_EXTREF_KEY;
1937	search_key.offset = key->objectid;
1938	ret = backref_in_log(root->log_root, &search_key, key->objectid, &name);
1939	if (ret < 0) {
1940		goto out;
1941	} else if (ret) {
1942		/* The dentry will be added later. */
1943		ret = 0;
1944		update_size = false;
1945		goto out;
1946	}
1947	btrfs_release_path(path);
1948	ret = insert_one_name(trans, root, key->objectid, key->offset,
1949			      &name, &log_key);
1950	if (ret && ret != -ENOENT && ret != -EEXIST)
1951		goto out;
1952	if (!ret)
1953		name_added = true;
1954	update_size = false;
1955	ret = 0;
1956
1957out:
1958	if (!ret && update_size) {
1959		btrfs_i_size_write(BTRFS_I(dir), dir->i_size + name.len * 2);
1960		ret = btrfs_update_inode(trans, BTRFS_I(dir));
1961	}
1962	kfree(name.name);
1963	iput(dir);
1964	if (!ret && name_added)
1965		ret = 1;
1966	return ret;
1967}
1968
1969/* Replay one dir item from a BTRFS_DIR_INDEX_KEY key. */
1970static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1971					struct btrfs_root *root,
1972					struct btrfs_path *path,
1973					struct extent_buffer *eb, int slot,
1974					struct btrfs_key *key)
1975{
1976	int ret;
1977	struct btrfs_dir_item *di;
1978
1979	/* We only log dir index keys, which only contain a single dir item. */
1980	ASSERT(key->type == BTRFS_DIR_INDEX_KEY);
1981
1982	di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1983	ret = replay_one_name(trans, root, path, eb, di, key);
1984	if (ret < 0)
1985		return ret;
1986
1987	/*
1988	 * If this entry refers to a non-directory (directories can not have a
1989	 * link count > 1) and it was added in the transaction that was not
1990	 * committed, make sure we fixup the link count of the inode the entry
1991	 * points to. Otherwise something like the following would result in a
1992	 * directory pointing to an inode with a wrong link that does not account
1993	 * for this dir entry:
1994	 *
1995	 * mkdir testdir
1996	 * touch testdir/foo
1997	 * touch testdir/bar
1998	 * sync
1999	 *
2000	 * ln testdir/bar testdir/bar_link
2001	 * ln testdir/foo testdir/foo_link
2002	 * xfs_io -c "fsync" testdir/bar
2003	 *
2004	 * <power failure>
2005	 *
2006	 * mount fs, log replay happens
2007	 *
2008	 * File foo would remain with a link count of 1 when it has two entries
2009	 * pointing to it in the directory testdir. This would make it impossible
2010	 * to ever delete the parent directory has it would result in stale
2011	 * dentries that can never be deleted.
2012	 */
2013	if (ret == 1 && btrfs_dir_ftype(eb, di) != BTRFS_FT_DIR) {
2014		struct btrfs_path *fixup_path;
2015		struct btrfs_key di_key;
2016
2017		fixup_path = btrfs_alloc_path();
2018		if (!fixup_path)
2019			return -ENOMEM;
2020
2021		btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2022		ret = link_to_fixup_dir(trans, root, fixup_path, di_key.objectid);
2023		btrfs_free_path(fixup_path);
2024	}
2025
2026	return ret;
2027}
2028
2029/*
2030 * directory replay has two parts.  There are the standard directory
2031 * items in the log copied from the subvolume, and range items
2032 * created in the log while the subvolume was logged.
2033 *
2034 * The range items tell us which parts of the key space the log
2035 * is authoritative for.  During replay, if a key in the subvolume
2036 * directory is in a logged range item, but not actually in the log
2037 * that means it was deleted from the directory before the fsync
2038 * and should be removed.
2039 */
2040static noinline int find_dir_range(struct btrfs_root *root,
2041				   struct btrfs_path *path,
2042				   u64 dirid,
2043				   u64 *start_ret, u64 *end_ret)
2044{
2045	struct btrfs_key key;
2046	u64 found_end;
2047	struct btrfs_dir_log_item *item;
2048	int ret;
2049	int nritems;
2050
2051	if (*start_ret == (u64)-1)
2052		return 1;
2053
2054	key.objectid = dirid;
2055	key.type = BTRFS_DIR_LOG_INDEX_KEY;
2056	key.offset = *start_ret;
2057
2058	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2059	if (ret < 0)
2060		goto out;
2061	if (ret > 0) {
2062		if (path->slots[0] == 0)
2063			goto out;
2064		path->slots[0]--;
2065	}
2066	if (ret != 0)
2067		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2068
2069	if (key.type != BTRFS_DIR_LOG_INDEX_KEY || key.objectid != dirid) {
2070		ret = 1;
2071		goto next;
2072	}
2073	item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2074			      struct btrfs_dir_log_item);
2075	found_end = btrfs_dir_log_end(path->nodes[0], item);
2076
2077	if (*start_ret >= key.offset && *start_ret <= found_end) {
2078		ret = 0;
2079		*start_ret = key.offset;
2080		*end_ret = found_end;
2081		goto out;
2082	}
2083	ret = 1;
2084next:
2085	/* check the next slot in the tree to see if it is a valid item */
2086	nritems = btrfs_header_nritems(path->nodes[0]);
2087	path->slots[0]++;
2088	if (path->slots[0] >= nritems) {
2089		ret = btrfs_next_leaf(root, path);
2090		if (ret)
2091			goto out;
2092	}
2093
2094	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2095
2096	if (key.type != BTRFS_DIR_LOG_INDEX_KEY || key.objectid != dirid) {
2097		ret = 1;
2098		goto out;
2099	}
2100	item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2101			      struct btrfs_dir_log_item);
2102	found_end = btrfs_dir_log_end(path->nodes[0], item);
2103	*start_ret = key.offset;
2104	*end_ret = found_end;
2105	ret = 0;
2106out:
2107	btrfs_release_path(path);
2108	return ret;
2109}
2110
2111/*
2112 * this looks for a given directory item in the log.  If the directory
2113 * item is not in the log, the item is removed and the inode it points
2114 * to is unlinked
2115 */
2116static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
2117				      struct btrfs_root *log,
2118				      struct btrfs_path *path,
2119				      struct btrfs_path *log_path,
2120				      struct inode *dir,
2121				      struct btrfs_key *dir_key)
2122{
2123	struct btrfs_root *root = BTRFS_I(dir)->root;
2124	int ret;
2125	struct extent_buffer *eb;
2126	int slot;
2127	struct btrfs_dir_item *di;
2128	struct fscrypt_str name = { 0 };
2129	struct inode *inode = NULL;
2130	struct btrfs_key location;
2131
2132	/*
2133	 * Currently we only log dir index keys. Even if we replay a log created
2134	 * by an older kernel that logged both dir index and dir item keys, all
2135	 * we need to do is process the dir index keys, we (and our caller) can
2136	 * safely ignore dir item keys (key type BTRFS_DIR_ITEM_KEY).
2137	 */
2138	ASSERT(dir_key->type == BTRFS_DIR_INDEX_KEY);
2139
2140	eb = path->nodes[0];
2141	slot = path->slots[0];
2142	di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
2143	ret = read_alloc_one_name(eb, di + 1, btrfs_dir_name_len(eb, di), &name);
2144	if (ret)
2145		goto out;
2146
2147	if (log) {
2148		struct btrfs_dir_item *log_di;
2149
2150		log_di = btrfs_lookup_dir_index_item(trans, log, log_path,
2151						     dir_key->objectid,
2152						     dir_key->offset, &name, 0);
2153		if (IS_ERR(log_di)) {
2154			ret = PTR_ERR(log_di);
2155			goto out;
2156		} else if (log_di) {
2157			/* The dentry exists in the log, we have nothing to do. */
2158			ret = 0;
2159			goto out;
2160		}
2161	}
2162
2163	btrfs_dir_item_key_to_cpu(eb, di, &location);
2164	btrfs_release_path(path);
2165	btrfs_release_path(log_path);
2166	inode = read_one_inode(root, location.objectid);
2167	if (!inode) {
2168		ret = -EIO;
2169		goto out;
2170	}
2171
2172	ret = link_to_fixup_dir(trans, root, path, location.objectid);
2173	if (ret)
2174		goto out;
2175
2176	inc_nlink(inode);
2177	ret = unlink_inode_for_log_replay(trans, BTRFS_I(dir), BTRFS_I(inode),
2178					  &name);
2179	/*
2180	 * Unlike dir item keys, dir index keys can only have one name (entry) in
2181	 * them, as there are no key collisions since each key has a unique offset
2182	 * (an index number), so we're done.
2183	 */
2184out:
2185	btrfs_release_path(path);
2186	btrfs_release_path(log_path);
2187	kfree(name.name);
2188	iput(inode);
2189	return ret;
2190}
2191
2192static int replay_xattr_deletes(struct btrfs_trans_handle *trans,
2193			      struct btrfs_root *root,
2194			      struct btrfs_root *log,
2195			      struct btrfs_path *path,
2196			      const u64 ino)
2197{
2198	struct btrfs_key search_key;
2199	struct btrfs_path *log_path;
2200	int i;
2201	int nritems;
2202	int ret;
2203
2204	log_path = btrfs_alloc_path();
2205	if (!log_path)
2206		return -ENOMEM;
2207
2208	search_key.objectid = ino;
2209	search_key.type = BTRFS_XATTR_ITEM_KEY;
2210	search_key.offset = 0;
2211again:
2212	ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
2213	if (ret < 0)
2214		goto out;
2215process_leaf:
2216	nritems = btrfs_header_nritems(path->nodes[0]);
2217	for (i = path->slots[0]; i < nritems; i++) {
2218		struct btrfs_key key;
2219		struct btrfs_dir_item *di;
2220		struct btrfs_dir_item *log_di;
2221		u32 total_size;
2222		u32 cur;
2223
2224		btrfs_item_key_to_cpu(path->nodes[0], &key, i);
2225		if (key.objectid != ino || key.type != BTRFS_XATTR_ITEM_KEY) {
2226			ret = 0;
2227			goto out;
2228		}
2229
2230		di = btrfs_item_ptr(path->nodes[0], i, struct btrfs_dir_item);
2231		total_size = btrfs_item_size(path->nodes[0], i);
2232		cur = 0;
2233		while (cur < total_size) {
2234			u16 name_len = btrfs_dir_name_len(path->nodes[0], di);
2235			u16 data_len = btrfs_dir_data_len(path->nodes[0], di);
2236			u32 this_len = sizeof(*di) + name_len + data_len;
2237			char *name;
2238
2239			name = kmalloc(name_len, GFP_NOFS);
2240			if (!name) {
2241				ret = -ENOMEM;
2242				goto out;
2243			}
2244			read_extent_buffer(path->nodes[0], name,
2245					   (unsigned long)(di + 1), name_len);
2246
2247			log_di = btrfs_lookup_xattr(NULL, log, log_path, ino,
2248						    name, name_len, 0);
2249			btrfs_release_path(log_path);
2250			if (!log_di) {
2251				/* Doesn't exist in log tree, so delete it. */
2252				btrfs_release_path(path);
2253				di = btrfs_lookup_xattr(trans, root, path, ino,
2254							name, name_len, -1);
2255				kfree(name);
2256				if (IS_ERR(di)) {
2257					ret = PTR_ERR(di);
2258					goto out;
2259				}
2260				ASSERT(di);
2261				ret = btrfs_delete_one_dir_name(trans, root,
2262								path, di);
2263				if (ret)
2264					goto out;
2265				btrfs_release_path(path);
2266				search_key = key;
2267				goto again;
2268			}
2269			kfree(name);
2270			if (IS_ERR(log_di)) {
2271				ret = PTR_ERR(log_di);
2272				goto out;
2273			}
2274			cur += this_len;
2275			di = (struct btrfs_dir_item *)((char *)di + this_len);
2276		}
2277	}
2278	ret = btrfs_next_leaf(root, path);
2279	if (ret > 0)
2280		ret = 0;
2281	else if (ret == 0)
2282		goto process_leaf;
2283out:
2284	btrfs_free_path(log_path);
2285	btrfs_release_path(path);
2286	return ret;
2287}
2288
2289
2290/*
2291 * deletion replay happens before we copy any new directory items
2292 * out of the log or out of backreferences from inodes.  It
2293 * scans the log to find ranges of keys that log is authoritative for,
2294 * and then scans the directory to find items in those ranges that are
2295 * not present in the log.
2296 *
2297 * Anything we don't find in the log is unlinked and removed from the
2298 * directory.
2299 */
2300static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
2301				       struct btrfs_root *root,
2302				       struct btrfs_root *log,
2303				       struct btrfs_path *path,
2304				       u64 dirid, int del_all)
2305{
2306	u64 range_start;
2307	u64 range_end;
2308	int ret = 0;
2309	struct btrfs_key dir_key;
2310	struct btrfs_key found_key;
2311	struct btrfs_path *log_path;
2312	struct inode *dir;
2313
2314	dir_key.objectid = dirid;
2315	dir_key.type = BTRFS_DIR_INDEX_KEY;
2316	log_path = btrfs_alloc_path();
2317	if (!log_path)
2318		return -ENOMEM;
2319
2320	dir = read_one_inode(root, dirid);
2321	/* it isn't an error if the inode isn't there, that can happen
2322	 * because we replay the deletes before we copy in the inode item
2323	 * from the log
2324	 */
2325	if (!dir) {
2326		btrfs_free_path(log_path);
2327		return 0;
2328	}
2329
2330	range_start = 0;
2331	range_end = 0;
2332	while (1) {
2333		if (del_all)
2334			range_end = (u64)-1;
2335		else {
2336			ret = find_dir_range(log, path, dirid,
2337					     &range_start, &range_end);
2338			if (ret < 0)
2339				goto out;
2340			else if (ret > 0)
2341				break;
2342		}
2343
2344		dir_key.offset = range_start;
2345		while (1) {
2346			int nritems;
2347			ret = btrfs_search_slot(NULL, root, &dir_key, path,
2348						0, 0);
2349			if (ret < 0)
2350				goto out;
2351
2352			nritems = btrfs_header_nritems(path->nodes[0]);
2353			if (path->slots[0] >= nritems) {
2354				ret = btrfs_next_leaf(root, path);
2355				if (ret == 1)
2356					break;
2357				else if (ret < 0)
2358					goto out;
2359			}
2360			btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2361					      path->slots[0]);
2362			if (found_key.objectid != dirid ||
2363			    found_key.type != dir_key.type) {
2364				ret = 0;
2365				goto out;
2366			}
2367
2368			if (found_key.offset > range_end)
2369				break;
2370
2371			ret = check_item_in_log(trans, log, path,
2372						log_path, dir,
2373						&found_key);
2374			if (ret)
2375				goto out;
2376			if (found_key.offset == (u64)-1)
2377				break;
2378			dir_key.offset = found_key.offset + 1;
2379		}
2380		btrfs_release_path(path);
2381		if (range_end == (u64)-1)
2382			break;
2383		range_start = range_end + 1;
2384	}
2385	ret = 0;
2386out:
2387	btrfs_release_path(path);
2388	btrfs_free_path(log_path);
2389	iput(dir);
2390	return ret;
2391}
2392
2393/*
2394 * the process_func used to replay items from the log tree.  This
2395 * gets called in two different stages.  The first stage just looks
2396 * for inodes and makes sure they are all copied into the subvolume.
2397 *
2398 * The second stage copies all the other item types from the log into
2399 * the subvolume.  The two stage approach is slower, but gets rid of
2400 * lots of complexity around inodes referencing other inodes that exist
2401 * only in the log (references come from either directory items or inode
2402 * back refs).
2403 */
2404static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
2405			     struct walk_control *wc, u64 gen, int level)
2406{
2407	int nritems;
2408	struct btrfs_tree_parent_check check = {
2409		.transid = gen,
2410		.level = level
2411	};
2412	struct btrfs_path *path;
2413	struct btrfs_root *root = wc->replay_dest;
2414	struct btrfs_key key;
2415	int i;
2416	int ret;
2417
2418	ret = btrfs_read_extent_buffer(eb, &check);
2419	if (ret)
2420		return ret;
2421
2422	level = btrfs_header_level(eb);
2423
2424	if (level != 0)
2425		return 0;
2426
2427	path = btrfs_alloc_path();
2428	if (!path)
2429		return -ENOMEM;
2430
2431	nritems = btrfs_header_nritems(eb);
2432	for (i = 0; i < nritems; i++) {
2433		btrfs_item_key_to_cpu(eb, &key, i);
2434
2435		/* inode keys are done during the first stage */
2436		if (key.type == BTRFS_INODE_ITEM_KEY &&
2437		    wc->stage == LOG_WALK_REPLAY_INODES) {
2438			struct btrfs_inode_item *inode_item;
2439			u32 mode;
2440
2441			inode_item = btrfs_item_ptr(eb, i,
2442					    struct btrfs_inode_item);
2443			/*
2444			 * If we have a tmpfile (O_TMPFILE) that got fsync'ed
2445			 * and never got linked before the fsync, skip it, as
2446			 * replaying it is pointless since it would be deleted
2447			 * later. We skip logging tmpfiles, but it's always
2448			 * possible we are replaying a log created with a kernel
2449			 * that used to log tmpfiles.
2450			 */
2451			if (btrfs_inode_nlink(eb, inode_item) == 0) {
2452				wc->ignore_cur_inode = true;
2453				continue;
2454			} else {
2455				wc->ignore_cur_inode = false;
2456			}
2457			ret = replay_xattr_deletes(wc->trans, root, log,
2458						   path, key.objectid);
2459			if (ret)
2460				break;
2461			mode = btrfs_inode_mode(eb, inode_item);
2462			if (S_ISDIR(mode)) {
2463				ret = replay_dir_deletes(wc->trans,
2464					 root, log, path, key.objectid, 0);
2465				if (ret)
2466					break;
2467			}
2468			ret = overwrite_item(wc->trans, root, path,
2469					     eb, i, &key);
2470			if (ret)
2471				break;
2472
2473			/*
2474			 * Before replaying extents, truncate the inode to its
2475			 * size. We need to do it now and not after log replay
2476			 * because before an fsync we can have prealloc extents
2477			 * added beyond the inode's i_size. If we did it after,
2478			 * through orphan cleanup for example, we would drop
2479			 * those prealloc extents just after replaying them.
2480			 */
2481			if (S_ISREG(mode)) {
2482				struct btrfs_drop_extents_args drop_args = { 0 };
2483				struct inode *inode;
2484				u64 from;
2485
2486				inode = read_one_inode(root, key.objectid);
2487				if (!inode) {
2488					ret = -EIO;
2489					break;
2490				}
2491				from = ALIGN(i_size_read(inode),
2492					     root->fs_info->sectorsize);
2493				drop_args.start = from;
2494				drop_args.end = (u64)-1;
2495				drop_args.drop_cache = true;
2496				ret = btrfs_drop_extents(wc->trans, root,
2497							 BTRFS_I(inode),
2498							 &drop_args);
2499				if (!ret) {
2500					inode_sub_bytes(inode,
2501							drop_args.bytes_found);
2502					/* Update the inode's nbytes. */
2503					ret = btrfs_update_inode(wc->trans,
2504								 BTRFS_I(inode));
2505				}
2506				iput(inode);
2507				if (ret)
2508					break;
2509			}
2510
2511			ret = link_to_fixup_dir(wc->trans, root,
2512						path, key.objectid);
2513			if (ret)
2514				break;
2515		}
2516
2517		if (wc->ignore_cur_inode)
2518			continue;
2519
2520		if (key.type == BTRFS_DIR_INDEX_KEY &&
2521		    wc->stage == LOG_WALK_REPLAY_DIR_INDEX) {
2522			ret = replay_one_dir_item(wc->trans, root, path,
2523						  eb, i, &key);
2524			if (ret)
2525				break;
2526		}
2527
2528		if (wc->stage < LOG_WALK_REPLAY_ALL)
2529			continue;
2530
2531		/* these keys are simply copied */
2532		if (key.type == BTRFS_XATTR_ITEM_KEY) {
2533			ret = overwrite_item(wc->trans, root, path,
2534					     eb, i, &key);
2535			if (ret)
2536				break;
2537		} else if (key.type == BTRFS_INODE_REF_KEY ||
2538			   key.type == BTRFS_INODE_EXTREF_KEY) {
2539			ret = add_inode_ref(wc->trans, root, log, path,
2540					    eb, i, &key);
2541			if (ret && ret != -ENOENT)
2542				break;
2543			ret = 0;
2544		} else if (key.type == BTRFS_EXTENT_DATA_KEY) {
2545			ret = replay_one_extent(wc->trans, root, path,
2546						eb, i, &key);
2547			if (ret)
2548				break;
2549		}
2550		/*
2551		 * We don't log BTRFS_DIR_ITEM_KEY keys anymore, only the
2552		 * BTRFS_DIR_INDEX_KEY items which we use to derive the
2553		 * BTRFS_DIR_ITEM_KEY items. If we are replaying a log from an
2554		 * older kernel with such keys, ignore them.
2555		 */
2556	}
2557	btrfs_free_path(path);
2558	return ret;
2559}
2560
2561/*
2562 * Correctly adjust the reserved bytes occupied by a log tree extent buffer
2563 */
2564static void unaccount_log_buffer(struct btrfs_fs_info *fs_info, u64 start)
2565{
2566	struct btrfs_block_group *cache;
2567
2568	cache = btrfs_lookup_block_group(fs_info, start);
2569	if (!cache) {
2570		btrfs_err(fs_info, "unable to find block group for %llu", start);
2571		return;
2572	}
2573
2574	spin_lock(&cache->space_info->lock);
2575	spin_lock(&cache->lock);
2576	cache->reserved -= fs_info->nodesize;
2577	cache->space_info->bytes_reserved -= fs_info->nodesize;
2578	spin_unlock(&cache->lock);
2579	spin_unlock(&cache->space_info->lock);
2580
2581	btrfs_put_block_group(cache);
2582}
2583
2584static int clean_log_buffer(struct btrfs_trans_handle *trans,
2585			    struct extent_buffer *eb)
2586{
2587	int ret;
2588
2589	btrfs_tree_lock(eb);
2590	btrfs_clear_buffer_dirty(trans, eb);
2591	wait_on_extent_buffer_writeback(eb);
2592	btrfs_tree_unlock(eb);
2593
2594	if (trans) {
2595		ret = btrfs_pin_reserved_extent(trans, eb);
2596		if (ret)
2597			return ret;
2598	} else {
2599		unaccount_log_buffer(eb->fs_info, eb->start);
2600	}
2601
2602	return 0;
2603}
2604
2605static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
2606				   struct btrfs_root *root,
2607				   struct btrfs_path *path, int *level,
2608				   struct walk_control *wc)
2609{
2610	struct btrfs_fs_info *fs_info = root->fs_info;
2611	u64 bytenr;
2612	u64 ptr_gen;
2613	struct extent_buffer *next;
2614	struct extent_buffer *cur;
2615	int ret = 0;
2616
2617	while (*level > 0) {
2618		struct btrfs_tree_parent_check check = { 0 };
2619
2620		cur = path->nodes[*level];
2621
2622		WARN_ON(btrfs_header_level(cur) != *level);
2623
2624		if (path->slots[*level] >=
2625		    btrfs_header_nritems(cur))
2626			break;
2627
2628		bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
2629		ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
2630		check.transid = ptr_gen;
2631		check.level = *level - 1;
2632		check.has_first_key = true;
2633		btrfs_node_key_to_cpu(cur, &check.first_key, path->slots[*level]);
2634
2635		next = btrfs_find_create_tree_block(fs_info, bytenr,
2636						    btrfs_header_owner(cur),
2637						    *level - 1);
2638		if (IS_ERR(next))
2639			return PTR_ERR(next);
2640
2641		if (*level == 1) {
2642			ret = wc->process_func(root, next, wc, ptr_gen,
2643					       *level - 1);
2644			if (ret) {
2645				free_extent_buffer(next);
2646				return ret;
2647			}
2648
2649			path->slots[*level]++;
2650			if (wc->free) {
2651				ret = btrfs_read_extent_buffer(next, &check);
2652				if (ret) {
2653					free_extent_buffer(next);
2654					return ret;
2655				}
2656
2657				ret = clean_log_buffer(trans, next);
2658				if (ret) {
2659					free_extent_buffer(next);
2660					return ret;
2661				}
2662			}
2663			free_extent_buffer(next);
2664			continue;
2665		}
2666		ret = btrfs_read_extent_buffer(next, &check);
2667		if (ret) {
2668			free_extent_buffer(next);
2669			return ret;
2670		}
2671
2672		if (path->nodes[*level-1])
2673			free_extent_buffer(path->nodes[*level-1]);
2674		path->nodes[*level-1] = next;
2675		*level = btrfs_header_level(next);
2676		path->slots[*level] = 0;
2677		cond_resched();
2678	}
2679	path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
2680
2681	cond_resched();
2682	return 0;
2683}
2684
2685static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
2686				 struct btrfs_root *root,
2687				 struct btrfs_path *path, int *level,
2688				 struct walk_control *wc)
2689{
2690	int i;
2691	int slot;
2692	int ret;
2693
2694	for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
2695		slot = path->slots[i];
2696		if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
2697			path->slots[i]++;
2698			*level = i;
2699			WARN_ON(*level == 0);
2700			return 0;
2701		} else {
2702			ret = wc->process_func(root, path->nodes[*level], wc,
2703				 btrfs_header_generation(path->nodes[*level]),
2704				 *level);
2705			if (ret)
2706				return ret;
2707
2708			if (wc->free) {
2709				ret = clean_log_buffer(trans, path->nodes[*level]);
2710				if (ret)
2711					return ret;
2712			}
2713			free_extent_buffer(path->nodes[*level]);
2714			path->nodes[*level] = NULL;
2715			*level = i + 1;
2716		}
2717	}
2718	return 1;
2719}
2720
2721/*
2722 * drop the reference count on the tree rooted at 'snap'.  This traverses
2723 * the tree freeing any blocks that have a ref count of zero after being
2724 * decremented.
2725 */
2726static int walk_log_tree(struct btrfs_trans_handle *trans,
2727			 struct btrfs_root *log, struct walk_control *wc)
2728{
2729	int ret = 0;
2730	int wret;
2731	int level;
2732	struct btrfs_path *path;
2733	int orig_level;
2734
2735	path = btrfs_alloc_path();
2736	if (!path)
2737		return -ENOMEM;
2738
2739	level = btrfs_header_level(log->node);
2740	orig_level = level;
2741	path->nodes[level] = log->node;
2742	atomic_inc(&log->node->refs);
2743	path->slots[level] = 0;
2744
2745	while (1) {
2746		wret = walk_down_log_tree(trans, log, path, &level, wc);
2747		if (wret > 0)
2748			break;
2749		if (wret < 0) {
2750			ret = wret;
2751			goto out;
2752		}
2753
2754		wret = walk_up_log_tree(trans, log, path, &level, wc);
2755		if (wret > 0)
2756			break;
2757		if (wret < 0) {
2758			ret = wret;
2759			goto out;
2760		}
2761	}
2762
2763	/* was the root node processed? if not, catch it here */
2764	if (path->nodes[orig_level]) {
2765		ret = wc->process_func(log, path->nodes[orig_level], wc,
2766			 btrfs_header_generation(path->nodes[orig_level]),
2767			 orig_level);
2768		if (ret)
2769			goto out;
2770		if (wc->free)
2771			ret = clean_log_buffer(trans, path->nodes[orig_level]);
2772	}
2773
2774out:
2775	btrfs_free_path(path);
2776	return ret;
2777}
2778
2779/*
2780 * helper function to update the item for a given subvolumes log root
2781 * in the tree of log roots
2782 */
2783static int update_log_root(struct btrfs_trans_handle *trans,
2784			   struct btrfs_root *log,
2785			   struct btrfs_root_item *root_item)
2786{
2787	struct btrfs_fs_info *fs_info = log->fs_info;
2788	int ret;
2789
2790	if (log->log_transid == 1) {
2791		/* insert root item on the first sync */
2792		ret = btrfs_insert_root(trans, fs_info->log_root_tree,
2793				&log->root_key, root_item);
2794	} else {
2795		ret = btrfs_update_root(trans, fs_info->log_root_tree,
2796				&log->root_key, root_item);
2797	}
2798	return ret;
2799}
2800
2801static void wait_log_commit(struct btrfs_root *root, int transid)
2802{
2803	DEFINE_WAIT(wait);
2804	int index = transid % 2;
2805
2806	/*
2807	 * we only allow two pending log transactions at a time,
2808	 * so we know that if ours is more than 2 older than the
2809	 * current transaction, we're done
2810	 */
2811	for (;;) {
2812		prepare_to_wait(&root->log_commit_wait[index],
2813				&wait, TASK_UNINTERRUPTIBLE);
2814
2815		if (!(root->log_transid_committed < transid &&
2816		      atomic_read(&root->log_commit[index])))
2817			break;
2818
2819		mutex_unlock(&root->log_mutex);
2820		schedule();
2821		mutex_lock(&root->log_mutex);
2822	}
2823	finish_wait(&root->log_commit_wait[index], &wait);
2824}
2825
2826static void wait_for_writer(struct btrfs_root *root)
2827{
2828	DEFINE_WAIT(wait);
2829
2830	for (;;) {
2831		prepare_to_wait(&root->log_writer_wait, &wait,
2832				TASK_UNINTERRUPTIBLE);
2833		if (!atomic_read(&root->log_writers))
2834			break;
2835
2836		mutex_unlock(&root->log_mutex);
2837		schedule();
2838		mutex_lock(&root->log_mutex);
2839	}
2840	finish_wait(&root->log_writer_wait, &wait);
2841}
2842
2843void btrfs_init_log_ctx(struct btrfs_log_ctx *ctx, struct btrfs_inode *inode)
2844{
2845	ctx->log_ret = 0;
2846	ctx->log_transid = 0;
2847	ctx->log_new_dentries = false;
2848	ctx->logging_new_name = false;
2849	ctx->logging_new_delayed_dentries = false;
2850	ctx->logged_before = false;
2851	ctx->inode = inode;
2852	INIT_LIST_HEAD(&ctx->list);
2853	INIT_LIST_HEAD(&ctx->ordered_extents);
2854	INIT_LIST_HEAD(&ctx->conflict_inodes);
2855	ctx->num_conflict_inodes = 0;
2856	ctx->logging_conflict_inodes = false;
2857	ctx->scratch_eb = NULL;
2858}
2859
2860void btrfs_init_log_ctx_scratch_eb(struct btrfs_log_ctx *ctx)
2861{
2862	struct btrfs_inode *inode = ctx->inode;
2863
2864	if (!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) &&
2865	    !test_bit(BTRFS_INODE_COPY_EVERYTHING, &inode->runtime_flags))
2866		return;
2867
2868	/*
2869	 * Don't care about allocation failure. This is just for optimization,
2870	 * if we fail to allocate here, we will try again later if needed.
2871	 */
2872	ctx->scratch_eb = alloc_dummy_extent_buffer(inode->root->fs_info, 0);
2873}
2874
2875void btrfs_release_log_ctx_extents(struct btrfs_log_ctx *ctx)
2876{
2877	struct btrfs_ordered_extent *ordered;
2878	struct btrfs_ordered_extent *tmp;
2879
2880	btrfs_assert_inode_locked(ctx->inode);
2881
2882	list_for_each_entry_safe(ordered, tmp, &ctx->ordered_extents, log_list) {
2883		list_del_init(&ordered->log_list);
2884		btrfs_put_ordered_extent(ordered);
2885	}
2886}
2887
2888
2889static inline void btrfs_remove_log_ctx(struct btrfs_root *root,
2890					struct btrfs_log_ctx *ctx)
2891{
2892	mutex_lock(&root->log_mutex);
2893	list_del_init(&ctx->list);
2894	mutex_unlock(&root->log_mutex);
2895}
2896
2897/* 
2898 * Invoked in log mutex context, or be sure there is no other task which
2899 * can access the list.
2900 */
2901static inline void btrfs_remove_all_log_ctxs(struct btrfs_root *root,
2902					     int index, int error)
2903{
2904	struct btrfs_log_ctx *ctx;
2905	struct btrfs_log_ctx *safe;
2906
2907	list_for_each_entry_safe(ctx, safe, &root->log_ctxs[index], list) {
2908		list_del_init(&ctx->list);
2909		ctx->log_ret = error;
2910	}
2911}
2912
2913/*
2914 * Sends a given tree log down to the disk and updates the super blocks to
2915 * record it.  When this call is done, you know that any inodes previously
2916 * logged are safely on disk only if it returns 0.
2917 *
2918 * Any other return value means you need to call btrfs_commit_transaction.
2919 * Some of the edge cases for fsyncing directories that have had unlinks
2920 * or renames done in the past mean that sometimes the only safe
2921 * fsync is to commit the whole FS.  When btrfs_sync_log returns -EAGAIN,
2922 * that has happened.
2923 */
2924int btrfs_sync_log(struct btrfs_trans_handle *trans,
2925		   struct btrfs_root *root, struct btrfs_log_ctx *ctx)
2926{
2927	int index1;
2928	int index2;
2929	int mark;
2930	int ret;
2931	struct btrfs_fs_info *fs_info = root->fs_info;
2932	struct btrfs_root *log = root->log_root;
2933	struct btrfs_root *log_root_tree = fs_info->log_root_tree;
2934	struct btrfs_root_item new_root_item;
2935	int log_transid = 0;
2936	struct btrfs_log_ctx root_log_ctx;
2937	struct blk_plug plug;
2938	u64 log_root_start;
2939	u64 log_root_level;
2940
2941	mutex_lock(&root->log_mutex);
2942	log_transid = ctx->log_transid;
2943	if (root->log_transid_committed >= log_transid) {
2944		mutex_unlock(&root->log_mutex);
2945		return ctx->log_ret;
2946	}
2947
2948	index1 = log_transid % 2;
2949	if (atomic_read(&root->log_commit[index1])) {
2950		wait_log_commit(root, log_transid);
2951		mutex_unlock(&root->log_mutex);
2952		return ctx->log_ret;
2953	}
2954	ASSERT(log_transid == root->log_transid);
2955	atomic_set(&root->log_commit[index1], 1);
2956
2957	/* wait for previous tree log sync to complete */
2958	if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
2959		wait_log_commit(root, log_transid - 1);
2960
2961	while (1) {
2962		int batch = atomic_read(&root->log_batch);
2963		/* when we're on an ssd, just kick the log commit out */
2964		if (!btrfs_test_opt(fs_info, SSD) &&
2965		    test_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state)) {
2966			mutex_unlock(&root->log_mutex);
2967			schedule_timeout_uninterruptible(1);
2968			mutex_lock(&root->log_mutex);
2969		}
2970		wait_for_writer(root);
2971		if (batch == atomic_read(&root->log_batch))
2972			break;
2973	}
2974
2975	/* bail out if we need to do a full commit */
2976	if (btrfs_need_log_full_commit(trans)) {
2977		ret = BTRFS_LOG_FORCE_COMMIT;
2978		mutex_unlock(&root->log_mutex);
2979		goto out;
2980	}
2981
2982	if (log_transid % 2 == 0)
2983		mark = EXTENT_DIRTY;
2984	else
2985		mark = EXTENT_NEW;
2986
2987	/* we start IO on  all the marked extents here, but we don't actually
2988	 * wait for them until later.
2989	 */
2990	blk_start_plug(&plug);
2991	ret = btrfs_write_marked_extents(fs_info, &log->dirty_log_pages, mark);
2992	/*
2993	 * -EAGAIN happens when someone, e.g., a concurrent transaction
2994	 *  commit, writes a dirty extent in this tree-log commit. This
2995	 *  concurrent write will create a hole writing out the extents,
2996	 *  and we cannot proceed on a zoned filesystem, requiring
2997	 *  sequential writing. While we can bail out to a full commit
2998	 *  here, but we can continue hoping the concurrent writing fills
2999	 *  the hole.
3000	 */
3001	if (ret == -EAGAIN && btrfs_is_zoned(fs_info))
3002		ret = 0;
3003	if (ret) {
3004		blk_finish_plug(&plug);
3005		btrfs_set_log_full_commit(trans);
3006		mutex_unlock(&root->log_mutex);
3007		goto out;
3008	}
3009
3010	/*
3011	 * We _must_ update under the root->log_mutex in order to make sure we
3012	 * have a consistent view of the log root we are trying to commit at
3013	 * this moment.
3014	 *
3015	 * We _must_ copy this into a local copy, because we are not holding the
3016	 * log_root_tree->log_mutex yet.  This is important because when we
3017	 * commit the log_root_tree we must have a consistent view of the
3018	 * log_root_tree when we update the super block to point at the
3019	 * log_root_tree bytenr.  If we update the log_root_tree here we'll race
3020	 * with the commit and possibly point at the new block which we may not
3021	 * have written out.
3022	 */
3023	btrfs_set_root_node(&log->root_item, log->node);
3024	memcpy(&new_root_item, &log->root_item, sizeof(new_root_item));
3025
3026	btrfs_set_root_log_transid(root, root->log_transid + 1);
3027	log->log_transid = root->log_transid;
3028	root->log_start_pid = 0;
3029	/*
3030	 * IO has been started, blocks of the log tree have WRITTEN flag set
3031	 * in their headers. new modifications of the log will be written to
3032	 * new positions. so it's safe to allow log writers to go in.
3033	 */
3034	mutex_unlock(&root->log_mutex);
3035
3036	if (btrfs_is_zoned(fs_info)) {
3037		mutex_lock(&fs_info->tree_root->log_mutex);
3038		if (!log_root_tree->node) {
3039			ret = btrfs_alloc_log_tree_node(trans, log_root_tree);
3040			if (ret) {
3041				mutex_unlock(&fs_info->tree_root->log_mutex);
3042				blk_finish_plug(&plug);
3043				goto out;
3044			}
3045		}
3046		mutex_unlock(&fs_info->tree_root->log_mutex);
3047	}
3048
3049	btrfs_init_log_ctx(&root_log_ctx, NULL);
3050
3051	mutex_lock(&log_root_tree->log_mutex);
3052
3053	index2 = log_root_tree->log_transid % 2;
3054	list_add_tail(&root_log_ctx.list, &log_root_tree->log_ctxs[index2]);
3055	root_log_ctx.log_transid = log_root_tree->log_transid;
3056
3057	/*
3058	 * Now we are safe to update the log_root_tree because we're under the
3059	 * log_mutex, and we're a current writer so we're holding the commit
3060	 * open until we drop the log_mutex.
3061	 */
3062	ret = update_log_root(trans, log, &new_root_item);
3063	if (ret) {
3064		list_del_init(&root_log_ctx.list);
3065		blk_finish_plug(&plug);
3066		btrfs_set_log_full_commit(trans);
3067		if (ret != -ENOSPC)
3068			btrfs_err(fs_info,
3069				  "failed to update log for root %llu ret %d",
3070				  btrfs_root_id(root), ret);
3071		btrfs_wait_tree_log_extents(log, mark);
3072		mutex_unlock(&log_root_tree->log_mutex);
3073		goto out;
3074	}
3075
3076	if (log_root_tree->log_transid_committed >= root_log_ctx.log_transid) {
3077		blk_finish_plug(&plug);
3078		list_del_init(&root_log_ctx.list);
3079		mutex_unlock(&log_root_tree->log_mutex);
3080		ret = root_log_ctx.log_ret;
3081		goto out;
3082	}
3083
3084	if (atomic_read(&log_root_tree->log_commit[index2])) {
3085		blk_finish_plug(&plug);
3086		ret = btrfs_wait_tree_log_extents(log, mark);
3087		wait_log_commit(log_root_tree,
3088				root_log_ctx.log_transid);
3089		mutex_unlock(&log_root_tree->log_mutex);
3090		if (!ret)
3091			ret = root_log_ctx.log_ret;
3092		goto out;
3093	}
3094	ASSERT(root_log_ctx.log_transid == log_root_tree->log_transid);
3095	atomic_set(&log_root_tree->log_commit[index2], 1);
3096
3097	if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
3098		wait_log_commit(log_root_tree,
3099				root_log_ctx.log_transid - 1);
3100	}
3101
3102	/*
3103	 * now that we've moved on to the tree of log tree roots,
3104	 * check the full commit flag again
3105	 */
3106	if (btrfs_need_log_full_commit(trans)) {
3107		blk_finish_plug(&plug);
3108		btrfs_wait_tree_log_extents(log, mark);
3109		mutex_unlock(&log_root_tree->log_mutex);
3110		ret = BTRFS_LOG_FORCE_COMMIT;
3111		goto out_wake_log_root;
3112	}
3113
3114	ret = btrfs_write_marked_extents(fs_info,
3115					 &log_root_tree->dirty_log_pages,
3116					 EXTENT_DIRTY | EXTENT_NEW);
3117	blk_finish_plug(&plug);
3118	/*
3119	 * As described above, -EAGAIN indicates a hole in the extents. We
3120	 * cannot wait for these write outs since the waiting cause a
3121	 * deadlock. Bail out to the full commit instead.
3122	 */
3123	if (ret == -EAGAIN && btrfs_is_zoned(fs_info)) {
3124		btrfs_set_log_full_commit(trans);
3125		btrfs_wait_tree_log_extents(log, mark);
3126		mutex_unlock(&log_root_tree->log_mutex);
3127		goto out_wake_log_root;
3128	} else if (ret) {
3129		btrfs_set_log_full_commit(trans);
3130		mutex_unlock(&log_root_tree->log_mutex);
3131		goto out_wake_log_root;
3132	}
3133	ret = btrfs_wait_tree_log_extents(log, mark);
3134	if (!ret)
3135		ret = btrfs_wait_tree_log_extents(log_root_tree,
3136						  EXTENT_NEW | EXTENT_DIRTY);
3137	if (ret) {
3138		btrfs_set_log_full_commit(trans);
3139		mutex_unlock(&log_root_tree->log_mutex);
3140		goto out_wake_log_root;
3141	}
3142
3143	log_root_start = log_root_tree->node->start;
3144	log_root_level = btrfs_header_level(log_root_tree->node);
3145	log_root_tree->log_transid++;
3146	mutex_unlock(&log_root_tree->log_mutex);
3147
3148	/*
3149	 * Here we are guaranteed that nobody is going to write the superblock
3150	 * for the current transaction before us and that neither we do write
3151	 * our superblock before the previous transaction finishes its commit
3152	 * and writes its superblock, because:
3153	 *
3154	 * 1) We are holding a handle on the current transaction, so no body
3155	 *    can commit it until we release the handle;
3156	 *
3157	 * 2) Before writing our superblock we acquire the tree_log_mutex, so
3158	 *    if the previous transaction is still committing, and hasn't yet
3159	 *    written its superblock, we wait for it to do it, because a
3160	 *    transaction commit acquires the tree_log_mutex when the commit
3161	 *    begins and releases it only after writing its superblock.
3162	 */
3163	mutex_lock(&fs_info->tree_log_mutex);
3164
3165	/*
3166	 * The previous transaction writeout phase could have failed, and thus
3167	 * marked the fs in an error state.  We must not commit here, as we
3168	 * could have updated our generation in the super_for_commit and
3169	 * writing the super here would result in transid mismatches.  If there
3170	 * is an error here just bail.
3171	 */
3172	if (BTRFS_FS_ERROR(fs_info)) {
3173		ret = -EIO;
3174		btrfs_set_log_full_commit(trans);
3175		btrfs_abort_transaction(trans, ret);
3176		mutex_unlock(&fs_info->tree_log_mutex);
3177		goto out_wake_log_root;
3178	}
3179
3180	btrfs_set_super_log_root(fs_info->super_for_commit, log_root_start);
3181	btrfs_set_super_log_root_level(fs_info->super_for_commit, log_root_level);
3182	ret = write_all_supers(fs_info, 1);
3183	mutex_unlock(&fs_info->tree_log_mutex);
3184	if (ret) {
3185		btrfs_set_log_full_commit(trans);
3186		btrfs_abort_transaction(trans, ret);
3187		goto out_wake_log_root;
3188	}
3189
3190	/*
3191	 * We know there can only be one task here, since we have not yet set
3192	 * root->log_commit[index1] to 0 and any task attempting to sync the
3193	 * log must wait for the previous log transaction to commit if it's
3194	 * still in progress or wait for the current log transaction commit if
3195	 * someone else already started it. We use <= and not < because the
3196	 * first log transaction has an ID of 0.
3197	 */
3198	ASSERT(btrfs_get_root_last_log_commit(root) <= log_transid);
3199	btrfs_set_root_last_log_commit(root, log_transid);
3200
3201out_wake_log_root:
3202	mutex_lock(&log_root_tree->log_mutex);
3203	btrfs_remove_all_log_ctxs(log_root_tree, index2, ret);
3204
3205	log_root_tree->log_transid_committed++;
3206	atomic_set(&log_root_tree->log_commit[index2], 0);
3207	mutex_unlock(&log_root_tree->log_mutex);
3208
3209	/*
3210	 * The barrier before waitqueue_active (in cond_wake_up) is needed so
3211	 * all the updates above are seen by the woken threads. It might not be
3212	 * necessary, but proving that seems to be hard.
3213	 */
3214	cond_wake_up(&log_root_tree->log_commit_wait[index2]);
3215out:
3216	mutex_lock(&root->log_mutex);
3217	btrfs_remove_all_log_ctxs(root, index1, ret);
3218	root->log_transid_committed++;
3219	atomic_set(&root->log_commit[index1], 0);
3220	mutex_unlock(&root->log_mutex);
3221
3222	/*
3223	 * The barrier before waitqueue_active (in cond_wake_up) is needed so
3224	 * all the updates above are seen by the woken threads. It might not be
3225	 * necessary, but proving that seems to be hard.
3226	 */
3227	cond_wake_up(&root->log_commit_wait[index1]);
3228	return ret;
3229}
3230
3231static void free_log_tree(struct btrfs_trans_handle *trans,
3232			  struct btrfs_root *log)
3233{
3234	int ret;
3235	struct walk_control wc = {
3236		.free = 1,
3237		.process_func = process_one_buffer
3238	};
3239
3240	if (log->node) {
3241		ret = walk_log_tree(trans, log, &wc);
3242		if (ret) {
3243			/*
3244			 * We weren't able to traverse the entire log tree, the
3245			 * typical scenario is getting an -EIO when reading an
3246			 * extent buffer of the tree, due to a previous writeback
3247			 * failure of it.
3248			 */
3249			set_bit(BTRFS_FS_STATE_LOG_CLEANUP_ERROR,
3250				&log->fs_info->fs_state);
3251
3252			/*
3253			 * Some extent buffers of the log tree may still be dirty
3254			 * and not yet written back to storage, because we may
3255			 * have updates to a log tree without syncing a log tree,
3256			 * such as during rename and link operations. So flush
3257			 * them out and wait for their writeback to complete, so
3258			 * that we properly cleanup their state and pages.
3259			 */
3260			btrfs_write_marked_extents(log->fs_info,
3261						   &log->dirty_log_pages,
3262						   EXTENT_DIRTY | EXTENT_NEW);
3263			btrfs_wait_tree_log_extents(log,
3264						    EXTENT_DIRTY | EXTENT_NEW);
3265
3266			if (trans)
3267				btrfs_abort_transaction(trans, ret);
3268			else
3269				btrfs_handle_fs_error(log->fs_info, ret, NULL);
3270		}
3271	}
3272
3273	extent_io_tree_release(&log->dirty_log_pages);
3274	extent_io_tree_release(&log->log_csum_range);
3275
3276	btrfs_put_root(log);
3277}
3278
3279/*
3280 * free all the extents used by the tree log.  This should be called
3281 * at commit time of the full transaction
3282 */
3283int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
3284{
3285	if (root->log_root) {
3286		free_log_tree(trans, root->log_root);
3287		root->log_root = NULL;
3288		clear_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state);
3289	}
3290	return 0;
3291}
3292
3293int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
3294			     struct btrfs_fs_info *fs_info)
3295{
3296	if (fs_info->log_root_tree) {
3297		free_log_tree(trans, fs_info->log_root_tree);
3298		fs_info->log_root_tree = NULL;
3299		clear_bit(BTRFS_ROOT_HAS_LOG_TREE, &fs_info->tree_root->state);
3300	}
3301	return 0;
3302}
3303
3304/*
3305 * Check if an inode was logged in the current transaction. This correctly deals
3306 * with the case where the inode was logged but has a logged_trans of 0, which
3307 * happens if the inode is evicted and loaded again, as logged_trans is an in
3308 * memory only field (not persisted).
3309 *
3310 * Returns 1 if the inode was logged before in the transaction, 0 if it was not,
3311 * and < 0 on error.
3312 */
3313static int inode_logged(const struct btrfs_trans_handle *trans,
3314			struct btrfs_inode *inode,
3315			struct btrfs_path *path_in)
3316{
3317	struct btrfs_path *path = path_in;
3318	struct btrfs_key key;
3319	int ret;
3320
3321	if (inode->logged_trans == trans->transid)
3322		return 1;
3323
3324	/*
3325	 * If logged_trans is not 0, then we know the inode logged was not logged
3326	 * in this transaction, so we can return false right away.
3327	 */
3328	if (inode->logged_trans > 0)
3329		return 0;
3330
3331	/*
3332	 * If no log tree was created for this root in this transaction, then
3333	 * the inode can not have been logged in this transaction. In that case
3334	 * set logged_trans to anything greater than 0 and less than the current
3335	 * transaction's ID, to avoid the search below in a future call in case
3336	 * a log tree gets created after this.
3337	 */
3338	if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &inode->root->state)) {
3339		inode->logged_trans = trans->transid - 1;
3340		return 0;
3341	}
3342
3343	/*
3344	 * We have a log tree and the inode's logged_trans is 0. We can't tell
3345	 * for sure if the inode was logged before in this transaction by looking
3346	 * only at logged_trans. We could be pessimistic and assume it was, but
3347	 * that can lead to unnecessarily logging an inode during rename and link
3348	 * operations, and then further updating the log in followup rename and
3349	 * link operations, specially if it's a directory, which adds latency
3350	 * visible to applications doing a series of rename or link operations.
3351	 *
3352	 * A logged_trans of 0 here can mean several things:
3353	 *
3354	 * 1) The inode was never logged since the filesystem was mounted, and may
3355	 *    or may have not been evicted and loaded again;
3356	 *
3357	 * 2) The inode was logged in a previous transaction, then evicted and
3358	 *    then loaded again;
3359	 *
3360	 * 3) The inode was logged in the current transaction, then evicted and
3361	 *    then loaded again.
3362	 *
3363	 * For cases 1) and 2) we don't want to return true, but we need to detect
3364	 * case 3) and return true. So we do a search in the log root for the inode
3365	 * item.
3366	 */
3367	key.objectid = btrfs_ino(inode);
3368	key.type = BTRFS_INODE_ITEM_KEY;
3369	key.offset = 0;
3370
3371	if (!path) {
3372		path = btrfs_alloc_path();
3373		if (!path)
3374			return -ENOMEM;
3375	}
3376
3377	ret = btrfs_search_slot(NULL, inode->root->log_root, &key, path, 0, 0);
3378
3379	if (path_in)
3380		btrfs_release_path(path);
3381	else
3382		btrfs_free_path(path);
3383
3384	/*
3385	 * Logging an inode always results in logging its inode item. So if we
3386	 * did not find the item we know the inode was not logged for sure.
3387	 */
3388	if (ret < 0) {
3389		return ret;
3390	} else if (ret > 0) {
3391		/*
3392		 * Set logged_trans to a value greater than 0 and less then the
3393		 * current transaction to avoid doing the search in future calls.
3394		 */
3395		inode->logged_trans = trans->transid - 1;
3396		return 0;
3397	}
3398
3399	/*
3400	 * The inode was previously logged and then evicted, set logged_trans to
3401	 * the current transacion's ID, to avoid future tree searches as long as
3402	 * the inode is not evicted again.
3403	 */
3404	inode->logged_trans = trans->transid;
3405
3406	/*
3407	 * If it's a directory, then we must set last_dir_index_offset to the
3408	 * maximum possible value, so that the next attempt to log the inode does
3409	 * not skip checking if dir index keys found in modified subvolume tree
3410	 * leaves have been logged before, otherwise it would result in attempts
3411	 * to insert duplicate dir index keys in the log tree. This must be done
3412	 * because last_dir_index_offset is an in-memory only field, not persisted
3413	 * in the inode item or any other on-disk structure, so its value is lost
3414	 * once the inode is evicted.
3415	 */
3416	if (S_ISDIR(inode->vfs_inode.i_mode))
3417		inode->last_dir_index_offset = (u64)-1;
3418
3419	return 1;
3420}
3421
3422/*
3423 * Delete a directory entry from the log if it exists.
3424 *
3425 * Returns < 0 on error
3426 *           1 if the entry does not exists
3427 *           0 if the entry existed and was successfully deleted
3428 */
3429static int del_logged_dentry(struct btrfs_trans_handle *trans,
3430			     struct btrfs_root *log,
3431			     struct btrfs_path *path,
3432			     u64 dir_ino,
3433			     const struct fscrypt_str *name,
3434			     u64 index)
3435{
3436	struct btrfs_dir_item *di;
3437
3438	/*
3439	 * We only log dir index items of a directory, so we don't need to look
3440	 * for dir item keys.
3441	 */
3442	di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
3443					 index, name, -1);
3444	if (IS_ERR(di))
3445		return PTR_ERR(di);
3446	else if (!di)
3447		return 1;
3448
3449	/*
3450	 * We do not need to update the size field of the directory's
3451	 * inode item because on log replay we update the field to reflect
3452	 * all existing entries in the directory (see overwrite_item()).
3453	 */
3454	return btrfs_delete_one_dir_name(trans, log, path, di);
3455}
3456
3457/*
3458 * If both a file and directory are logged, and unlinks or renames are
3459 * mixed in, we have a few interesting corners:
3460 *
3461 * create file X in dir Y
3462 * link file X to X.link in dir Y
3463 * fsync file X
3464 * unlink file X but leave X.link
3465 * fsync dir Y
3466 *
3467 * After a crash we would expect only X.link to exist.  But file X
3468 * didn't get fsync'd again so the log has back refs for X and X.link.
3469 *
3470 * We solve this by removing directory entries and inode backrefs from the
3471 * log when a file that was logged in the current transaction is
3472 * unlinked.  Any later fsync will include the updated log entries, and
3473 * we'll be able to reconstruct the proper directory items from backrefs.
3474 *
3475 * This optimizations allows us to avoid relogging the entire inode
3476 * or the entire directory.
3477 */
3478void btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
3479				  struct btrfs_root *root,
3480				  const struct fscrypt_str *name,
3481				  struct btrfs_inode *dir, u64 index)
3482{
3483	struct btrfs_path *path;
3484	int ret;
3485
3486	ret = inode_logged(trans, dir, NULL);
3487	if (ret == 0)
3488		return;
3489	else if (ret < 0) {
3490		btrfs_set_log_full_commit(trans);
3491		return;
3492	}
3493
3494	ret = join_running_log_trans(root);
3495	if (ret)
3496		return;
3497
3498	mutex_lock(&dir->log_mutex);
3499
3500	path = btrfs_alloc_path();
3501	if (!path) {
3502		ret = -ENOMEM;
3503		goto out_unlock;
3504	}
3505
3506	ret = del_logged_dentry(trans, root->log_root, path, btrfs_ino(dir),
3507				name, index);
3508	btrfs_free_path(path);
3509out_unlock:
3510	mutex_unlock(&dir->log_mutex);
3511	if (ret < 0)
3512		btrfs_set_log_full_commit(trans);
3513	btrfs_end_log_trans(root);
3514}
3515
3516/* see comments for btrfs_del_dir_entries_in_log */
3517void btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
3518				struct btrfs_root *root,
3519				const struct fscrypt_str *name,
3520				struct btrfs_inode *inode, u64 dirid)
3521{
3522	struct btrfs_root *log;
3523	u64 index;
3524	int ret;
3525
3526	ret = inode_logged(trans, inode, NULL);
3527	if (ret == 0)
3528		return;
3529	else if (ret < 0) {
3530		btrfs_set_log_full_commit(trans);
3531		return;
3532	}
3533
3534	ret = join_running_log_trans(root);
3535	if (ret)
3536		return;
3537	log = root->log_root;
3538	mutex_lock(&inode->log_mutex);
3539
3540	ret = btrfs_del_inode_ref(trans, log, name, btrfs_ino(inode),
3541				  dirid, &index);
3542	mutex_unlock(&inode->log_mutex);
3543	if (ret < 0 && ret != -ENOENT)
3544		btrfs_set_log_full_commit(trans);
3545	btrfs_end_log_trans(root);
3546}
3547
3548/*
3549 * creates a range item in the log for 'dirid'.  first_offset and
3550 * last_offset tell us which parts of the key space the log should
3551 * be considered authoritative for.
3552 */
3553static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
3554				       struct btrfs_root *log,
3555				       struct btrfs_path *path,
3556				       u64 dirid,
3557				       u64 first_offset, u64 last_offset)
3558{
3559	int ret;
3560	struct btrfs_key key;
3561	struct btrfs_dir_log_item *item;
3562
3563	key.objectid = dirid;
3564	key.offset = first_offset;
3565	key.type = BTRFS_DIR_LOG_INDEX_KEY;
3566	ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
3567	/*
3568	 * -EEXIST is fine and can happen sporadically when we are logging a
3569	 * directory and have concurrent insertions in the subvolume's tree for
3570	 * items from other inodes and that result in pushing off some dir items
3571	 * from one leaf to another in order to accommodate for the new items.
3572	 * This results in logging the same dir index range key.
3573	 */
3574	if (ret && ret != -EEXIST)
3575		return ret;
3576
3577	item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3578			      struct btrfs_dir_log_item);
3579	if (ret == -EEXIST) {
3580		const u64 curr_end = btrfs_dir_log_end(path->nodes[0], item);
3581
3582		/*
3583		 * btrfs_del_dir_entries_in_log() might have been called during
3584		 * an unlink between the initial insertion of this key and the
3585		 * current update, or we might be logging a single entry deletion
3586		 * during a rename, so set the new last_offset to the max value.
3587		 */
3588		last_offset = max(last_offset, curr_end);
3589	}
3590	btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
3591	btrfs_mark_buffer_dirty(trans, path->nodes[0]);
3592	btrfs_release_path(path);
3593	return 0;
3594}
3595
3596static int flush_dir_items_batch(struct btrfs_trans_handle *trans,
3597				 struct btrfs_inode *inode,
3598				 struct extent_buffer *src,
3599				 struct btrfs_path *dst_path,
3600				 int start_slot,
3601				 int count)
3602{
3603	struct btrfs_root *log = inode->root->log_root;
3604	char *ins_data = NULL;
3605	struct btrfs_item_batch batch;
3606	struct extent_buffer *dst;
3607	unsigned long src_offset;
3608	unsigned long dst_offset;
3609	u64 last_index;
3610	struct btrfs_key key;
3611	u32 item_size;
3612	int ret;
3613	int i;
3614
3615	ASSERT(count > 0);
3616	batch.nr = count;
3617
3618	if (count == 1) {
3619		btrfs_item_key_to_cpu(src, &key, start_slot);
3620		item_size = btrfs_item_size(src, start_slot);
3621		batch.keys = &key;
3622		batch.data_sizes = &item_size;
3623		batch.total_data_size = item_size;
3624	} else {
3625		struct btrfs_key *ins_keys;
3626		u32 *ins_sizes;
3627
3628		ins_data = kmalloc(count * sizeof(u32) +
3629				   count * sizeof(struct btrfs_key), GFP_NOFS);
3630		if (!ins_data)
3631			return -ENOMEM;
3632
3633		ins_sizes = (u32 *)ins_data;
3634		ins_keys = (struct btrfs_key *)(ins_data + count * sizeof(u32));
3635		batch.keys = ins_keys;
3636		batch.data_sizes = ins_sizes;
3637		batch.total_data_size = 0;
3638
3639		for (i = 0; i < count; i++) {
3640			const int slot = start_slot + i;
3641
3642			btrfs_item_key_to_cpu(src, &ins_keys[i], slot);
3643			ins_sizes[i] = btrfs_item_size(src, slot);
3644			batch.total_data_size += ins_sizes[i];
3645		}
3646	}
3647
3648	ret = btrfs_insert_empty_items(trans, log, dst_path, &batch);
3649	if (ret)
3650		goto out;
3651
3652	dst = dst_path->nodes[0];
3653	/*
3654	 * Copy all the items in bulk, in a single copy operation. Item data is
3655	 * organized such that it's placed at the end of a leaf and from right
3656	 * to left. For example, the data for the second item ends at an offset
3657	 * that matches the offset where the data for the first item starts, the
3658	 * data for the third item ends at an offset that matches the offset
3659	 * where the data of the second items starts, and so on.
3660	 * Therefore our source and destination start offsets for copy match the
3661	 * offsets of the last items (highest slots).
3662	 */
3663	dst_offset = btrfs_item_ptr_offset(dst, dst_path->slots[0] + count - 1);
3664	src_offset = btrfs_item_ptr_offset(src, start_slot + count - 1);
3665	copy_extent_buffer(dst, src, dst_offset, src_offset, batch.total_data_size);
3666	btrfs_release_path(dst_path);
3667
3668	last_index = batch.keys[count - 1].offset;
3669	ASSERT(last_index > inode->last_dir_index_offset);
3670
3671	/*
3672	 * If for some unexpected reason the last item's index is not greater
3673	 * than the last index we logged, warn and force a transaction commit.
3674	 */
3675	if (WARN_ON(last_index <= inode->last_dir_index_offset))
3676		ret = BTRFS_LOG_FORCE_COMMIT;
3677	else
3678		inode->last_dir_index_offset = last_index;
3679
3680	if (btrfs_get_first_dir_index_to_log(inode) == 0)
3681		btrfs_set_first_dir_index_to_log(inode, batch.keys[0].offset);
3682out:
3683	kfree(ins_data);
3684
3685	return ret;
3686}
3687
3688static int clone_leaf(struct btrfs_path *path, struct btrfs_log_ctx *ctx)
3689{
3690	const int slot = path->slots[0];
3691
3692	if (ctx->scratch_eb) {
3693		copy_extent_buffer_full(ctx->scratch_eb, path->nodes[0]);
3694	} else {
3695		ctx->scratch_eb = btrfs_clone_extent_buffer(path->nodes[0]);
3696		if (!ctx->scratch_eb)
3697			return -ENOMEM;
3698	}
3699
3700	btrfs_release_path(path);
3701	path->nodes[0] = ctx->scratch_eb;
3702	path->slots[0] = slot;
3703	/*
3704	 * Add extra ref to scratch eb so that it is not freed when callers
3705	 * release the path, so we can reuse it later if needed.
3706	 */
3707	atomic_inc(&ctx->scratch_eb->refs);
3708
3709	return 0;
3710}
3711
3712static int process_dir_items_leaf(struct btrfs_trans_handle *trans,
3713				  struct btrfs_inode *inode,
3714				  struct btrfs_path *path,
3715				  struct btrfs_path *dst_path,
3716				  struct btrfs_log_ctx *ctx,
3717				  u64 *last_old_dentry_offset)
3718{
3719	struct btrfs_root *log = inode->root->log_root;
3720	struct extent_buffer *src;
3721	const int nritems = btrfs_header_nritems(path->nodes[0]);
3722	const u64 ino = btrfs_ino(inode);
3723	bool last_found = false;
3724	int batch_start = 0;
3725	int batch_size = 0;
3726	int ret;
3727
3728	/*
3729	 * We need to clone the leaf, release the read lock on it, and use the
3730	 * clone before modifying the log tree. See the comment at copy_items()
3731	 * about why we need to do this.
3732	 */
3733	ret = clone_leaf(path, ctx);
3734	if (ret < 0)
3735		return ret;
3736
3737	src = path->nodes[0];
3738
3739	for (int i = path->slots[0]; i < nritems; i++) {
3740		struct btrfs_dir_item *di;
3741		struct btrfs_key key;
3742		int ret;
3743
3744		btrfs_item_key_to_cpu(src, &key, i);
3745
3746		if (key.objectid != ino || key.type != BTRFS_DIR_INDEX_KEY) {
3747			last_found = true;
3748			break;
3749		}
3750
3751		di = btrfs_item_ptr(src, i, struct btrfs_dir_item);
3752
3753		/*
3754		 * Skip ranges of items that consist only of dir item keys created
3755		 * in past transactions. However if we find a gap, we must log a
3756		 * dir index range item for that gap, so that index keys in that
3757		 * gap are deleted during log replay.
3758		 */
3759		if (btrfs_dir_transid(src, di) < trans->transid) {
3760			if (key.offset > *last_old_dentry_offset + 1) {
3761				ret = insert_dir_log_key(trans, log, dst_path,
3762						 ino, *last_old_dentry_offset + 1,
3763						 key.offset - 1);
3764				if (ret < 0)
3765					return ret;
3766			}
3767
3768			*last_old_dentry_offset = key.offset;
3769			continue;
3770		}
3771
3772		/* If we logged this dir index item before, we can skip it. */
3773		if (key.offset <= inode->last_dir_index_offset)
3774			continue;
3775
3776		/*
3777		 * We must make sure that when we log a directory entry, the
3778		 * corresponding inode, after log replay, has a matching link
3779		 * count. For example:
3780		 *
3781		 * touch foo
3782		 * mkdir mydir
3783		 * sync
3784		 * ln foo mydir/bar
3785		 * xfs_io -c "fsync" mydir
3786		 * <crash>
3787		 * <mount fs and log replay>
3788		 *
3789		 * Would result in a fsync log that when replayed, our file inode
3790		 * would have a link count of 1, but we get two directory entries
3791		 * pointing to the same inode. After removing one of the names,
3792		 * it would not be possible to remove the other name, which
3793		 * resulted always in stale file handle errors, and would not be
3794		 * possible to rmdir the parent directory, since its i_size could
3795		 * never be decremented to the value BTRFS_EMPTY_DIR_SIZE,
3796		 * resulting in -ENOTEMPTY errors.
3797		 */
3798		if (!ctx->log_new_dentries) {
3799			struct btrfs_key di_key;
3800
3801			btrfs_dir_item_key_to_cpu(src, di, &di_key);
3802			if (di_key.type != BTRFS_ROOT_ITEM_KEY)
3803				ctx->log_new_dentries = true;
3804		}
3805
3806		if (batch_size == 0)
3807			batch_start = i;
3808		batch_size++;
3809	}
3810
3811	if (batch_size > 0) {
3812		int ret;
3813
3814		ret = flush_dir_items_batch(trans, inode, src, dst_path,
3815					    batch_start, batch_size);
3816		if (ret < 0)
3817			return ret;
3818	}
3819
3820	return last_found ? 1 : 0;
3821}
3822
3823/*
3824 * log all the items included in the current transaction for a given
3825 * directory.  This also creates the range items in the log tree required
3826 * to replay anything deleted before the fsync
3827 */
3828static noinline int log_dir_items(struct btrfs_trans_handle *trans,
3829			  struct btrfs_inode *inode,
3830			  struct btrfs_path *path,
3831			  struct btrfs_path *dst_path,
3832			  struct btrfs_log_ctx *ctx,
3833			  u64 min_offset, u64 *last_offset_ret)
3834{
3835	struct btrfs_key min_key;
3836	struct btrfs_root *root = inode->root;
3837	struct btrfs_root *log = root->log_root;
3838	int ret;
3839	u64 last_old_dentry_offset = min_offset - 1;
3840	u64 last_offset = (u64)-1;
3841	u64 ino = btrfs_ino(inode);
3842
3843	min_key.objectid = ino;
3844	min_key.type = BTRFS_DIR_INDEX_KEY;
3845	min_key.offset = min_offset;
3846
3847	ret = btrfs_search_forward(root, &min_key, path, trans->transid);
3848
3849	/*
3850	 * we didn't find anything from this transaction, see if there
3851	 * is anything at all
3852	 */
3853	if (ret != 0 || min_key.objectid != ino ||
3854	    min_key.type != BTRFS_DIR_INDEX_KEY) {
3855		min_key.objectid = ino;
3856		min_key.type = BTRFS_DIR_INDEX_KEY;
3857		min_key.offset = (u64)-1;
3858		btrfs_release_path(path);
3859		ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3860		if (ret < 0) {
3861			btrfs_release_path(path);
3862			return ret;
3863		}
3864		ret = btrfs_previous_item(root, path, ino, BTRFS_DIR_INDEX_KEY);
3865
3866		/* if ret == 0 there are items for this type,
3867		 * create a range to tell us the last key of this type.
3868		 * otherwise, there are no items in this directory after
3869		 * *min_offset, and we create a range to indicate that.
3870		 */
3871		if (ret == 0) {
3872			struct btrfs_key tmp;
3873
3874			btrfs_item_key_to_cpu(path->nodes[0], &tmp,
3875					      path->slots[0]);
3876			if (tmp.type == BTRFS_DIR_INDEX_KEY)
3877				last_old_dentry_offset = tmp.offset;
3878		} else if (ret > 0) {
3879			ret = 0;
3880		}
3881
3882		goto done;
3883	}
3884
3885	/* go backward to find any previous key */
3886	ret = btrfs_previous_item(root, path, ino, BTRFS_DIR_INDEX_KEY);
3887	if (ret == 0) {
3888		struct btrfs_key tmp;
3889
3890		btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3891		/*
3892		 * The dir index key before the first one we found that needs to
3893		 * be logged might be in a previous leaf, and there might be a
3894		 * gap between these keys, meaning that we had deletions that
3895		 * happened. So the key range item we log (key type
3896		 * BTRFS_DIR_LOG_INDEX_KEY) must cover a range that starts at the
3897		 * previous key's offset plus 1, so that those deletes are replayed.
3898		 */
3899		if (tmp.type == BTRFS_DIR_INDEX_KEY)
3900			last_old_dentry_offset = tmp.offset;
3901	} else if (ret < 0) {
3902		goto done;
3903	}
3904
3905	btrfs_release_path(path);
3906
3907	/*
3908	 * Find the first key from this transaction again or the one we were at
3909	 * in the loop below in case we had to reschedule. We may be logging the
3910	 * directory without holding its VFS lock, which happen when logging new
3911	 * dentries (through log_new_dir_dentries()) or in some cases when we
3912	 * need to log the parent directory of an inode. This means a dir index
3913	 * key might be deleted from the inode's root, and therefore we may not
3914	 * find it anymore. If we can't find it, just move to the next key. We
3915	 * can not bail out and ignore, because if we do that we will simply
3916	 * not log dir index keys that come after the one that was just deleted
3917	 * and we can end up logging a dir index range that ends at (u64)-1
3918	 * (@last_offset is initialized to that), resulting in removing dir
3919	 * entries we should not remove at log replay time.
3920	 */
3921search:
3922	ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3923	if (ret > 0) {
3924		ret = btrfs_next_item(root, path);
3925		if (ret > 0) {
3926			/* There are no more keys in the inode's root. */
3927			ret = 0;
3928			goto done;
3929		}
3930	}
3931	if (ret < 0)
3932		goto done;
3933
3934	/*
3935	 * we have a block from this transaction, log every item in it
3936	 * from our directory
3937	 */
3938	while (1) {
3939		ret = process_dir_items_leaf(trans, inode, path, dst_path, ctx,
3940					     &last_old_dentry_offset);
3941		if (ret != 0) {
3942			if (ret > 0)
3943				ret = 0;
3944			goto done;
3945		}
3946		path->slots[0] = btrfs_header_nritems(path->nodes[0]);
3947
3948		/*
3949		 * look ahead to the next item and see if it is also
3950		 * from this directory and from this transaction
3951		 */
3952		ret = btrfs_next_leaf(root, path);
3953		if (ret) {
3954			if (ret == 1) {
3955				last_offset = (u64)-1;
3956				ret = 0;
3957			}
3958			goto done;
3959		}
3960		btrfs_item_key_to_cpu(path->nodes[0], &min_key, path->slots[0]);
3961		if (min_key.objectid != ino || min_key.type != BTRFS_DIR_INDEX_KEY) {
3962			last_offset = (u64)-1;
3963			goto done;
3964		}
3965		if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
3966			/*
3967			 * The next leaf was not changed in the current transaction
3968			 * and has at least one dir index key.
3969			 * We check for the next key because there might have been
3970			 * one or more deletions between the last key we logged and
3971			 * that next key. So the key range item we log (key type
3972			 * BTRFS_DIR_LOG_INDEX_KEY) must end at the next key's
3973			 * offset minus 1, so that those deletes are replayed.
3974			 */
3975			last_offset = min_key.offset - 1;
3976			goto done;
3977		}
3978		if (need_resched()) {
3979			btrfs_release_path(path);
3980			cond_resched();
3981			goto search;
3982		}
3983	}
3984done:
3985	btrfs_release_path(path);
3986	btrfs_release_path(dst_path);
3987
3988	if (ret == 0) {
3989		*last_offset_ret = last_offset;
3990		/*
3991		 * In case the leaf was changed in the current transaction but
3992		 * all its dir items are from a past transaction, the last item
3993		 * in the leaf is a dir item and there's no gap between that last
3994		 * dir item and the first one on the next leaf (which did not
3995		 * change in the current transaction), then we don't need to log
3996		 * a range, last_old_dentry_offset is == to last_offset.
3997		 */
3998		ASSERT(last_old_dentry_offset <= last_offset);
3999		if (last_old_dentry_offset < last_offset)
4000			ret = insert_dir_log_key(trans, log, path, ino,
4001						 last_old_dentry_offset + 1,
4002						 last_offset);
4003	}
4004
4005	return ret;
4006}
4007
4008/*
4009 * If the inode was logged before and it was evicted, then its
4010 * last_dir_index_offset is (u64)-1, so we don't the value of the last index
4011 * key offset. If that's the case, search for it and update the inode. This
4012 * is to avoid lookups in the log tree every time we try to insert a dir index
4013 * key from a leaf changed in the current transaction, and to allow us to always
4014 * do batch insertions of dir index keys.
4015 */
4016static int update_last_dir_index_offset(struct btrfs_inode *inode,
4017					struct btrfs_path *path,
4018					const struct btrfs_log_ctx *ctx)
4019{
4020	const u64 ino = btrfs_ino(inode);
4021	struct btrfs_key key;
4022	int ret;
4023
4024	lockdep_assert_held(&inode->log_mutex);
4025
4026	if (inode->last_dir_index_offset != (u64)-1)
4027		return 0;
4028
4029	if (!ctx->logged_before) {
4030		inode->last_dir_index_offset = BTRFS_DIR_START_INDEX - 1;
4031		return 0;
4032	}
4033
4034	key.objectid = ino;
4035	key.type = BTRFS_DIR_INDEX_KEY;
4036	key.offset = (u64)-1;
4037
4038	ret = btrfs_search_slot(NULL, inode->root->log_root, &key, path, 0, 0);
4039	/*
4040	 * An error happened or we actually have an index key with an offset
4041	 * value of (u64)-1. Bail out, we're done.
4042	 */
4043	if (ret <= 0)
4044		goto out;
4045
4046	ret = 0;
4047	inode->last_dir_index_offset = BTRFS_DIR_START_INDEX - 1;
4048
4049	/*
4050	 * No dir index items, bail out and leave last_dir_index_offset with
4051	 * the value right before the first valid index value.
4052	 */
4053	if (path->slots[0] == 0)
4054		goto out;
4055
4056	/*
4057	 * btrfs_search_slot() left us at one slot beyond the slot with the last
4058	 * index key, or beyond the last key of the directory that is not an
4059	 * index key. If we have an index key before, set last_dir_index_offset
4060	 * to its offset value, otherwise leave it with a value right before the
4061	 * first valid index value, as it means we have an empty directory.
4062	 */
4063	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
4064	if (key.objectid == ino && key.type == BTRFS_DIR_INDEX_KEY)
4065		inode->last_dir_index_offset = key.offset;
4066
4067out:
4068	btrfs_release_path(path);
4069
4070	return ret;
4071}
4072
4073/*
4074 * logging directories is very similar to logging inodes, We find all the items
4075 * from the current transaction and write them to the log.
4076 *
4077 * The recovery code scans the directory in the subvolume, and if it finds a
4078 * key in the range logged that is not present in the log tree, then it means
4079 * that dir entry was unlinked during the transaction.
4080 *
4081 * In order for that scan to work, we must include one key smaller than
4082 * the smallest logged by this transaction and one key larger than the largest
4083 * key logged by this transaction.
4084 */
4085static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
4086			  struct btrfs_inode *inode,
4087			  struct btrfs_path *path,
4088			  struct btrfs_path *dst_path,
4089			  struct btrfs_log_ctx *ctx)
4090{
4091	u64 min_key;
4092	u64 max_key;
4093	int ret;
4094
4095	ret = update_last_dir_index_offset(inode, path, ctx);
4096	if (ret)
4097		return ret;
4098
4099	min_key = BTRFS_DIR_START_INDEX;
4100	max_key = 0;
4101
4102	while (1) {
4103		ret = log_dir_items(trans, inode, path, dst_path,
4104				ctx, min_key, &max_key);
4105		if (ret)
4106			return ret;
4107		if (max_key == (u64)-1)
4108			break;
4109		min_key = max_key + 1;
4110	}
4111
4112	return 0;
4113}
4114
4115/*
4116 * a helper function to drop items from the log before we relog an
4117 * inode.  max_key_type indicates the highest item type to remove.
4118 * This cannot be run for file data extents because it does not
4119 * free the extents they point to.
4120 */
4121static int drop_inode_items(struct btrfs_trans_handle *trans,
4122				  struct btrfs_root *log,
4123				  struct btrfs_path *path,
4124				  struct btrfs_inode *inode,
4125				  int max_key_type)
4126{
4127	int ret;
4128	struct btrfs_key key;
4129	struct btrfs_key found_key;
4130	int start_slot;
4131
4132	key.objectid = btrfs_ino(inode);
4133	key.type = max_key_type;
4134	key.offset = (u64)-1;
4135
4136	while (1) {
4137		ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
4138		if (ret < 0) {
4139			break;
4140		} else if (ret > 0) {
4141			if (path->slots[0] == 0)
4142				break;
4143			path->slots[0]--;
4144		}
4145
4146		btrfs_item_key_to_cpu(path->nodes[0], &found_key,
4147				      path->slots[0]);
4148
4149		if (found_key.objectid != key.objectid)
4150			break;
4151
4152		found_key.offset = 0;
4153		found_key.type = 0;
4154		ret = btrfs_bin_search(path->nodes[0], 0, &found_key, &start_slot);
4155		if (ret < 0)
4156			break;
4157
4158		ret = btrfs_del_items(trans, log, path, start_slot,
4159				      path->slots[0] - start_slot + 1);
4160		/*
4161		 * If start slot isn't 0 then we don't need to re-search, we've
4162		 * found the last guy with the objectid in this tree.
4163		 */
4164		if (ret || start_slot != 0)
4165			break;
4166		btrfs_release_path(path);
4167	}
4168	btrfs_release_path(path);
4169	if (ret > 0)
4170		ret = 0;
4171	return ret;
4172}
4173
4174static int truncate_inode_items(struct btrfs_trans_handle *trans,
4175				struct btrfs_root *log_root,
4176				struct btrfs_inode *inode,
4177				u64 new_size, u32 min_type)
4178{
4179	struct btrfs_truncate_control control = {
4180		.new_size = new_size,
4181		.ino = btrfs_ino(inode),
4182		.min_type = min_type,
4183		.skip_ref_updates = true,
4184	};
4185
4186	return btrfs_truncate_inode_items(trans, log_root, &control);
4187}
4188
4189static void fill_inode_item(struct btrfs_trans_handle *trans,
4190			    struct extent_buffer *leaf,
4191			    struct btrfs_inode_item *item,
4192			    struct inode *inode, int log_inode_only,
4193			    u64 logged_isize)
4194{
4195	struct btrfs_map_token token;
4196	u64 flags;
4197
4198	btrfs_init_map_token(&token, leaf);
4199
4200	if (log_inode_only) {
4201		/* set the generation to zero so the recover code
4202		 * can tell the difference between an logging
4203		 * just to say 'this inode exists' and a logging
4204		 * to say 'update this inode with these values'
4205		 */
4206		btrfs_set_token_inode_generation(&token, item, 0);
4207		btrfs_set_token_inode_size(&token, item, logged_isize);
4208	} else {
4209		btrfs_set_token_inode_generation(&token, item,
4210						 BTRFS_I(inode)->generation);
4211		btrfs_set_token_inode_size(&token, item, inode->i_size);
4212	}
4213
4214	btrfs_set_token_inode_uid(&token, item, i_uid_read(inode));
4215	btrfs_set_token_inode_gid(&token, item, i_gid_read(inode));
4216	btrfs_set_token_inode_mode(&token, item, inode->i_mode);
4217	btrfs_set_token_inode_nlink(&token, item, inode->i_nlink);
4218
4219	btrfs_set_token_timespec_sec(&token, &item->atime,
4220				     inode_get_atime_sec(inode));
4221	btrfs_set_token_timespec_nsec(&token, &item->atime,
4222				      inode_get_atime_nsec(inode));
4223
4224	btrfs_set_token_timespec_sec(&token, &item->mtime,
4225				     inode_get_mtime_sec(inode));
4226	btrfs_set_token_timespec_nsec(&token, &item->mtime,
4227				      inode_get_mtime_nsec(inode));
4228
4229	btrfs_set_token_timespec_sec(&token, &item->ctime,
4230				     inode_get_ctime_sec(inode));
4231	btrfs_set_token_timespec_nsec(&token, &item->ctime,
4232				      inode_get_ctime_nsec(inode));
4233
4234	/*
4235	 * We do not need to set the nbytes field, in fact during a fast fsync
4236	 * its value may not even be correct, since a fast fsync does not wait
4237	 * for ordered extent completion, which is where we update nbytes, it
4238	 * only waits for writeback to complete. During log replay as we find
4239	 * file extent items and replay them, we adjust the nbytes field of the
4240	 * inode item in subvolume tree as needed (see overwrite_item()).
4241	 */
4242
4243	btrfs_set_token_inode_sequence(&token, item, inode_peek_iversion(inode));
4244	btrfs_set_token_inode_transid(&token, item, trans->transid);
4245	btrfs_set_token_inode_rdev(&token, item, inode->i_rdev);
4246	flags = btrfs_inode_combine_flags(BTRFS_I(inode)->flags,
4247					  BTRFS_I(inode)->ro_flags);
4248	btrfs_set_token_inode_flags(&token, item, flags);
4249	btrfs_set_token_inode_block_group(&token, item, 0);
4250}
4251
4252static int log_inode_item(struct btrfs_trans_handle *trans,
4253			  struct btrfs_root *log, struct btrfs_path *path,
4254			  struct btrfs_inode *inode, bool inode_item_dropped)
4255{
4256	struct btrfs_inode_item *inode_item;
4257	struct btrfs_key key;
4258	int ret;
4259
4260	btrfs_get_inode_key(inode, &key);
4261	/*
4262	 * If we are doing a fast fsync and the inode was logged before in the
4263	 * current transaction, then we know the inode was previously logged and
4264	 * it exists in the log tree. For performance reasons, in this case use
4265	 * btrfs_search_slot() directly with ins_len set to 0 so that we never
4266	 * attempt a write lock on the leaf's parent, which adds unnecessary lock
4267	 * contention in case there are concurrent fsyncs for other inodes of the
4268	 * same subvolume. Using btrfs_insert_empty_item() when the inode item
4269	 * already exists can also result in unnecessarily splitting a leaf.
4270	 */
4271	if (!inode_item_dropped && inode->logged_trans == trans->transid) {
4272		ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
4273		ASSERT(ret <= 0);
4274		if (ret > 0)
4275			ret = -ENOENT;
4276	} else {
4277		/*
4278		 * This means it is the first fsync in the current transaction,
4279		 * so the inode item is not in the log and we need to insert it.
4280		 * We can never get -EEXIST because we are only called for a fast
4281		 * fsync and in case an inode eviction happens after the inode was
4282		 * logged before in the current transaction, when we load again
4283		 * the inode, we set BTRFS_INODE_NEEDS_FULL_SYNC on its runtime
4284		 * flags and set ->logged_trans to 0.
4285		 */
4286		ret = btrfs_insert_empty_item(trans, log, path, &key,
4287					      sizeof(*inode_item));
4288		ASSERT(ret != -EEXIST);
4289	}
4290	if (ret)
4291		return ret;
4292	inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
4293				    struct btrfs_inode_item);
4294	fill_inode_item(trans, path->nodes[0], inode_item, &inode->vfs_inode,
4295			0, 0);
4296	btrfs_release_path(path);
4297	return 0;
4298}
4299
4300static int log_csums(struct btrfs_trans_handle *trans,
4301		     struct btrfs_inode *inode,
4302		     struct btrfs_root *log_root,
4303		     struct btrfs_ordered_sum *sums)
4304{
4305	const u64 lock_end = sums->logical + sums->len - 1;
4306	struct extent_state *cached_state = NULL;
4307	int ret;
4308
4309	/*
4310	 * If this inode was not used for reflink operations in the current
4311	 * transaction with new extents, then do the fast path, no need to
4312	 * worry about logging checksum items with overlapping ranges.
4313	 */
4314	if (inode->last_reflink_trans < trans->transid)
4315		return btrfs_csum_file_blocks(trans, log_root, sums);
4316
4317	/*
4318	 * Serialize logging for checksums. This is to avoid racing with the
4319	 * same checksum being logged by another task that is logging another
4320	 * file which happens to refer to the same extent as well. Such races
4321	 * can leave checksum items in the log with overlapping ranges.
4322	 */
4323	ret = lock_extent(&log_root->log_csum_range, sums->logical, lock_end,
4324			  &cached_state);
4325	if (ret)
4326		return ret;
4327	/*
4328	 * Due to extent cloning, we might have logged a csum item that covers a
4329	 * subrange of a cloned extent, and later we can end up logging a csum
4330	 * item for a larger subrange of the same extent or the entire range.
4331	 * This would leave csum items in the log tree that cover the same range
4332	 * and break the searches for checksums in the log tree, resulting in
4333	 * some checksums missing in the fs/subvolume tree. So just delete (or
4334	 * trim and adjust) any existing csum items in the log for this range.
4335	 */
4336	ret = btrfs_del_csums(trans, log_root, sums->logical, sums->len);
4337	if (!ret)
4338		ret = btrfs_csum_file_blocks(trans, log_root, sums);
4339
4340	unlock_extent(&log_root->log_csum_range, sums->logical, lock_end,
4341		      &cached_state);
4342
4343	return ret;
4344}
4345
4346static noinline int copy_items(struct btrfs_trans_handle *trans,
4347			       struct btrfs_inode *inode,
4348			       struct btrfs_path *dst_path,
4349			       struct btrfs_path *src_path,
4350			       int start_slot, int nr, int inode_only,
4351			       u64 logged_isize, struct btrfs_log_ctx *ctx)
4352{
4353	struct btrfs_root *log = inode->root->log_root;
4354	struct btrfs_file_extent_item *extent;
4355	struct extent_buffer *src;
4356	int ret;
4357	struct btrfs_key *ins_keys;
4358	u32 *ins_sizes;
4359	struct btrfs_item_batch batch;
4360	char *ins_data;
4361	int dst_index;
4362	const bool skip_csum = (inode->flags & BTRFS_INODE_NODATASUM);
4363	const u64 i_size = i_size_read(&inode->vfs_inode);
4364
4365	/*
4366	 * To keep lockdep happy and avoid deadlocks, clone the source leaf and
4367	 * use the clone. This is because otherwise we would be changing the log
4368	 * tree, to insert items from the subvolume tree or insert csum items,
4369	 * while holding a read lock on a leaf from the subvolume tree, which
4370	 * creates a nasty lock dependency when COWing log tree nodes/leaves:
4371	 *
4372	 * 1) Modifying the log tree triggers an extent buffer allocation while
4373	 *    holding a write lock on a parent extent buffer from the log tree.
4374	 *    Allocating the pages for an extent buffer, or the extent buffer
4375	 *    struct, can trigger inode eviction and finally the inode eviction
4376	 *    will trigger a release/remove of a delayed node, which requires
4377	 *    taking the delayed node's mutex;
4378	 *
4379	 * 2) Allocating a metadata extent for a log tree can trigger the async
4380	 *    reclaim thread and make us wait for it to release enough space and
4381	 *    unblock our reservation ticket. The reclaim thread can start
4382	 *    flushing delayed items, and that in turn results in the need to
4383	 *    lock delayed node mutexes and in the need to write lock extent
4384	 *    buffers of a subvolume tree - all this while holding a write lock
4385	 *    on the parent extent buffer in the log tree.
4386	 *
4387	 * So one task in scenario 1) running in parallel with another task in
4388	 * scenario 2) could lead to a deadlock, one wanting to lock a delayed
4389	 * node mutex while having a read lock on a leaf from the subvolume,
4390	 * while the other is holding the delayed node's mutex and wants to
4391	 * write lock the same subvolume leaf for flushing delayed items.
4392	 */
4393	ret = clone_leaf(src_path, ctx);
4394	if (ret < 0)
4395		return ret;
4396
4397	src = src_path->nodes[0];
4398
4399	ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
4400			   nr * sizeof(u32), GFP_NOFS);
4401	if (!ins_data)
4402		return -ENOMEM;
4403
4404	ins_sizes = (u32 *)ins_data;
4405	ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
4406	batch.keys = ins_keys;
4407	batch.data_sizes = ins_sizes;
4408	batch.total_data_size = 0;
4409	batch.nr = 0;
4410
4411	dst_index = 0;
4412	for (int i = 0; i < nr; i++) {
4413		const int src_slot = start_slot + i;
4414		struct btrfs_root *csum_root;
4415		struct btrfs_ordered_sum *sums;
4416		struct btrfs_ordered_sum *sums_next;
4417		LIST_HEAD(ordered_sums);
4418		u64 disk_bytenr;
4419		u64 disk_num_bytes;
4420		u64 extent_offset;
4421		u64 extent_num_bytes;
4422		bool is_old_extent;
4423
4424		btrfs_item_key_to_cpu(src, &ins_keys[dst_index], src_slot);
4425
4426		if (ins_keys[dst_index].type != BTRFS_EXTENT_DATA_KEY)
4427			goto add_to_batch;
4428
4429		extent = btrfs_item_ptr(src, src_slot,
4430					struct btrfs_file_extent_item);
4431
4432		is_old_extent = (btrfs_file_extent_generation(src, extent) <
4433				 trans->transid);
4434
4435		/*
4436		 * Don't copy extents from past generations. That would make us
4437		 * log a lot more metadata for common cases like doing only a
4438		 * few random writes into a file and then fsync it for the first
4439		 * time or after the full sync flag is set on the inode. We can
4440		 * get leaves full of extent items, most of which are from past
4441		 * generations, so we can skip them - as long as the inode has
4442		 * not been the target of a reflink operation in this transaction,
4443		 * as in that case it might have had file extent items with old
4444		 * generations copied into it. We also must always log prealloc
4445		 * extents that start at or beyond eof, otherwise we would lose
4446		 * them on log replay.
4447		 */
4448		if (is_old_extent &&
4449		    ins_keys[dst_index].offset < i_size &&
4450		    inode->last_reflink_trans < trans->transid)
4451			continue;
4452
4453		if (skip_csum)
4454			goto add_to_batch;
4455
4456		/* Only regular extents have checksums. */
4457		if (btrfs_file_extent_type(src, extent) != BTRFS_FILE_EXTENT_REG)
4458			goto add_to_batch;
4459
4460		/*
4461		 * If it's an extent created in a past transaction, then its
4462		 * checksums are already accessible from the committed csum tree,
4463		 * no need to log them.
4464		 */
4465		if (is_old_extent)
4466			goto add_to_batch;
4467
4468		disk_bytenr = btrfs_file_extent_disk_bytenr(src, extent);
4469		/* If it's an explicit hole, there are no checksums. */
4470		if (disk_bytenr == 0)
4471			goto add_to_batch;
4472
4473		disk_num_bytes = btrfs_file_extent_disk_num_bytes(src, extent);
4474
4475		if (btrfs_file_extent_compression(src, extent)) {
4476			extent_offset = 0;
4477			extent_num_bytes = disk_num_bytes;
4478		} else {
4479			extent_offset = btrfs_file_extent_offset(src, extent);
4480			extent_num_bytes = btrfs_file_extent_num_bytes(src, extent);
4481		}
4482
4483		csum_root = btrfs_csum_root(trans->fs_info, disk_bytenr);
4484		disk_bytenr += extent_offset;
4485		ret = btrfs_lookup_csums_list(csum_root, disk_bytenr,
4486					      disk_bytenr + extent_num_bytes - 1,
4487					      &ordered_sums, false);
4488		if (ret < 0)
4489			goto out;
4490		ret = 0;
4491
4492		list_for_each_entry_safe(sums, sums_next, &ordered_sums, list) {
4493			if (!ret)
4494				ret = log_csums(trans, inode, log, sums);
4495			list_del(&sums->list);
4496			kfree(sums);
4497		}
4498		if (ret)
4499			goto out;
4500
4501add_to_batch:
4502		ins_sizes[dst_index] = btrfs_item_size(src, src_slot);
4503		batch.total_data_size += ins_sizes[dst_index];
4504		batch.nr++;
4505		dst_index++;
4506	}
4507
4508	/*
4509	 * We have a leaf full of old extent items that don't need to be logged,
4510	 * so we don't need to do anything.
4511	 */
4512	if (batch.nr == 0)
4513		goto out;
4514
4515	ret = btrfs_insert_empty_items(trans, log, dst_path, &batch);
4516	if (ret)
4517		goto out;
4518
4519	dst_index = 0;
4520	for (int i = 0; i < nr; i++) {
4521		const int src_slot = start_slot + i;
4522		const int dst_slot = dst_path->slots[0] + dst_index;
4523		struct btrfs_key key;
4524		unsigned long src_offset;
4525		unsigned long dst_offset;
4526
4527		/*
4528		 * We're done, all the remaining items in the source leaf
4529		 * correspond to old file extent items.
4530		 */
4531		if (dst_index >= batch.nr)
4532			break;
4533
4534		btrfs_item_key_to_cpu(src, &key, src_slot);
4535
4536		if (key.type != BTRFS_EXTENT_DATA_KEY)
4537			goto copy_item;
4538
4539		extent = btrfs_item_ptr(src, src_slot,
4540					struct btrfs_file_extent_item);
4541
4542		/* See the comment in the previous loop, same logic. */
4543		if (btrfs_file_extent_generation(src, extent) < trans->transid &&
4544		    key.offset < i_size &&
4545		    inode->last_reflink_trans < trans->transid)
4546			continue;
4547
4548copy_item:
4549		dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0], dst_slot);
4550		src_offset = btrfs_item_ptr_offset(src, src_slot);
4551
4552		if (key.type == BTRFS_INODE_ITEM_KEY) {
4553			struct btrfs_inode_item *inode_item;
4554
4555			inode_item = btrfs_item_ptr(dst_path->nodes[0], dst_slot,
4556						    struct btrfs_inode_item);
4557			fill_inode_item(trans, dst_path->nodes[0], inode_item,
4558					&inode->vfs_inode,
4559					inode_only == LOG_INODE_EXISTS,
4560					logged_isize);
4561		} else {
4562			copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
4563					   src_offset, ins_sizes[dst_index]);
4564		}
4565
4566		dst_index++;
4567	}
4568
4569	btrfs_mark_buffer_dirty(trans, dst_path->nodes[0]);
4570	btrfs_release_path(dst_path);
4571out:
4572	kfree(ins_data);
4573
4574	return ret;
4575}
4576
4577static int extent_cmp(void *priv, const struct list_head *a,
4578		      const struct list_head *b)
4579{
4580	const struct extent_map *em1, *em2;
4581
4582	em1 = list_entry(a, struct extent_map, list);
4583	em2 = list_entry(b, struct extent_map, list);
4584
4585	if (em1->start < em2->start)
4586		return -1;
4587	else if (em1->start > em2->start)
4588		return 1;
4589	return 0;
4590}
4591
4592static int log_extent_csums(struct btrfs_trans_handle *trans,
4593			    struct btrfs_inode *inode,
4594			    struct btrfs_root *log_root,
4595			    const struct extent_map *em,
4596			    struct btrfs_log_ctx *ctx)
4597{
4598	struct btrfs_ordered_extent *ordered;
4599	struct btrfs_root *csum_root;
4600	u64 block_start;
4601	u64 csum_offset;
4602	u64 csum_len;
4603	u64 mod_start = em->start;
4604	u64 mod_len = em->len;
4605	LIST_HEAD(ordered_sums);
4606	int ret = 0;
4607
4608	if (inode->flags & BTRFS_INODE_NODATASUM ||
4609	    (em->flags & EXTENT_FLAG_PREALLOC) ||
4610	    em->disk_bytenr == EXTENT_MAP_HOLE)
4611		return 0;
4612
4613	list_for_each_entry(ordered, &ctx->ordered_extents, log_list) {
4614		const u64 ordered_end = ordered->file_offset + ordered->num_bytes;
4615		const u64 mod_end = mod_start + mod_len;
4616		struct btrfs_ordered_sum *sums;
4617
4618		if (mod_len == 0)
4619			break;
4620
4621		if (ordered_end <= mod_start)
4622			continue;
4623		if (mod_end <= ordered->file_offset)
4624			break;
4625
4626		/*
4627		 * We are going to copy all the csums on this ordered extent, so
4628		 * go ahead and adjust mod_start and mod_len in case this ordered
4629		 * extent has already been logged.
4630		 */
4631		if (ordered->file_offset > mod_start) {
4632			if (ordered_end >= mod_end)
4633				mod_len = ordered->file_offset - mod_start;
4634			/*
4635			 * If we have this case
4636			 *
4637			 * |--------- logged extent ---------|
4638			 *       |----- ordered extent ----|
4639			 *
4640			 * Just don't mess with mod_start and mod_len, we'll
4641			 * just end up logging more csums than we need and it
4642			 * will be ok.
4643			 */
4644		} else {
4645			if (ordered_end < mod_end) {
4646				mod_len = mod_end - ordered_end;
4647				mod_start = ordered_end;
4648			} else {
4649				mod_len = 0;
4650			}
4651		}
4652
4653		/*
4654		 * To keep us from looping for the above case of an ordered
4655		 * extent that falls inside of the logged extent.
4656		 */
4657		if (test_and_set_bit(BTRFS_ORDERED_LOGGED_CSUM, &ordered->flags))
4658			continue;
4659
4660		list_for_each_entry(sums, &ordered->list, list) {
4661			ret = log_csums(trans, inode, log_root, sums);
4662			if (ret)
4663				return ret;
4664		}
4665	}
4666
4667	/* We're done, found all csums in the ordered extents. */
4668	if (mod_len == 0)
4669		return 0;
4670
4671	/* If we're compressed we have to save the entire range of csums. */
4672	if (extent_map_is_compressed(em)) {
4673		csum_offset = 0;
4674		csum_len = em->disk_num_bytes;
4675	} else {
4676		csum_offset = mod_start - em->start;
4677		csum_len = mod_len;
4678	}
4679
4680	/* block start is already adjusted for the file extent offset. */
4681	block_start = extent_map_block_start(em);
4682	csum_root = btrfs_csum_root(trans->fs_info, block_start);
4683	ret = btrfs_lookup_csums_list(csum_root, block_start + csum_offset,
4684				      block_start + csum_offset + csum_len - 1,
4685				      &ordered_sums, false);
4686	if (ret < 0)
4687		return ret;
4688	ret = 0;
4689
4690	while (!list_empty(&ordered_sums)) {
4691		struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
4692						   struct btrfs_ordered_sum,
4693						   list);
4694		if (!ret)
4695			ret = log_csums(trans, inode, log_root, sums);
4696		list_del(&sums->list);
4697		kfree(sums);
4698	}
4699
4700	return ret;
4701}
4702
4703static int log_one_extent(struct btrfs_trans_handle *trans,
4704			  struct btrfs_inode *inode,
4705			  const struct extent_map *em,
4706			  struct btrfs_path *path,
4707			  struct btrfs_log_ctx *ctx)
4708{
4709	struct btrfs_drop_extents_args drop_args = { 0 };
4710	struct btrfs_root *log = inode->root->log_root;
4711	struct btrfs_file_extent_item fi = { 0 };
4712	struct extent_buffer *leaf;
4713	struct btrfs_key key;
4714	enum btrfs_compression_type compress_type;
4715	u64 extent_offset = em->offset;
4716	u64 block_start = extent_map_block_start(em);
4717	u64 block_len;
4718	int ret;
4719
4720	btrfs_set_stack_file_extent_generation(&fi, trans->transid);
4721	if (em->flags & EXTENT_FLAG_PREALLOC)
4722		btrfs_set_stack_file_extent_type(&fi, BTRFS_FILE_EXTENT_PREALLOC);
4723	else
4724		btrfs_set_stack_file_extent_type(&fi, BTRFS_FILE_EXTENT_REG);
4725
4726	block_len = em->disk_num_bytes;
4727	compress_type = extent_map_compression(em);
4728	if (compress_type != BTRFS_COMPRESS_NONE) {
4729		btrfs_set_stack_file_extent_disk_bytenr(&fi, block_start);
4730		btrfs_set_stack_file_extent_disk_num_bytes(&fi, block_len);
4731	} else if (em->disk_bytenr < EXTENT_MAP_LAST_BYTE) {
4732		btrfs_set_stack_file_extent_disk_bytenr(&fi, block_start - extent_offset);
 
4733		btrfs_set_stack_file_extent_disk_num_bytes(&fi, block_len);
4734	}
4735
4736	btrfs_set_stack_file_extent_offset(&fi, extent_offset);
4737	btrfs_set_stack_file_extent_num_bytes(&fi, em->len);
4738	btrfs_set_stack_file_extent_ram_bytes(&fi, em->ram_bytes);
4739	btrfs_set_stack_file_extent_compression(&fi, compress_type);
4740
4741	ret = log_extent_csums(trans, inode, log, em, ctx);
4742	if (ret)
4743		return ret;
4744
4745	/*
4746	 * If this is the first time we are logging the inode in the current
4747	 * transaction, we can avoid btrfs_drop_extents(), which is expensive
4748	 * because it does a deletion search, which always acquires write locks
4749	 * for extent buffers at levels 2, 1 and 0. This not only wastes time
4750	 * but also adds significant contention in a log tree, since log trees
4751	 * are small, with a root at level 2 or 3 at most, due to their short
4752	 * life span.
4753	 */
4754	if (ctx->logged_before) {
4755		drop_args.path = path;
4756		drop_args.start = em->start;
4757		drop_args.end = em->start + em->len;
4758		drop_args.replace_extent = true;
4759		drop_args.extent_item_size = sizeof(fi);
4760		ret = btrfs_drop_extents(trans, log, inode, &drop_args);
4761		if (ret)
4762			return ret;
4763	}
4764
4765	if (!drop_args.extent_inserted) {
4766		key.objectid = btrfs_ino(inode);
4767		key.type = BTRFS_EXTENT_DATA_KEY;
4768		key.offset = em->start;
4769
4770		ret = btrfs_insert_empty_item(trans, log, path, &key,
4771					      sizeof(fi));
4772		if (ret)
4773			return ret;
4774	}
4775	leaf = path->nodes[0];
4776	write_extent_buffer(leaf, &fi,
4777			    btrfs_item_ptr_offset(leaf, path->slots[0]),
4778			    sizeof(fi));
4779	btrfs_mark_buffer_dirty(trans, leaf);
4780
4781	btrfs_release_path(path);
4782
4783	return ret;
4784}
4785
4786/*
4787 * Log all prealloc extents beyond the inode's i_size to make sure we do not
4788 * lose them after doing a full/fast fsync and replaying the log. We scan the
4789 * subvolume's root instead of iterating the inode's extent map tree because
4790 * otherwise we can log incorrect extent items based on extent map conversion.
4791 * That can happen due to the fact that extent maps are merged when they
4792 * are not in the extent map tree's list of modified extents.
4793 */
4794static int btrfs_log_prealloc_extents(struct btrfs_trans_handle *trans,
4795				      struct btrfs_inode *inode,
4796				      struct btrfs_path *path,
4797				      struct btrfs_log_ctx *ctx)
4798{
4799	struct btrfs_root *root = inode->root;
4800	struct btrfs_key key;
4801	const u64 i_size = i_size_read(&inode->vfs_inode);
4802	const u64 ino = btrfs_ino(inode);
4803	struct btrfs_path *dst_path = NULL;
4804	bool dropped_extents = false;
4805	u64 truncate_offset = i_size;
4806	struct extent_buffer *leaf;
4807	int slot;
4808	int ins_nr = 0;
4809	int start_slot = 0;
4810	int ret;
4811
4812	if (!(inode->flags & BTRFS_INODE_PREALLOC))
4813		return 0;
4814
4815	key.objectid = ino;
4816	key.type = BTRFS_EXTENT_DATA_KEY;
4817	key.offset = i_size;
4818	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4819	if (ret < 0)
4820		goto out;
4821
4822	/*
4823	 * We must check if there is a prealloc extent that starts before the
4824	 * i_size and crosses the i_size boundary. This is to ensure later we
4825	 * truncate down to the end of that extent and not to the i_size, as
4826	 * otherwise we end up losing part of the prealloc extent after a log
4827	 * replay and with an implicit hole if there is another prealloc extent
4828	 * that starts at an offset beyond i_size.
4829	 */
4830	ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
4831	if (ret < 0)
4832		goto out;
4833
4834	if (ret == 0) {
4835		struct btrfs_file_extent_item *ei;
4836
4837		leaf = path->nodes[0];
4838		slot = path->slots[0];
4839		ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4840
4841		if (btrfs_file_extent_type(leaf, ei) ==
4842		    BTRFS_FILE_EXTENT_PREALLOC) {
4843			u64 extent_end;
4844
4845			btrfs_item_key_to_cpu(leaf, &key, slot);
4846			extent_end = key.offset +
4847				btrfs_file_extent_num_bytes(leaf, ei);
4848
4849			if (extent_end > i_size)
4850				truncate_offset = extent_end;
4851		}
4852	} else {
4853		ret = 0;
4854	}
4855
4856	while (true) {
4857		leaf = path->nodes[0];
4858		slot = path->slots[0];
4859
4860		if (slot >= btrfs_header_nritems(leaf)) {
4861			if (ins_nr > 0) {
4862				ret = copy_items(trans, inode, dst_path, path,
4863						 start_slot, ins_nr, 1, 0, ctx);
4864				if (ret < 0)
4865					goto out;
4866				ins_nr = 0;
4867			}
4868			ret = btrfs_next_leaf(root, path);
4869			if (ret < 0)
4870				goto out;
4871			if (ret > 0) {
4872				ret = 0;
4873				break;
4874			}
4875			continue;
4876		}
4877
4878		btrfs_item_key_to_cpu(leaf, &key, slot);
4879		if (key.objectid > ino)
4880			break;
4881		if (WARN_ON_ONCE(key.objectid < ino) ||
4882		    key.type < BTRFS_EXTENT_DATA_KEY ||
4883		    key.offset < i_size) {
4884			path->slots[0]++;
4885			continue;
4886		}
4887		/*
4888		 * Avoid overlapping items in the log tree. The first time we
4889		 * get here, get rid of everything from a past fsync. After
4890		 * that, if the current extent starts before the end of the last
4891		 * extent we copied, truncate the last one. This can happen if
4892		 * an ordered extent completion modifies the subvolume tree
4893		 * while btrfs_next_leaf() has the tree unlocked.
4894		 */
4895		if (!dropped_extents || key.offset < truncate_offset) {
4896			ret = truncate_inode_items(trans, root->log_root, inode,
4897						   min(key.offset, truncate_offset),
4898						   BTRFS_EXTENT_DATA_KEY);
4899			if (ret)
4900				goto out;
4901			dropped_extents = true;
4902		}
4903		truncate_offset = btrfs_file_extent_end(path);
4904		if (ins_nr == 0)
4905			start_slot = slot;
4906		ins_nr++;
4907		path->slots[0]++;
4908		if (!dst_path) {
4909			dst_path = btrfs_alloc_path();
4910			if (!dst_path) {
4911				ret = -ENOMEM;
4912				goto out;
4913			}
4914		}
4915	}
4916	if (ins_nr > 0)
4917		ret = copy_items(trans, inode, dst_path, path,
4918				 start_slot, ins_nr, 1, 0, ctx);
4919out:
4920	btrfs_release_path(path);
4921	btrfs_free_path(dst_path);
4922	return ret;
4923}
4924
4925static int btrfs_log_changed_extents(struct btrfs_trans_handle *trans,
4926				     struct btrfs_inode *inode,
4927				     struct btrfs_path *path,
4928				     struct btrfs_log_ctx *ctx)
4929{
4930	struct btrfs_ordered_extent *ordered;
4931	struct btrfs_ordered_extent *tmp;
4932	struct extent_map *em, *n;
4933	LIST_HEAD(extents);
4934	struct extent_map_tree *tree = &inode->extent_tree;
4935	int ret = 0;
4936	int num = 0;
4937
4938	write_lock(&tree->lock);
4939
4940	list_for_each_entry_safe(em, n, &tree->modified_extents, list) {
4941		list_del_init(&em->list);
4942		/*
4943		 * Just an arbitrary number, this can be really CPU intensive
4944		 * once we start getting a lot of extents, and really once we
4945		 * have a bunch of extents we just want to commit since it will
4946		 * be faster.
4947		 */
4948		if (++num > 32768) {
4949			list_del_init(&tree->modified_extents);
4950			ret = -EFBIG;
4951			goto process;
4952		}
4953
4954		if (em->generation < trans->transid)
4955			continue;
4956
4957		/* We log prealloc extents beyond eof later. */
4958		if ((em->flags & EXTENT_FLAG_PREALLOC) &&
4959		    em->start >= i_size_read(&inode->vfs_inode))
4960			continue;
4961
4962		/* Need a ref to keep it from getting evicted from cache */
4963		refcount_inc(&em->refs);
4964		em->flags |= EXTENT_FLAG_LOGGING;
4965		list_add_tail(&em->list, &extents);
4966		num++;
4967	}
4968
4969	list_sort(NULL, &extents, extent_cmp);
4970process:
4971	while (!list_empty(&extents)) {
4972		em = list_entry(extents.next, struct extent_map, list);
4973
4974		list_del_init(&em->list);
4975
4976		/*
4977		 * If we had an error we just need to delete everybody from our
4978		 * private list.
4979		 */
4980		if (ret) {
4981			clear_em_logging(inode, em);
4982			free_extent_map(em);
4983			continue;
4984		}
4985
4986		write_unlock(&tree->lock);
4987
4988		ret = log_one_extent(trans, inode, em, path, ctx);
4989		write_lock(&tree->lock);
4990		clear_em_logging(inode, em);
4991		free_extent_map(em);
4992	}
4993	WARN_ON(!list_empty(&extents));
4994	write_unlock(&tree->lock);
4995
4996	if (!ret)
4997		ret = btrfs_log_prealloc_extents(trans, inode, path, ctx);
4998	if (ret)
4999		return ret;
5000
5001	/*
5002	 * We have logged all extents successfully, now make sure the commit of
5003	 * the current transaction waits for the ordered extents to complete
5004	 * before it commits and wipes out the log trees, otherwise we would
5005	 * lose data if an ordered extents completes after the transaction
5006	 * commits and a power failure happens after the transaction commit.
5007	 */
5008	list_for_each_entry_safe(ordered, tmp, &ctx->ordered_extents, log_list) {
5009		list_del_init(&ordered->log_list);
5010		set_bit(BTRFS_ORDERED_LOGGED, &ordered->flags);
5011
5012		if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
5013			spin_lock_irq(&inode->ordered_tree_lock);
5014			if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
5015				set_bit(BTRFS_ORDERED_PENDING, &ordered->flags);
5016				atomic_inc(&trans->transaction->pending_ordered);
5017			}
5018			spin_unlock_irq(&inode->ordered_tree_lock);
5019		}
5020		btrfs_put_ordered_extent(ordered);
5021	}
5022
5023	return 0;
5024}
5025
5026static int logged_inode_size(struct btrfs_root *log, struct btrfs_inode *inode,
5027			     struct btrfs_path *path, u64 *size_ret)
5028{
5029	struct btrfs_key key;
5030	int ret;
5031
5032	key.objectid = btrfs_ino(inode);
5033	key.type = BTRFS_INODE_ITEM_KEY;
5034	key.offset = 0;
5035
5036	ret = btrfs_search_slot(NULL, log, &key, path, 0, 0);
5037	if (ret < 0) {
5038		return ret;
5039	} else if (ret > 0) {
5040		*size_ret = 0;
5041	} else {
5042		struct btrfs_inode_item *item;
5043
5044		item = btrfs_item_ptr(path->nodes[0], path->slots[0],
5045				      struct btrfs_inode_item);
5046		*size_ret = btrfs_inode_size(path->nodes[0], item);
5047		/*
5048		 * If the in-memory inode's i_size is smaller then the inode
5049		 * size stored in the btree, return the inode's i_size, so
5050		 * that we get a correct inode size after replaying the log
5051		 * when before a power failure we had a shrinking truncate
5052		 * followed by addition of a new name (rename / new hard link).
5053		 * Otherwise return the inode size from the btree, to avoid
5054		 * data loss when replaying a log due to previously doing a
5055		 * write that expands the inode's size and logging a new name
5056		 * immediately after.
5057		 */
5058		if (*size_ret > inode->vfs_inode.i_size)
5059			*size_ret = inode->vfs_inode.i_size;
5060	}
5061
5062	btrfs_release_path(path);
5063	return 0;
5064}
5065
5066/*
5067 * At the moment we always log all xattrs. This is to figure out at log replay
5068 * time which xattrs must have their deletion replayed. If a xattr is missing
5069 * in the log tree and exists in the fs/subvol tree, we delete it. This is
5070 * because if a xattr is deleted, the inode is fsynced and a power failure
5071 * happens, causing the log to be replayed the next time the fs is mounted,
5072 * we want the xattr to not exist anymore (same behaviour as other filesystems
5073 * with a journal, ext3/4, xfs, f2fs, etc).
5074 */
5075static int btrfs_log_all_xattrs(struct btrfs_trans_handle *trans,
5076				struct btrfs_inode *inode,
5077				struct btrfs_path *path,
5078				struct btrfs_path *dst_path,
5079				struct btrfs_log_ctx *ctx)
5080{
5081	struct btrfs_root *root = inode->root;
5082	int ret;
5083	struct btrfs_key key;
5084	const u64 ino = btrfs_ino(inode);
5085	int ins_nr = 0;
5086	int start_slot = 0;
5087	bool found_xattrs = false;
5088
5089	if (test_bit(BTRFS_INODE_NO_XATTRS, &inode->runtime_flags))
5090		return 0;
5091
5092	key.objectid = ino;
5093	key.type = BTRFS_XATTR_ITEM_KEY;
5094	key.offset = 0;
5095
5096	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5097	if (ret < 0)
5098		return ret;
5099
5100	while (true) {
5101		int slot = path->slots[0];
5102		struct extent_buffer *leaf = path->nodes[0];
5103		int nritems = btrfs_header_nritems(leaf);
5104
5105		if (slot >= nritems) {
5106			if (ins_nr > 0) {
5107				ret = copy_items(trans, inode, dst_path, path,
5108						 start_slot, ins_nr, 1, 0, ctx);
5109				if (ret < 0)
5110					return ret;
5111				ins_nr = 0;
5112			}
5113			ret = btrfs_next_leaf(root, path);
5114			if (ret < 0)
5115				return ret;
5116			else if (ret > 0)
5117				break;
5118			continue;
5119		}
5120
5121		btrfs_item_key_to_cpu(leaf, &key, slot);
5122		if (key.objectid != ino || key.type != BTRFS_XATTR_ITEM_KEY)
5123			break;
5124
5125		if (ins_nr == 0)
5126			start_slot = slot;
5127		ins_nr++;
5128		path->slots[0]++;
5129		found_xattrs = true;
5130		cond_resched();
5131	}
5132	if (ins_nr > 0) {
5133		ret = copy_items(trans, inode, dst_path, path,
5134				 start_slot, ins_nr, 1, 0, ctx);
5135		if (ret < 0)
5136			return ret;
5137	}
5138
5139	if (!found_xattrs)
5140		set_bit(BTRFS_INODE_NO_XATTRS, &inode->runtime_flags);
5141
5142	return 0;
5143}
5144
5145/*
5146 * When using the NO_HOLES feature if we punched a hole that causes the
5147 * deletion of entire leafs or all the extent items of the first leaf (the one
5148 * that contains the inode item and references) we may end up not processing
5149 * any extents, because there are no leafs with a generation matching the
5150 * current transaction that have extent items for our inode. So we need to find
5151 * if any holes exist and then log them. We also need to log holes after any
5152 * truncate operation that changes the inode's size.
5153 */
5154static int btrfs_log_holes(struct btrfs_trans_handle *trans,
5155			   struct btrfs_inode *inode,
5156			   struct btrfs_path *path)
5157{
5158	struct btrfs_root *root = inode->root;
5159	struct btrfs_fs_info *fs_info = root->fs_info;
5160	struct btrfs_key key;
5161	const u64 ino = btrfs_ino(inode);
5162	const u64 i_size = i_size_read(&inode->vfs_inode);
5163	u64 prev_extent_end = 0;
5164	int ret;
5165
5166	if (!btrfs_fs_incompat(fs_info, NO_HOLES) || i_size == 0)
5167		return 0;
5168
5169	key.objectid = ino;
5170	key.type = BTRFS_EXTENT_DATA_KEY;
5171	key.offset = 0;
5172
5173	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5174	if (ret < 0)
5175		return ret;
5176
5177	while (true) {
5178		struct extent_buffer *leaf = path->nodes[0];
5179
5180		if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
5181			ret = btrfs_next_leaf(root, path);
5182			if (ret < 0)
5183				return ret;
5184			if (ret > 0) {
5185				ret = 0;
5186				break;
5187			}
5188			leaf = path->nodes[0];
5189		}
5190
5191		btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
5192		if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
5193			break;
5194
5195		/* We have a hole, log it. */
5196		if (prev_extent_end < key.offset) {
5197			const u64 hole_len = key.offset - prev_extent_end;
5198
5199			/*
5200			 * Release the path to avoid deadlocks with other code
5201			 * paths that search the root while holding locks on
5202			 * leafs from the log root.
5203			 */
5204			btrfs_release_path(path);
5205			ret = btrfs_insert_hole_extent(trans, root->log_root,
5206						       ino, prev_extent_end,
5207						       hole_len);
5208			if (ret < 0)
5209				return ret;
5210
5211			/*
5212			 * Search for the same key again in the root. Since it's
5213			 * an extent item and we are holding the inode lock, the
5214			 * key must still exist. If it doesn't just emit warning
5215			 * and return an error to fall back to a transaction
5216			 * commit.
5217			 */
5218			ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5219			if (ret < 0)
5220				return ret;
5221			if (WARN_ON(ret > 0))
5222				return -ENOENT;
5223			leaf = path->nodes[0];
5224		}
5225
5226		prev_extent_end = btrfs_file_extent_end(path);
5227		path->slots[0]++;
5228		cond_resched();
5229	}
5230
5231	if (prev_extent_end < i_size) {
5232		u64 hole_len;
5233
5234		btrfs_release_path(path);
5235		hole_len = ALIGN(i_size - prev_extent_end, fs_info->sectorsize);
5236		ret = btrfs_insert_hole_extent(trans, root->log_root, ino,
5237					       prev_extent_end, hole_len);
5238		if (ret < 0)
5239			return ret;
5240	}
5241
5242	return 0;
5243}
5244
5245/*
5246 * When we are logging a new inode X, check if it doesn't have a reference that
5247 * matches the reference from some other inode Y created in a past transaction
5248 * and that was renamed in the current transaction. If we don't do this, then at
5249 * log replay time we can lose inode Y (and all its files if it's a directory):
5250 *
5251 * mkdir /mnt/x
5252 * echo "hello world" > /mnt/x/foobar
5253 * sync
5254 * mv /mnt/x /mnt/y
5255 * mkdir /mnt/x                 # or touch /mnt/x
5256 * xfs_io -c fsync /mnt/x
5257 * <power fail>
5258 * mount fs, trigger log replay
5259 *
5260 * After the log replay procedure, we would lose the first directory and all its
5261 * files (file foobar).
5262 * For the case where inode Y is not a directory we simply end up losing it:
5263 *
5264 * echo "123" > /mnt/foo
5265 * sync
5266 * mv /mnt/foo /mnt/bar
5267 * echo "abc" > /mnt/foo
5268 * xfs_io -c fsync /mnt/foo
5269 * <power fail>
5270 *
5271 * We also need this for cases where a snapshot entry is replaced by some other
5272 * entry (file or directory) otherwise we end up with an unreplayable log due to
5273 * attempts to delete the snapshot entry (entry of type BTRFS_ROOT_ITEM_KEY) as
5274 * if it were a regular entry:
5275 *
5276 * mkdir /mnt/x
5277 * btrfs subvolume snapshot /mnt /mnt/x/snap
5278 * btrfs subvolume delete /mnt/x/snap
5279 * rmdir /mnt/x
5280 * mkdir /mnt/x
5281 * fsync /mnt/x or fsync some new file inside it
5282 * <power fail>
5283 *
5284 * The snapshot delete, rmdir of x, mkdir of a new x and the fsync all happen in
5285 * the same transaction.
5286 */
5287static int btrfs_check_ref_name_override(struct extent_buffer *eb,
5288					 const int slot,
5289					 const struct btrfs_key *key,
5290					 struct btrfs_inode *inode,
5291					 u64 *other_ino, u64 *other_parent)
5292{
5293	int ret;
5294	struct btrfs_path *search_path;
5295	char *name = NULL;
5296	u32 name_len = 0;
5297	u32 item_size = btrfs_item_size(eb, slot);
5298	u32 cur_offset = 0;
5299	unsigned long ptr = btrfs_item_ptr_offset(eb, slot);
5300
5301	search_path = btrfs_alloc_path();
5302	if (!search_path)
5303		return -ENOMEM;
5304	search_path->search_commit_root = 1;
5305	search_path->skip_locking = 1;
5306
5307	while (cur_offset < item_size) {
5308		u64 parent;
5309		u32 this_name_len;
5310		u32 this_len;
5311		unsigned long name_ptr;
5312		struct btrfs_dir_item *di;
5313		struct fscrypt_str name_str;
5314
5315		if (key->type == BTRFS_INODE_REF_KEY) {
5316			struct btrfs_inode_ref *iref;
5317
5318			iref = (struct btrfs_inode_ref *)(ptr + cur_offset);
5319			parent = key->offset;
5320			this_name_len = btrfs_inode_ref_name_len(eb, iref);
5321			name_ptr = (unsigned long)(iref + 1);
5322			this_len = sizeof(*iref) + this_name_len;
5323		} else {
5324			struct btrfs_inode_extref *extref;
5325
5326			extref = (struct btrfs_inode_extref *)(ptr +
5327							       cur_offset);
5328			parent = btrfs_inode_extref_parent(eb, extref);
5329			this_name_len = btrfs_inode_extref_name_len(eb, extref);
5330			name_ptr = (unsigned long)&extref->name;
5331			this_len = sizeof(*extref) + this_name_len;
5332		}
5333
5334		if (this_name_len > name_len) {
5335			char *new_name;
5336
5337			new_name = krealloc(name, this_name_len, GFP_NOFS);
5338			if (!new_name) {
5339				ret = -ENOMEM;
5340				goto out;
5341			}
5342			name_len = this_name_len;
5343			name = new_name;
5344		}
5345
5346		read_extent_buffer(eb, name, name_ptr, this_name_len);
5347
5348		name_str.name = name;
5349		name_str.len = this_name_len;
5350		di = btrfs_lookup_dir_item(NULL, inode->root, search_path,
5351				parent, &name_str, 0);
5352		if (di && !IS_ERR(di)) {
5353			struct btrfs_key di_key;
5354
5355			btrfs_dir_item_key_to_cpu(search_path->nodes[0],
5356						  di, &di_key);
5357			if (di_key.type == BTRFS_INODE_ITEM_KEY) {
5358				if (di_key.objectid != key->objectid) {
5359					ret = 1;
5360					*other_ino = di_key.objectid;
5361					*other_parent = parent;
5362				} else {
5363					ret = 0;
5364				}
5365			} else {
5366				ret = -EAGAIN;
5367			}
5368			goto out;
5369		} else if (IS_ERR(di)) {
5370			ret = PTR_ERR(di);
5371			goto out;
5372		}
5373		btrfs_release_path(search_path);
5374
5375		cur_offset += this_len;
5376	}
5377	ret = 0;
5378out:
5379	btrfs_free_path(search_path);
5380	kfree(name);
5381	return ret;
5382}
5383
5384/*
5385 * Check if we need to log an inode. This is used in contexts where while
5386 * logging an inode we need to log another inode (either that it exists or in
5387 * full mode). This is used instead of btrfs_inode_in_log() because the later
5388 * requires the inode to be in the log and have the log transaction committed,
5389 * while here we do not care if the log transaction was already committed - our
5390 * caller will commit the log later - and we want to avoid logging an inode
5391 * multiple times when multiple tasks have joined the same log transaction.
5392 */
5393static bool need_log_inode(const struct btrfs_trans_handle *trans,
5394			   struct btrfs_inode *inode)
5395{
5396	/*
5397	 * If a directory was not modified, no dentries added or removed, we can
5398	 * and should avoid logging it.
5399	 */
5400	if (S_ISDIR(inode->vfs_inode.i_mode) && inode->last_trans < trans->transid)
5401		return false;
5402
5403	/*
5404	 * If this inode does not have new/updated/deleted xattrs since the last
5405	 * time it was logged and is flagged as logged in the current transaction,
5406	 * we can skip logging it. As for new/deleted names, those are updated in
5407	 * the log by link/unlink/rename operations.
5408	 * In case the inode was logged and then evicted and reloaded, its
5409	 * logged_trans will be 0, in which case we have to fully log it since
5410	 * logged_trans is a transient field, not persisted.
5411	 */
5412	if (inode_logged(trans, inode, NULL) == 1 &&
5413	    !test_bit(BTRFS_INODE_COPY_EVERYTHING, &inode->runtime_flags))
5414		return false;
5415
5416	return true;
5417}
5418
5419struct btrfs_dir_list {
5420	u64 ino;
5421	struct list_head list;
5422};
5423
5424/*
5425 * Log the inodes of the new dentries of a directory.
5426 * See process_dir_items_leaf() for details about why it is needed.
5427 * This is a recursive operation - if an existing dentry corresponds to a
5428 * directory, that directory's new entries are logged too (same behaviour as
5429 * ext3/4, xfs, f2fs, reiserfs, nilfs2). Note that when logging the inodes
5430 * the dentries point to we do not acquire their VFS lock, otherwise lockdep
5431 * complains about the following circular lock dependency / possible deadlock:
5432 *
5433 *        CPU0                                        CPU1
5434 *        ----                                        ----
5435 * lock(&type->i_mutex_dir_key#3/2);
5436 *                                            lock(sb_internal#2);
5437 *                                            lock(&type->i_mutex_dir_key#3/2);
5438 * lock(&sb->s_type->i_mutex_key#14);
5439 *
5440 * Where sb_internal is the lock (a counter that works as a lock) acquired by
5441 * sb_start_intwrite() in btrfs_start_transaction().
5442 * Not acquiring the VFS lock of the inodes is still safe because:
5443 *
5444 * 1) For regular files we log with a mode of LOG_INODE_EXISTS. It's possible
5445 *    that while logging the inode new references (names) are added or removed
5446 *    from the inode, leaving the logged inode item with a link count that does
5447 *    not match the number of logged inode reference items. This is fine because
5448 *    at log replay time we compute the real number of links and correct the
5449 *    link count in the inode item (see replay_one_buffer() and
5450 *    link_to_fixup_dir());
5451 *
5452 * 2) For directories we log with a mode of LOG_INODE_ALL. It's possible that
5453 *    while logging the inode's items new index items (key type
5454 *    BTRFS_DIR_INDEX_KEY) are added to fs/subvol tree and the logged inode item
5455 *    has a size that doesn't match the sum of the lengths of all the logged
5456 *    names - this is ok, not a problem, because at log replay time we set the
5457 *    directory's i_size to the correct value (see replay_one_name() and
5458 *    overwrite_item()).
5459 */
5460static int log_new_dir_dentries(struct btrfs_trans_handle *trans,
5461				struct btrfs_inode *start_inode,
5462				struct btrfs_log_ctx *ctx)
5463{
5464	struct btrfs_root *root = start_inode->root;
 
5465	struct btrfs_path *path;
5466	LIST_HEAD(dir_list);
5467	struct btrfs_dir_list *dir_elem;
5468	u64 ino = btrfs_ino(start_inode);
5469	struct btrfs_inode *curr_inode = start_inode;
5470	int ret = 0;
5471
5472	/*
5473	 * If we are logging a new name, as part of a link or rename operation,
5474	 * don't bother logging new dentries, as we just want to log the names
5475	 * of an inode and that any new parents exist.
5476	 */
5477	if (ctx->logging_new_name)
5478		return 0;
5479
5480	path = btrfs_alloc_path();
5481	if (!path)
5482		return -ENOMEM;
5483
5484	/* Pairs with btrfs_add_delayed_iput below. */
5485	ihold(&curr_inode->vfs_inode);
5486
5487	while (true) {
5488		struct inode *vfs_inode;
5489		struct btrfs_key key;
5490		struct btrfs_key found_key;
5491		u64 next_index;
5492		bool continue_curr_inode = true;
5493		int iter_ret;
5494
5495		key.objectid = ino;
5496		key.type = BTRFS_DIR_INDEX_KEY;
5497		key.offset = btrfs_get_first_dir_index_to_log(curr_inode);
5498		next_index = key.offset;
5499again:
5500		btrfs_for_each_slot(root->log_root, &key, &found_key, path, iter_ret) {
5501			struct extent_buffer *leaf = path->nodes[0];
5502			struct btrfs_dir_item *di;
5503			struct btrfs_key di_key;
5504			struct inode *di_inode;
5505			int log_mode = LOG_INODE_EXISTS;
5506			int type;
5507
5508			if (found_key.objectid != ino ||
5509			    found_key.type != BTRFS_DIR_INDEX_KEY) {
5510				continue_curr_inode = false;
5511				break;
5512			}
5513
5514			next_index = found_key.offset + 1;
5515
5516			di = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dir_item);
5517			type = btrfs_dir_ftype(leaf, di);
5518			if (btrfs_dir_transid(leaf, di) < trans->transid)
5519				continue;
5520			btrfs_dir_item_key_to_cpu(leaf, di, &di_key);
5521			if (di_key.type == BTRFS_ROOT_ITEM_KEY)
5522				continue;
5523
5524			btrfs_release_path(path);
5525			di_inode = btrfs_iget_logging(di_key.objectid, root);
5526			if (IS_ERR(di_inode)) {
5527				ret = PTR_ERR(di_inode);
5528				goto out;
5529			}
5530
5531			if (!need_log_inode(trans, BTRFS_I(di_inode))) {
5532				btrfs_add_delayed_iput(BTRFS_I(di_inode));
5533				break;
5534			}
5535
5536			ctx->log_new_dentries = false;
5537			if (type == BTRFS_FT_DIR)
5538				log_mode = LOG_INODE_ALL;
5539			ret = btrfs_log_inode(trans, BTRFS_I(di_inode),
5540					      log_mode, ctx);
5541			btrfs_add_delayed_iput(BTRFS_I(di_inode));
5542			if (ret)
5543				goto out;
5544			if (ctx->log_new_dentries) {
5545				dir_elem = kmalloc(sizeof(*dir_elem), GFP_NOFS);
5546				if (!dir_elem) {
5547					ret = -ENOMEM;
5548					goto out;
5549				}
5550				dir_elem->ino = di_key.objectid;
5551				list_add_tail(&dir_elem->list, &dir_list);
5552			}
5553			break;
5554		}
5555
5556		btrfs_release_path(path);
5557
5558		if (iter_ret < 0) {
5559			ret = iter_ret;
5560			goto out;
5561		} else if (iter_ret > 0) {
5562			continue_curr_inode = false;
5563		} else {
5564			key = found_key;
5565		}
5566
5567		if (continue_curr_inode && key.offset < (u64)-1) {
5568			key.offset++;
5569			goto again;
5570		}
5571
5572		btrfs_set_first_dir_index_to_log(curr_inode, next_index);
5573
5574		if (list_empty(&dir_list))
5575			break;
5576
5577		dir_elem = list_first_entry(&dir_list, struct btrfs_dir_list, list);
5578		ino = dir_elem->ino;
5579		list_del(&dir_elem->list);
5580		kfree(dir_elem);
5581
5582		btrfs_add_delayed_iput(curr_inode);
5583		curr_inode = NULL;
5584
5585		vfs_inode = btrfs_iget_logging(ino, root);
5586		if (IS_ERR(vfs_inode)) {
5587			ret = PTR_ERR(vfs_inode);
5588			break;
5589		}
5590		curr_inode = BTRFS_I(vfs_inode);
5591	}
5592out:
5593	btrfs_free_path(path);
5594	if (curr_inode)
5595		btrfs_add_delayed_iput(curr_inode);
5596
5597	if (ret) {
5598		struct btrfs_dir_list *next;
5599
5600		list_for_each_entry_safe(dir_elem, next, &dir_list, list)
5601			kfree(dir_elem);
5602	}
5603
5604	return ret;
5605}
5606
5607struct btrfs_ino_list {
5608	u64 ino;
5609	u64 parent;
5610	struct list_head list;
5611};
5612
5613static void free_conflicting_inodes(struct btrfs_log_ctx *ctx)
5614{
5615	struct btrfs_ino_list *curr;
5616	struct btrfs_ino_list *next;
5617
5618	list_for_each_entry_safe(curr, next, &ctx->conflict_inodes, list) {
5619		list_del(&curr->list);
5620		kfree(curr);
5621	}
5622}
5623
5624static int conflicting_inode_is_dir(struct btrfs_root *root, u64 ino,
5625				    struct btrfs_path *path)
5626{
5627	struct btrfs_key key;
5628	int ret;
5629
5630	key.objectid = ino;
5631	key.type = BTRFS_INODE_ITEM_KEY;
5632	key.offset = 0;
5633
5634	path->search_commit_root = 1;
5635	path->skip_locking = 1;
5636
5637	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5638	if (WARN_ON_ONCE(ret > 0)) {
5639		/*
5640		 * We have previously found the inode through the commit root
5641		 * so this should not happen. If it does, just error out and
5642		 * fallback to a transaction commit.
5643		 */
5644		ret = -ENOENT;
5645	} else if (ret == 0) {
5646		struct btrfs_inode_item *item;
5647
5648		item = btrfs_item_ptr(path->nodes[0], path->slots[0],
5649				      struct btrfs_inode_item);
5650		if (S_ISDIR(btrfs_inode_mode(path->nodes[0], item)))
5651			ret = 1;
5652	}
5653
5654	btrfs_release_path(path);
5655	path->search_commit_root = 0;
5656	path->skip_locking = 0;
5657
5658	return ret;
5659}
5660
5661static int add_conflicting_inode(struct btrfs_trans_handle *trans,
5662				 struct btrfs_root *root,
5663				 struct btrfs_path *path,
5664				 u64 ino, u64 parent,
5665				 struct btrfs_log_ctx *ctx)
5666{
5667	struct btrfs_ino_list *ino_elem;
5668	struct inode *inode;
5669
5670	/*
5671	 * It's rare to have a lot of conflicting inodes, in practice it is not
5672	 * common to have more than 1 or 2. We don't want to collect too many,
5673	 * as we could end up logging too many inodes (even if only in
5674	 * LOG_INODE_EXISTS mode) and slow down other fsyncs or transaction
5675	 * commits.
5676	 */
5677	if (ctx->num_conflict_inodes >= MAX_CONFLICT_INODES)
5678		return BTRFS_LOG_FORCE_COMMIT;
5679
5680	inode = btrfs_iget_logging(ino, root);
5681	/*
5682	 * If the other inode that had a conflicting dir entry was deleted in
5683	 * the current transaction then we either:
5684	 *
5685	 * 1) Log the parent directory (later after adding it to the list) if
5686	 *    the inode is a directory. This is because it may be a deleted
5687	 *    subvolume/snapshot or it may be a regular directory that had
5688	 *    deleted subvolumes/snapshots (or subdirectories that had them),
5689	 *    and at the moment we can't deal with dropping subvolumes/snapshots
5690	 *    during log replay. So we just log the parent, which will result in
5691	 *    a fallback to a transaction commit if we are dealing with those
5692	 *    cases (last_unlink_trans will match the current transaction);
5693	 *
5694	 * 2) Do nothing if it's not a directory. During log replay we simply
5695	 *    unlink the conflicting dentry from the parent directory and then
5696	 *    add the dentry for our inode. Like this we can avoid logging the
5697	 *    parent directory (and maybe fallback to a transaction commit in
5698	 *    case it has a last_unlink_trans == trans->transid, due to moving
5699	 *    some inode from it to some other directory).
5700	 */
5701	if (IS_ERR(inode)) {
5702		int ret = PTR_ERR(inode);
5703
5704		if (ret != -ENOENT)
5705			return ret;
5706
5707		ret = conflicting_inode_is_dir(root, ino, path);
5708		/* Not a directory or we got an error. */
5709		if (ret <= 0)
5710			return ret;
5711
5712		/* Conflicting inode is a directory, so we'll log its parent. */
5713		ino_elem = kmalloc(sizeof(*ino_elem), GFP_NOFS);
5714		if (!ino_elem)
5715			return -ENOMEM;
5716		ino_elem->ino = ino;
5717		ino_elem->parent = parent;
5718		list_add_tail(&ino_elem->list, &ctx->conflict_inodes);
5719		ctx->num_conflict_inodes++;
5720
5721		return 0;
5722	}
5723
5724	/*
5725	 * If the inode was already logged skip it - otherwise we can hit an
5726	 * infinite loop. Example:
5727	 *
5728	 * From the commit root (previous transaction) we have the following
5729	 * inodes:
5730	 *
5731	 * inode 257 a directory
5732	 * inode 258 with references "zz" and "zz_link" on inode 257
5733	 * inode 259 with reference "a" on inode 257
5734	 *
5735	 * And in the current (uncommitted) transaction we have:
5736	 *
5737	 * inode 257 a directory, unchanged
5738	 * inode 258 with references "a" and "a2" on inode 257
5739	 * inode 259 with reference "zz_link" on inode 257
5740	 * inode 261 with reference "zz" on inode 257
5741	 *
5742	 * When logging inode 261 the following infinite loop could
5743	 * happen if we don't skip already logged inodes:
5744	 *
5745	 * - we detect inode 258 as a conflicting inode, with inode 261
5746	 *   on reference "zz", and log it;
5747	 *
5748	 * - we detect inode 259 as a conflicting inode, with inode 258
5749	 *   on reference "a", and log it;
5750	 *
5751	 * - we detect inode 258 as a conflicting inode, with inode 259
5752	 *   on reference "zz_link", and log it - again! After this we
5753	 *   repeat the above steps forever.
5754	 *
5755	 * Here we can use need_log_inode() because we only need to log the
5756	 * inode in LOG_INODE_EXISTS mode and rename operations update the log,
5757	 * so that the log ends up with the new name and without the old name.
5758	 */
5759	if (!need_log_inode(trans, BTRFS_I(inode))) {
5760		btrfs_add_delayed_iput(BTRFS_I(inode));
5761		return 0;
5762	}
5763
5764	btrfs_add_delayed_iput(BTRFS_I(inode));
5765
5766	ino_elem = kmalloc(sizeof(*ino_elem), GFP_NOFS);
5767	if (!ino_elem)
5768		return -ENOMEM;
5769	ino_elem->ino = ino;
5770	ino_elem->parent = parent;
5771	list_add_tail(&ino_elem->list, &ctx->conflict_inodes);
5772	ctx->num_conflict_inodes++;
5773
5774	return 0;
5775}
5776
5777static int log_conflicting_inodes(struct btrfs_trans_handle *trans,
5778				  struct btrfs_root *root,
5779				  struct btrfs_log_ctx *ctx)
5780{
 
5781	int ret = 0;
5782
5783	/*
5784	 * Conflicting inodes are logged by the first call to btrfs_log_inode(),
5785	 * otherwise we could have unbounded recursion of btrfs_log_inode()
5786	 * calls. This check guarantees we can have only 1 level of recursion.
5787	 */
5788	if (ctx->logging_conflict_inodes)
5789		return 0;
5790
5791	ctx->logging_conflict_inodes = true;
5792
5793	/*
5794	 * New conflicting inodes may be found and added to the list while we
5795	 * are logging a conflicting inode, so keep iterating while the list is
5796	 * not empty.
5797	 */
5798	while (!list_empty(&ctx->conflict_inodes)) {
5799		struct btrfs_ino_list *curr;
5800		struct inode *inode;
5801		u64 ino;
5802		u64 parent;
5803
5804		curr = list_first_entry(&ctx->conflict_inodes,
5805					struct btrfs_ino_list, list);
5806		ino = curr->ino;
5807		parent = curr->parent;
5808		list_del(&curr->list);
5809		kfree(curr);
5810
5811		inode = btrfs_iget_logging(ino, root);
5812		/*
5813		 * If the other inode that had a conflicting dir entry was
5814		 * deleted in the current transaction, we need to log its parent
5815		 * directory. See the comment at add_conflicting_inode().
5816		 */
5817		if (IS_ERR(inode)) {
5818			ret = PTR_ERR(inode);
5819			if (ret != -ENOENT)
5820				break;
5821
5822			inode = btrfs_iget_logging(parent, root);
5823			if (IS_ERR(inode)) {
5824				ret = PTR_ERR(inode);
5825				break;
5826			}
5827
5828			/*
5829			 * Always log the directory, we cannot make this
5830			 * conditional on need_log_inode() because the directory
5831			 * might have been logged in LOG_INODE_EXISTS mode or
5832			 * the dir index of the conflicting inode is not in a
5833			 * dir index key range logged for the directory. So we
5834			 * must make sure the deletion is recorded.
5835			 */
5836			ret = btrfs_log_inode(trans, BTRFS_I(inode),
5837					      LOG_INODE_ALL, ctx);
5838			btrfs_add_delayed_iput(BTRFS_I(inode));
5839			if (ret)
5840				break;
5841			continue;
5842		}
5843
5844		/*
5845		 * Here we can use need_log_inode() because we only need to log
5846		 * the inode in LOG_INODE_EXISTS mode and rename operations
5847		 * update the log, so that the log ends up with the new name and
5848		 * without the old name.
5849		 *
5850		 * We did this check at add_conflicting_inode(), but here we do
5851		 * it again because if some other task logged the inode after
5852		 * that, we can avoid doing it again.
5853		 */
5854		if (!need_log_inode(trans, BTRFS_I(inode))) {
5855			btrfs_add_delayed_iput(BTRFS_I(inode));
5856			continue;
5857		}
5858
5859		/*
5860		 * We are safe logging the other inode without acquiring its
5861		 * lock as long as we log with the LOG_INODE_EXISTS mode. We
5862		 * are safe against concurrent renames of the other inode as
5863		 * well because during a rename we pin the log and update the
5864		 * log with the new name before we unpin it.
5865		 */
5866		ret = btrfs_log_inode(trans, BTRFS_I(inode), LOG_INODE_EXISTS, ctx);
5867		btrfs_add_delayed_iput(BTRFS_I(inode));
5868		if (ret)
5869			break;
5870	}
5871
5872	ctx->logging_conflict_inodes = false;
5873	if (ret)
5874		free_conflicting_inodes(ctx);
5875
5876	return ret;
5877}
5878
5879static int copy_inode_items_to_log(struct btrfs_trans_handle *trans,
5880				   struct btrfs_inode *inode,
5881				   struct btrfs_key *min_key,
5882				   const struct btrfs_key *max_key,
5883				   struct btrfs_path *path,
5884				   struct btrfs_path *dst_path,
5885				   const u64 logged_isize,
5886				   const int inode_only,
5887				   struct btrfs_log_ctx *ctx,
5888				   bool *need_log_inode_item)
5889{
5890	const u64 i_size = i_size_read(&inode->vfs_inode);
5891	struct btrfs_root *root = inode->root;
5892	int ins_start_slot = 0;
5893	int ins_nr = 0;
5894	int ret;
5895
5896	while (1) {
5897		ret = btrfs_search_forward(root, min_key, path, trans->transid);
5898		if (ret < 0)
5899			return ret;
5900		if (ret > 0) {
5901			ret = 0;
5902			break;
5903		}
5904again:
5905		/* Note, ins_nr might be > 0 here, cleanup outside the loop */
5906		if (min_key->objectid != max_key->objectid)
5907			break;
5908		if (min_key->type > max_key->type)
5909			break;
5910
5911		if (min_key->type == BTRFS_INODE_ITEM_KEY) {
5912			*need_log_inode_item = false;
5913		} else if (min_key->type == BTRFS_EXTENT_DATA_KEY &&
5914			   min_key->offset >= i_size) {
5915			/*
5916			 * Extents at and beyond eof are logged with
5917			 * btrfs_log_prealloc_extents().
5918			 * Only regular files have BTRFS_EXTENT_DATA_KEY keys,
5919			 * and no keys greater than that, so bail out.
5920			 */
5921			break;
5922		} else if ((min_key->type == BTRFS_INODE_REF_KEY ||
5923			    min_key->type == BTRFS_INODE_EXTREF_KEY) &&
5924			   (inode->generation == trans->transid ||
5925			    ctx->logging_conflict_inodes)) {
5926			u64 other_ino = 0;
5927			u64 other_parent = 0;
5928
5929			ret = btrfs_check_ref_name_override(path->nodes[0],
5930					path->slots[0], min_key, inode,
5931					&other_ino, &other_parent);
5932			if (ret < 0) {
5933				return ret;
5934			} else if (ret > 0 &&
5935				   other_ino != btrfs_ino(ctx->inode)) {
5936				if (ins_nr > 0) {
5937					ins_nr++;
5938				} else {
5939					ins_nr = 1;
5940					ins_start_slot = path->slots[0];
5941				}
5942				ret = copy_items(trans, inode, dst_path, path,
5943						 ins_start_slot, ins_nr,
5944						 inode_only, logged_isize, ctx);
5945				if (ret < 0)
5946					return ret;
5947				ins_nr = 0;
5948
5949				btrfs_release_path(path);
5950				ret = add_conflicting_inode(trans, root, path,
5951							    other_ino,
5952							    other_parent, ctx);
5953				if (ret)
5954					return ret;
5955				goto next_key;
5956			}
5957		} else if (min_key->type == BTRFS_XATTR_ITEM_KEY) {
5958			/* Skip xattrs, logged later with btrfs_log_all_xattrs() */
5959			if (ins_nr == 0)
5960				goto next_slot;
5961			ret = copy_items(trans, inode, dst_path, path,
5962					 ins_start_slot,
5963					 ins_nr, inode_only, logged_isize, ctx);
5964			if (ret < 0)
5965				return ret;
5966			ins_nr = 0;
5967			goto next_slot;
5968		}
5969
5970		if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
5971			ins_nr++;
5972			goto next_slot;
5973		} else if (!ins_nr) {
5974			ins_start_slot = path->slots[0];
5975			ins_nr = 1;
5976			goto next_slot;
5977		}
5978
5979		ret = copy_items(trans, inode, dst_path, path, ins_start_slot,
5980				 ins_nr, inode_only, logged_isize, ctx);
5981		if (ret < 0)
5982			return ret;
5983		ins_nr = 1;
5984		ins_start_slot = path->slots[0];
5985next_slot:
5986		path->slots[0]++;
5987		if (path->slots[0] < btrfs_header_nritems(path->nodes[0])) {
5988			btrfs_item_key_to_cpu(path->nodes[0], min_key,
5989					      path->slots[0]);
5990			goto again;
5991		}
5992		if (ins_nr) {
5993			ret = copy_items(trans, inode, dst_path, path,
5994					 ins_start_slot, ins_nr, inode_only,
5995					 logged_isize, ctx);
5996			if (ret < 0)
5997				return ret;
5998			ins_nr = 0;
5999		}
6000		btrfs_release_path(path);
6001next_key:
6002		if (min_key->offset < (u64)-1) {
6003			min_key->offset++;
6004		} else if (min_key->type < max_key->type) {
6005			min_key->type++;
6006			min_key->offset = 0;
6007		} else {
6008			break;
6009		}
6010
6011		/*
6012		 * We may process many leaves full of items for our inode, so
6013		 * avoid monopolizing a cpu for too long by rescheduling while
6014		 * not holding locks on any tree.
6015		 */
6016		cond_resched();
6017	}
6018	if (ins_nr) {
6019		ret = copy_items(trans, inode, dst_path, path, ins_start_slot,
6020				 ins_nr, inode_only, logged_isize, ctx);
6021		if (ret)
6022			return ret;
6023	}
6024
6025	if (inode_only == LOG_INODE_ALL && S_ISREG(inode->vfs_inode.i_mode)) {
6026		/*
6027		 * Release the path because otherwise we might attempt to double
6028		 * lock the same leaf with btrfs_log_prealloc_extents() below.
6029		 */
6030		btrfs_release_path(path);
6031		ret = btrfs_log_prealloc_extents(trans, inode, dst_path, ctx);
6032	}
6033
6034	return ret;
6035}
6036
6037static int insert_delayed_items_batch(struct btrfs_trans_handle *trans,
6038				      struct btrfs_root *log,
6039				      struct btrfs_path *path,
6040				      const struct btrfs_item_batch *batch,
6041				      const struct btrfs_delayed_item *first_item)
6042{
6043	const struct btrfs_delayed_item *curr = first_item;
6044	int ret;
6045
6046	ret = btrfs_insert_empty_items(trans, log, path, batch);
6047	if (ret)
6048		return ret;
6049
6050	for (int i = 0; i < batch->nr; i++) {
6051		char *data_ptr;
6052
6053		data_ptr = btrfs_item_ptr(path->nodes[0], path->slots[0], char);
6054		write_extent_buffer(path->nodes[0], &curr->data,
6055				    (unsigned long)data_ptr, curr->data_len);
6056		curr = list_next_entry(curr, log_list);
6057		path->slots[0]++;
6058	}
6059
6060	btrfs_release_path(path);
6061
6062	return 0;
6063}
6064
6065static int log_delayed_insertion_items(struct btrfs_trans_handle *trans,
6066				       struct btrfs_inode *inode,
6067				       struct btrfs_path *path,
6068				       const struct list_head *delayed_ins_list,
6069				       struct btrfs_log_ctx *ctx)
6070{
6071	/* 195 (4095 bytes of keys and sizes) fits in a single 4K page. */
6072	const int max_batch_size = 195;
6073	const int leaf_data_size = BTRFS_LEAF_DATA_SIZE(trans->fs_info);
6074	const u64 ino = btrfs_ino(inode);
6075	struct btrfs_root *log = inode->root->log_root;
6076	struct btrfs_item_batch batch = {
6077		.nr = 0,
6078		.total_data_size = 0,
6079	};
6080	const struct btrfs_delayed_item *first = NULL;
6081	const struct btrfs_delayed_item *curr;
6082	char *ins_data;
6083	struct btrfs_key *ins_keys;
6084	u32 *ins_sizes;
6085	u64 curr_batch_size = 0;
6086	int batch_idx = 0;
6087	int ret;
6088
6089	/* We are adding dir index items to the log tree. */
6090	lockdep_assert_held(&inode->log_mutex);
6091
6092	/*
6093	 * We collect delayed items before copying index keys from the subvolume
6094	 * to the log tree. However just after we collected them, they may have
6095	 * been flushed (all of them or just some of them), and therefore we
6096	 * could have copied them from the subvolume tree to the log tree.
6097	 * So find the first delayed item that was not yet logged (they are
6098	 * sorted by index number).
6099	 */
6100	list_for_each_entry(curr, delayed_ins_list, log_list) {
6101		if (curr->index > inode->last_dir_index_offset) {
6102			first = curr;
6103			break;
6104		}
6105	}
6106
6107	/* Empty list or all delayed items were already logged. */
6108	if (!first)
6109		return 0;
6110
6111	ins_data = kmalloc(max_batch_size * sizeof(u32) +
6112			   max_batch_size * sizeof(struct btrfs_key), GFP_NOFS);
6113	if (!ins_data)
6114		return -ENOMEM;
6115	ins_sizes = (u32 *)ins_data;
6116	batch.data_sizes = ins_sizes;
6117	ins_keys = (struct btrfs_key *)(ins_data + max_batch_size * sizeof(u32));
6118	batch.keys = ins_keys;
6119
6120	curr = first;
6121	while (!list_entry_is_head(curr, delayed_ins_list, log_list)) {
6122		const u32 curr_size = curr->data_len + sizeof(struct btrfs_item);
6123
6124		if (curr_batch_size + curr_size > leaf_data_size ||
6125		    batch.nr == max_batch_size) {
6126			ret = insert_delayed_items_batch(trans, log, path,
6127							 &batch, first);
6128			if (ret)
6129				goto out;
6130			batch_idx = 0;
6131			batch.nr = 0;
6132			batch.total_data_size = 0;
6133			curr_batch_size = 0;
6134			first = curr;
6135		}
6136
6137		ins_sizes[batch_idx] = curr->data_len;
6138		ins_keys[batch_idx].objectid = ino;
6139		ins_keys[batch_idx].type = BTRFS_DIR_INDEX_KEY;
6140		ins_keys[batch_idx].offset = curr->index;
6141		curr_batch_size += curr_size;
6142		batch.total_data_size += curr->data_len;
6143		batch.nr++;
6144		batch_idx++;
6145		curr = list_next_entry(curr, log_list);
6146	}
6147
6148	ASSERT(batch.nr >= 1);
6149	ret = insert_delayed_items_batch(trans, log, path, &batch, first);
6150
6151	curr = list_last_entry(delayed_ins_list, struct btrfs_delayed_item,
6152			       log_list);
6153	inode->last_dir_index_offset = curr->index;
6154out:
6155	kfree(ins_data);
6156
6157	return ret;
6158}
6159
6160static int log_delayed_deletions_full(struct btrfs_trans_handle *trans,
6161				      struct btrfs_inode *inode,
6162				      struct btrfs_path *path,
6163				      const struct list_head *delayed_del_list,
6164				      struct btrfs_log_ctx *ctx)
6165{
6166	const u64 ino = btrfs_ino(inode);
6167	const struct btrfs_delayed_item *curr;
6168
6169	curr = list_first_entry(delayed_del_list, struct btrfs_delayed_item,
6170				log_list);
6171
6172	while (!list_entry_is_head(curr, delayed_del_list, log_list)) {
6173		u64 first_dir_index = curr->index;
6174		u64 last_dir_index;
6175		const struct btrfs_delayed_item *next;
6176		int ret;
6177
6178		/*
6179		 * Find a range of consecutive dir index items to delete. Like
6180		 * this we log a single dir range item spanning several contiguous
6181		 * dir items instead of logging one range item per dir index item.
6182		 */
6183		next = list_next_entry(curr, log_list);
6184		while (!list_entry_is_head(next, delayed_del_list, log_list)) {
6185			if (next->index != curr->index + 1)
6186				break;
6187			curr = next;
6188			next = list_next_entry(next, log_list);
6189		}
6190
6191		last_dir_index = curr->index;
6192		ASSERT(last_dir_index >= first_dir_index);
6193
6194		ret = insert_dir_log_key(trans, inode->root->log_root, path,
6195					 ino, first_dir_index, last_dir_index);
6196		if (ret)
6197			return ret;
6198		curr = list_next_entry(curr, log_list);
6199	}
6200
6201	return 0;
6202}
6203
6204static int batch_delete_dir_index_items(struct btrfs_trans_handle *trans,
6205					struct btrfs_inode *inode,
6206					struct btrfs_path *path,
 
6207					const struct list_head *delayed_del_list,
6208					const struct btrfs_delayed_item *first,
6209					const struct btrfs_delayed_item **last_ret)
6210{
6211	const struct btrfs_delayed_item *next;
6212	struct extent_buffer *leaf = path->nodes[0];
6213	const int last_slot = btrfs_header_nritems(leaf) - 1;
6214	int slot = path->slots[0] + 1;
6215	const u64 ino = btrfs_ino(inode);
6216
6217	next = list_next_entry(first, log_list);
6218
6219	while (slot < last_slot &&
6220	       !list_entry_is_head(next, delayed_del_list, log_list)) {
6221		struct btrfs_key key;
6222
6223		btrfs_item_key_to_cpu(leaf, &key, slot);
6224		if (key.objectid != ino ||
6225		    key.type != BTRFS_DIR_INDEX_KEY ||
6226		    key.offset != next->index)
6227			break;
6228
6229		slot++;
6230		*last_ret = next;
6231		next = list_next_entry(next, log_list);
6232	}
6233
6234	return btrfs_del_items(trans, inode->root->log_root, path,
6235			       path->slots[0], slot - path->slots[0]);
6236}
6237
6238static int log_delayed_deletions_incremental(struct btrfs_trans_handle *trans,
6239					     struct btrfs_inode *inode,
6240					     struct btrfs_path *path,
6241					     const struct list_head *delayed_del_list,
6242					     struct btrfs_log_ctx *ctx)
6243{
6244	struct btrfs_root *log = inode->root->log_root;
6245	const struct btrfs_delayed_item *curr;
6246	u64 last_range_start = 0;
6247	u64 last_range_end = 0;
6248	struct btrfs_key key;
6249
6250	key.objectid = btrfs_ino(inode);
6251	key.type = BTRFS_DIR_INDEX_KEY;
6252	curr = list_first_entry(delayed_del_list, struct btrfs_delayed_item,
6253				log_list);
6254
6255	while (!list_entry_is_head(curr, delayed_del_list, log_list)) {
6256		const struct btrfs_delayed_item *last = curr;
6257		u64 first_dir_index = curr->index;
6258		u64 last_dir_index;
6259		bool deleted_items = false;
6260		int ret;
6261
6262		key.offset = curr->index;
6263		ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
6264		if (ret < 0) {
6265			return ret;
6266		} else if (ret == 0) {
6267			ret = batch_delete_dir_index_items(trans, inode, path,
6268							   delayed_del_list, curr,
6269							   &last);
6270			if (ret)
6271				return ret;
6272			deleted_items = true;
6273		}
6274
6275		btrfs_release_path(path);
6276
6277		/*
6278		 * If we deleted items from the leaf, it means we have a range
6279		 * item logging their range, so no need to add one or update an
6280		 * existing one. Otherwise we have to log a dir range item.
6281		 */
6282		if (deleted_items)
6283			goto next_batch;
6284
6285		last_dir_index = last->index;
6286		ASSERT(last_dir_index >= first_dir_index);
6287		/*
6288		 * If this range starts right after where the previous one ends,
6289		 * then we want to reuse the previous range item and change its
6290		 * end offset to the end of this range. This is just to minimize
6291		 * leaf space usage, by avoiding adding a new range item.
6292		 */
6293		if (last_range_end != 0 && first_dir_index == last_range_end + 1)
6294			first_dir_index = last_range_start;
6295
6296		ret = insert_dir_log_key(trans, log, path, key.objectid,
6297					 first_dir_index, last_dir_index);
6298		if (ret)
6299			return ret;
6300
6301		last_range_start = first_dir_index;
6302		last_range_end = last_dir_index;
6303next_batch:
6304		curr = list_next_entry(last, log_list);
6305	}
6306
6307	return 0;
6308}
6309
6310static int log_delayed_deletion_items(struct btrfs_trans_handle *trans,
6311				      struct btrfs_inode *inode,
6312				      struct btrfs_path *path,
6313				      const struct list_head *delayed_del_list,
6314				      struct btrfs_log_ctx *ctx)
6315{
6316	/*
6317	 * We are deleting dir index items from the log tree or adding range
6318	 * items to it.
6319	 */
6320	lockdep_assert_held(&inode->log_mutex);
6321
6322	if (list_empty(delayed_del_list))
6323		return 0;
6324
6325	if (ctx->logged_before)
6326		return log_delayed_deletions_incremental(trans, inode, path,
6327							 delayed_del_list, ctx);
6328
6329	return log_delayed_deletions_full(trans, inode, path, delayed_del_list,
6330					  ctx);
6331}
6332
6333/*
6334 * Similar logic as for log_new_dir_dentries(), but it iterates over the delayed
6335 * items instead of the subvolume tree.
6336 */
6337static int log_new_delayed_dentries(struct btrfs_trans_handle *trans,
6338				    struct btrfs_inode *inode,
6339				    const struct list_head *delayed_ins_list,
6340				    struct btrfs_log_ctx *ctx)
6341{
6342	const bool orig_log_new_dentries = ctx->log_new_dentries;
 
6343	struct btrfs_delayed_item *item;
6344	int ret = 0;
6345
6346	/*
6347	 * No need for the log mutex, plus to avoid potential deadlocks or
6348	 * lockdep annotations due to nesting of delayed inode mutexes and log
6349	 * mutexes.
6350	 */
6351	lockdep_assert_not_held(&inode->log_mutex);
6352
6353	ASSERT(!ctx->logging_new_delayed_dentries);
6354	ctx->logging_new_delayed_dentries = true;
6355
6356	list_for_each_entry(item, delayed_ins_list, log_list) {
6357		struct btrfs_dir_item *dir_item;
6358		struct inode *di_inode;
6359		struct btrfs_key key;
6360		int log_mode = LOG_INODE_EXISTS;
6361
6362		dir_item = (struct btrfs_dir_item *)item->data;
6363		btrfs_disk_key_to_cpu(&key, &dir_item->location);
6364
6365		if (key.type == BTRFS_ROOT_ITEM_KEY)
6366			continue;
6367
6368		di_inode = btrfs_iget_logging(key.objectid, inode->root);
6369		if (IS_ERR(di_inode)) {
6370			ret = PTR_ERR(di_inode);
6371			break;
6372		}
6373
6374		if (!need_log_inode(trans, BTRFS_I(di_inode))) {
6375			btrfs_add_delayed_iput(BTRFS_I(di_inode));
6376			continue;
6377		}
6378
6379		if (btrfs_stack_dir_ftype(dir_item) == BTRFS_FT_DIR)
6380			log_mode = LOG_INODE_ALL;
6381
6382		ctx->log_new_dentries = false;
6383		ret = btrfs_log_inode(trans, BTRFS_I(di_inode), log_mode, ctx);
6384
6385		if (!ret && ctx->log_new_dentries)
6386			ret = log_new_dir_dentries(trans, BTRFS_I(di_inode), ctx);
6387
6388		btrfs_add_delayed_iput(BTRFS_I(di_inode));
6389
6390		if (ret)
6391			break;
6392	}
6393
6394	ctx->log_new_dentries = orig_log_new_dentries;
6395	ctx->logging_new_delayed_dentries = false;
6396
6397	return ret;
6398}
6399
6400/* log a single inode in the tree log.
6401 * At least one parent directory for this inode must exist in the tree
6402 * or be logged already.
6403 *
6404 * Any items from this inode changed by the current transaction are copied
6405 * to the log tree.  An extra reference is taken on any extents in this
6406 * file, allowing us to avoid a whole pile of corner cases around logging
6407 * blocks that have been removed from the tree.
6408 *
6409 * See LOG_INODE_ALL and related defines for a description of what inode_only
6410 * does.
6411 *
6412 * This handles both files and directories.
6413 */
6414static int btrfs_log_inode(struct btrfs_trans_handle *trans,
6415			   struct btrfs_inode *inode,
6416			   int inode_only,
6417			   struct btrfs_log_ctx *ctx)
6418{
6419	struct btrfs_path *path;
6420	struct btrfs_path *dst_path;
6421	struct btrfs_key min_key;
6422	struct btrfs_key max_key;
6423	struct btrfs_root *log = inode->root->log_root;
6424	int ret;
6425	bool fast_search = false;
6426	u64 ino = btrfs_ino(inode);
6427	struct extent_map_tree *em_tree = &inode->extent_tree;
6428	u64 logged_isize = 0;
6429	bool need_log_inode_item = true;
6430	bool xattrs_logged = false;
6431	bool inode_item_dropped = true;
6432	bool full_dir_logging = false;
6433	LIST_HEAD(delayed_ins_list);
6434	LIST_HEAD(delayed_del_list);
6435
6436	path = btrfs_alloc_path();
6437	if (!path)
6438		return -ENOMEM;
6439	dst_path = btrfs_alloc_path();
6440	if (!dst_path) {
6441		btrfs_free_path(path);
6442		return -ENOMEM;
6443	}
6444
6445	min_key.objectid = ino;
6446	min_key.type = BTRFS_INODE_ITEM_KEY;
6447	min_key.offset = 0;
6448
6449	max_key.objectid = ino;
6450
6451
6452	/* today the code can only do partial logging of directories */
6453	if (S_ISDIR(inode->vfs_inode.i_mode) ||
6454	    (!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
6455		       &inode->runtime_flags) &&
6456	     inode_only >= LOG_INODE_EXISTS))
6457		max_key.type = BTRFS_XATTR_ITEM_KEY;
6458	else
6459		max_key.type = (u8)-1;
6460	max_key.offset = (u64)-1;
6461
6462	if (S_ISDIR(inode->vfs_inode.i_mode) && inode_only == LOG_INODE_ALL)
6463		full_dir_logging = true;
6464
6465	/*
6466	 * If we are logging a directory while we are logging dentries of the
6467	 * delayed items of some other inode, then we need to flush the delayed
6468	 * items of this directory and not log the delayed items directly. This
6469	 * is to prevent more than one level of recursion into btrfs_log_inode()
6470	 * by having something like this:
6471	 *
6472	 *     $ mkdir -p a/b/c/d/e/f/g/h/...
6473	 *     $ xfs_io -c "fsync" a
6474	 *
6475	 * Where all directories in the path did not exist before and are
6476	 * created in the current transaction.
6477	 * So in such a case we directly log the delayed items of the main
6478	 * directory ("a") without flushing them first, while for each of its
6479	 * subdirectories we flush their delayed items before logging them.
6480	 * This prevents a potential unbounded recursion like this:
6481	 *
6482	 * btrfs_log_inode()
6483	 *   log_new_delayed_dentries()
6484	 *      btrfs_log_inode()
6485	 *        log_new_delayed_dentries()
6486	 *          btrfs_log_inode()
6487	 *            log_new_delayed_dentries()
6488	 *              (...)
6489	 *
6490	 * We have thresholds for the maximum number of delayed items to have in
6491	 * memory, and once they are hit, the items are flushed asynchronously.
6492	 * However the limit is quite high, so lets prevent deep levels of
6493	 * recursion to happen by limiting the maximum depth to be 1.
6494	 */
6495	if (full_dir_logging && ctx->logging_new_delayed_dentries) {
6496		ret = btrfs_commit_inode_delayed_items(trans, inode);
6497		if (ret)
6498			goto out;
6499	}
6500
6501	mutex_lock(&inode->log_mutex);
6502
6503	/*
6504	 * For symlinks, we must always log their content, which is stored in an
6505	 * inline extent, otherwise we could end up with an empty symlink after
6506	 * log replay, which is invalid on linux (symlink(2) returns -ENOENT if
6507	 * one attempts to create an empty symlink).
6508	 * We don't need to worry about flushing delalloc, because when we create
6509	 * the inline extent when the symlink is created (we never have delalloc
6510	 * for symlinks).
6511	 */
6512	if (S_ISLNK(inode->vfs_inode.i_mode))
6513		inode_only = LOG_INODE_ALL;
6514
6515	/*
6516	 * Before logging the inode item, cache the value returned by
6517	 * inode_logged(), because after that we have the need to figure out if
6518	 * the inode was previously logged in this transaction.
6519	 */
6520	ret = inode_logged(trans, inode, path);
6521	if (ret < 0)
6522		goto out_unlock;
6523	ctx->logged_before = (ret == 1);
6524	ret = 0;
6525
6526	/*
6527	 * This is for cases where logging a directory could result in losing a
6528	 * a file after replaying the log. For example, if we move a file from a
6529	 * directory A to a directory B, then fsync directory A, we have no way
6530	 * to known the file was moved from A to B, so logging just A would
6531	 * result in losing the file after a log replay.
6532	 */
6533	if (full_dir_logging && inode->last_unlink_trans >= trans->transid) {
6534		ret = BTRFS_LOG_FORCE_COMMIT;
6535		goto out_unlock;
6536	}
6537
6538	/*
6539	 * a brute force approach to making sure we get the most uptodate
6540	 * copies of everything.
6541	 */
6542	if (S_ISDIR(inode->vfs_inode.i_mode)) {
6543		clear_bit(BTRFS_INODE_COPY_EVERYTHING, &inode->runtime_flags);
6544		if (ctx->logged_before)
6545			ret = drop_inode_items(trans, log, path, inode,
6546					       BTRFS_XATTR_ITEM_KEY);
6547	} else {
6548		if (inode_only == LOG_INODE_EXISTS && ctx->logged_before) {
6549			/*
6550			 * Make sure the new inode item we write to the log has
6551			 * the same isize as the current one (if it exists).
6552			 * This is necessary to prevent data loss after log
6553			 * replay, and also to prevent doing a wrong expanding
6554			 * truncate - for e.g. create file, write 4K into offset
6555			 * 0, fsync, write 4K into offset 4096, add hard link,
6556			 * fsync some other file (to sync log), power fail - if
6557			 * we use the inode's current i_size, after log replay
6558			 * we get a 8Kb file, with the last 4Kb extent as a hole
6559			 * (zeroes), as if an expanding truncate happened,
6560			 * instead of getting a file of 4Kb only.
6561			 */
6562			ret = logged_inode_size(log, inode, path, &logged_isize);
6563			if (ret)
6564				goto out_unlock;
6565		}
6566		if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
6567			     &inode->runtime_flags)) {
6568			if (inode_only == LOG_INODE_EXISTS) {
6569				max_key.type = BTRFS_XATTR_ITEM_KEY;
6570				if (ctx->logged_before)
6571					ret = drop_inode_items(trans, log, path,
6572							       inode, max_key.type);
6573			} else {
6574				clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
6575					  &inode->runtime_flags);
6576				clear_bit(BTRFS_INODE_COPY_EVERYTHING,
6577					  &inode->runtime_flags);
6578				if (ctx->logged_before)
6579					ret = truncate_inode_items(trans, log,
6580								   inode, 0, 0);
6581			}
6582		} else if (test_and_clear_bit(BTRFS_INODE_COPY_EVERYTHING,
6583					      &inode->runtime_flags) ||
6584			   inode_only == LOG_INODE_EXISTS) {
6585			if (inode_only == LOG_INODE_ALL)
6586				fast_search = true;
6587			max_key.type = BTRFS_XATTR_ITEM_KEY;
6588			if (ctx->logged_before)
6589				ret = drop_inode_items(trans, log, path, inode,
6590						       max_key.type);
6591		} else {
6592			if (inode_only == LOG_INODE_ALL)
6593				fast_search = true;
6594			inode_item_dropped = false;
6595			goto log_extents;
6596		}
6597
6598	}
6599	if (ret)
6600		goto out_unlock;
6601
6602	/*
6603	 * If we are logging a directory in full mode, collect the delayed items
6604	 * before iterating the subvolume tree, so that we don't miss any new
6605	 * dir index items in case they get flushed while or right after we are
6606	 * iterating the subvolume tree.
6607	 */
6608	if (full_dir_logging && !ctx->logging_new_delayed_dentries)
6609		btrfs_log_get_delayed_items(inode, &delayed_ins_list,
6610					    &delayed_del_list);
6611
6612	ret = copy_inode_items_to_log(trans, inode, &min_key, &max_key,
6613				      path, dst_path, logged_isize,
6614				      inode_only, ctx,
6615				      &need_log_inode_item);
6616	if (ret)
6617		goto out_unlock;
6618
6619	btrfs_release_path(path);
6620	btrfs_release_path(dst_path);
6621	ret = btrfs_log_all_xattrs(trans, inode, path, dst_path, ctx);
6622	if (ret)
6623		goto out_unlock;
6624	xattrs_logged = true;
6625	if (max_key.type >= BTRFS_EXTENT_DATA_KEY && !fast_search) {
6626		btrfs_release_path(path);
6627		btrfs_release_path(dst_path);
6628		ret = btrfs_log_holes(trans, inode, path);
6629		if (ret)
6630			goto out_unlock;
6631	}
6632log_extents:
6633	btrfs_release_path(path);
6634	btrfs_release_path(dst_path);
6635	if (need_log_inode_item) {
6636		ret = log_inode_item(trans, log, dst_path, inode, inode_item_dropped);
6637		if (ret)
6638			goto out_unlock;
6639		/*
6640		 * If we are doing a fast fsync and the inode was logged before
6641		 * in this transaction, we don't need to log the xattrs because
6642		 * they were logged before. If xattrs were added, changed or
6643		 * deleted since the last time we logged the inode, then we have
6644		 * already logged them because the inode had the runtime flag
6645		 * BTRFS_INODE_COPY_EVERYTHING set.
6646		 */
6647		if (!xattrs_logged && inode->logged_trans < trans->transid) {
6648			ret = btrfs_log_all_xattrs(trans, inode, path, dst_path, ctx);
6649			if (ret)
6650				goto out_unlock;
6651			btrfs_release_path(path);
6652		}
6653	}
6654	if (fast_search) {
6655		ret = btrfs_log_changed_extents(trans, inode, dst_path, ctx);
6656		if (ret)
6657			goto out_unlock;
6658	} else if (inode_only == LOG_INODE_ALL) {
6659		struct extent_map *em, *n;
6660
6661		write_lock(&em_tree->lock);
6662		list_for_each_entry_safe(em, n, &em_tree->modified_extents, list)
6663			list_del_init(&em->list);
6664		write_unlock(&em_tree->lock);
6665	}
6666
6667	if (full_dir_logging) {
6668		ret = log_directory_changes(trans, inode, path, dst_path, ctx);
6669		if (ret)
6670			goto out_unlock;
6671		ret = log_delayed_insertion_items(trans, inode, path,
6672						  &delayed_ins_list, ctx);
6673		if (ret)
6674			goto out_unlock;
6675		ret = log_delayed_deletion_items(trans, inode, path,
6676						 &delayed_del_list, ctx);
6677		if (ret)
6678			goto out_unlock;
6679	}
6680
6681	spin_lock(&inode->lock);
6682	inode->logged_trans = trans->transid;
6683	/*
6684	 * Don't update last_log_commit if we logged that an inode exists.
6685	 * We do this for three reasons:
6686	 *
6687	 * 1) We might have had buffered writes to this inode that were
6688	 *    flushed and had their ordered extents completed in this
6689	 *    transaction, but we did not previously log the inode with
6690	 *    LOG_INODE_ALL. Later the inode was evicted and after that
6691	 *    it was loaded again and this LOG_INODE_EXISTS log operation
6692	 *    happened. We must make sure that if an explicit fsync against
6693	 *    the inode is performed later, it logs the new extents, an
6694	 *    updated inode item, etc, and syncs the log. The same logic
6695	 *    applies to direct IO writes instead of buffered writes.
6696	 *
6697	 * 2) When we log the inode with LOG_INODE_EXISTS, its inode item
6698	 *    is logged with an i_size of 0 or whatever value was logged
6699	 *    before. If later the i_size of the inode is increased by a
6700	 *    truncate operation, the log is synced through an fsync of
6701	 *    some other inode and then finally an explicit fsync against
6702	 *    this inode is made, we must make sure this fsync logs the
6703	 *    inode with the new i_size, the hole between old i_size and
6704	 *    the new i_size, and syncs the log.
6705	 *
6706	 * 3) If we are logging that an ancestor inode exists as part of
6707	 *    logging a new name from a link or rename operation, don't update
6708	 *    its last_log_commit - otherwise if an explicit fsync is made
6709	 *    against an ancestor, the fsync considers the inode in the log
6710	 *    and doesn't sync the log, resulting in the ancestor missing after
6711	 *    a power failure unless the log was synced as part of an fsync
6712	 *    against any other unrelated inode.
6713	 */
6714	if (inode_only != LOG_INODE_EXISTS)
6715		inode->last_log_commit = inode->last_sub_trans;
6716	spin_unlock(&inode->lock);
6717
6718	/*
6719	 * Reset the last_reflink_trans so that the next fsync does not need to
6720	 * go through the slower path when logging extents and their checksums.
6721	 */
6722	if (inode_only == LOG_INODE_ALL)
6723		inode->last_reflink_trans = 0;
6724
6725out_unlock:
6726	mutex_unlock(&inode->log_mutex);
6727out:
6728	btrfs_free_path(path);
6729	btrfs_free_path(dst_path);
6730
6731	if (ret)
6732		free_conflicting_inodes(ctx);
6733	else
6734		ret = log_conflicting_inodes(trans, inode->root, ctx);
6735
6736	if (full_dir_logging && !ctx->logging_new_delayed_dentries) {
6737		if (!ret)
6738			ret = log_new_delayed_dentries(trans, inode,
6739						       &delayed_ins_list, ctx);
6740
6741		btrfs_log_put_delayed_items(inode, &delayed_ins_list,
6742					    &delayed_del_list);
6743	}
6744
6745	return ret;
6746}
6747
6748static int btrfs_log_all_parents(struct btrfs_trans_handle *trans,
6749				 struct btrfs_inode *inode,
6750				 struct btrfs_log_ctx *ctx)
6751{
 
6752	int ret;
6753	struct btrfs_path *path;
6754	struct btrfs_key key;
6755	struct btrfs_root *root = inode->root;
6756	const u64 ino = btrfs_ino(inode);
6757
6758	path = btrfs_alloc_path();
6759	if (!path)
6760		return -ENOMEM;
6761	path->skip_locking = 1;
6762	path->search_commit_root = 1;
6763
6764	key.objectid = ino;
6765	key.type = BTRFS_INODE_REF_KEY;
6766	key.offset = 0;
6767	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6768	if (ret < 0)
6769		goto out;
6770
6771	while (true) {
6772		struct extent_buffer *leaf = path->nodes[0];
6773		int slot = path->slots[0];
6774		u32 cur_offset = 0;
6775		u32 item_size;
6776		unsigned long ptr;
6777
6778		if (slot >= btrfs_header_nritems(leaf)) {
6779			ret = btrfs_next_leaf(root, path);
6780			if (ret < 0)
6781				goto out;
6782			else if (ret > 0)
6783				break;
6784			continue;
6785		}
6786
6787		btrfs_item_key_to_cpu(leaf, &key, slot);
6788		/* BTRFS_INODE_EXTREF_KEY is BTRFS_INODE_REF_KEY + 1 */
6789		if (key.objectid != ino || key.type > BTRFS_INODE_EXTREF_KEY)
6790			break;
6791
6792		item_size = btrfs_item_size(leaf, slot);
6793		ptr = btrfs_item_ptr_offset(leaf, slot);
6794		while (cur_offset < item_size) {
6795			struct btrfs_key inode_key;
6796			struct inode *dir_inode;
6797
6798			inode_key.type = BTRFS_INODE_ITEM_KEY;
6799			inode_key.offset = 0;
6800
6801			if (key.type == BTRFS_INODE_EXTREF_KEY) {
6802				struct btrfs_inode_extref *extref;
6803
6804				extref = (struct btrfs_inode_extref *)
6805					(ptr + cur_offset);
6806				inode_key.objectid = btrfs_inode_extref_parent(
6807					leaf, extref);
6808				cur_offset += sizeof(*extref);
6809				cur_offset += btrfs_inode_extref_name_len(leaf,
6810					extref);
6811			} else {
6812				inode_key.objectid = key.offset;
6813				cur_offset = item_size;
6814			}
6815
6816			dir_inode = btrfs_iget_logging(inode_key.objectid, root);
 
6817			/*
6818			 * If the parent inode was deleted, return an error to
6819			 * fallback to a transaction commit. This is to prevent
6820			 * getting an inode that was moved from one parent A to
6821			 * a parent B, got its former parent A deleted and then
6822			 * it got fsync'ed, from existing at both parents after
6823			 * a log replay (and the old parent still existing).
6824			 * Example:
6825			 *
6826			 * mkdir /mnt/A
6827			 * mkdir /mnt/B
6828			 * touch /mnt/B/bar
6829			 * sync
6830			 * mv /mnt/B/bar /mnt/A/bar
6831			 * mv -T /mnt/A /mnt/B
6832			 * fsync /mnt/B/bar
6833			 * <power fail>
6834			 *
6835			 * If we ignore the old parent B which got deleted,
6836			 * after a log replay we would have file bar linked
6837			 * at both parents and the old parent B would still
6838			 * exist.
6839			 */
6840			if (IS_ERR(dir_inode)) {
6841				ret = PTR_ERR(dir_inode);
6842				goto out;
6843			}
6844
6845			if (!need_log_inode(trans, BTRFS_I(dir_inode))) {
6846				btrfs_add_delayed_iput(BTRFS_I(dir_inode));
6847				continue;
6848			}
6849
6850			ctx->log_new_dentries = false;
6851			ret = btrfs_log_inode(trans, BTRFS_I(dir_inode),
6852					      LOG_INODE_ALL, ctx);
6853			if (!ret && ctx->log_new_dentries)
6854				ret = log_new_dir_dentries(trans,
6855						   BTRFS_I(dir_inode), ctx);
6856			btrfs_add_delayed_iput(BTRFS_I(dir_inode));
6857			if (ret)
6858				goto out;
6859		}
6860		path->slots[0]++;
6861	}
6862	ret = 0;
6863out:
6864	btrfs_free_path(path);
6865	return ret;
6866}
6867
6868static int log_new_ancestors(struct btrfs_trans_handle *trans,
6869			     struct btrfs_root *root,
6870			     struct btrfs_path *path,
6871			     struct btrfs_log_ctx *ctx)
6872{
6873	struct btrfs_key found_key;
6874
6875	btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
6876
6877	while (true) {
 
6878		struct extent_buffer *leaf;
6879		int slot;
6880		struct btrfs_key search_key;
6881		struct inode *inode;
6882		u64 ino;
6883		int ret = 0;
6884
6885		btrfs_release_path(path);
6886
6887		ino = found_key.offset;
6888
6889		search_key.objectid = found_key.offset;
6890		search_key.type = BTRFS_INODE_ITEM_KEY;
6891		search_key.offset = 0;
6892		inode = btrfs_iget_logging(ino, root);
6893		if (IS_ERR(inode))
6894			return PTR_ERR(inode);
6895
6896		if (BTRFS_I(inode)->generation >= trans->transid &&
6897		    need_log_inode(trans, BTRFS_I(inode)))
6898			ret = btrfs_log_inode(trans, BTRFS_I(inode),
6899					      LOG_INODE_EXISTS, ctx);
6900		btrfs_add_delayed_iput(BTRFS_I(inode));
6901		if (ret)
6902			return ret;
6903
6904		if (search_key.objectid == BTRFS_FIRST_FREE_OBJECTID)
6905			break;
6906
6907		search_key.type = BTRFS_INODE_REF_KEY;
6908		ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
6909		if (ret < 0)
6910			return ret;
6911
6912		leaf = path->nodes[0];
6913		slot = path->slots[0];
6914		if (slot >= btrfs_header_nritems(leaf)) {
6915			ret = btrfs_next_leaf(root, path);
6916			if (ret < 0)
6917				return ret;
6918			else if (ret > 0)
6919				return -ENOENT;
6920			leaf = path->nodes[0];
6921			slot = path->slots[0];
6922		}
6923
6924		btrfs_item_key_to_cpu(leaf, &found_key, slot);
6925		if (found_key.objectid != search_key.objectid ||
6926		    found_key.type != BTRFS_INODE_REF_KEY)
6927			return -ENOENT;
6928	}
6929	return 0;
6930}
6931
6932static int log_new_ancestors_fast(struct btrfs_trans_handle *trans,
6933				  struct btrfs_inode *inode,
6934				  struct dentry *parent,
6935				  struct btrfs_log_ctx *ctx)
6936{
6937	struct btrfs_root *root = inode->root;
6938	struct dentry *old_parent = NULL;
6939	struct super_block *sb = inode->vfs_inode.i_sb;
6940	int ret = 0;
6941
6942	while (true) {
6943		if (!parent || d_really_is_negative(parent) ||
6944		    sb != parent->d_sb)
6945			break;
6946
6947		inode = BTRFS_I(d_inode(parent));
6948		if (root != inode->root)
6949			break;
6950
6951		if (inode->generation >= trans->transid &&
6952		    need_log_inode(trans, inode)) {
6953			ret = btrfs_log_inode(trans, inode,
6954					      LOG_INODE_EXISTS, ctx);
6955			if (ret)
6956				break;
6957		}
6958		if (IS_ROOT(parent))
6959			break;
6960
6961		parent = dget_parent(parent);
6962		dput(old_parent);
6963		old_parent = parent;
6964	}
6965	dput(old_parent);
6966
6967	return ret;
6968}
6969
6970static int log_all_new_ancestors(struct btrfs_trans_handle *trans,
6971				 struct btrfs_inode *inode,
6972				 struct dentry *parent,
6973				 struct btrfs_log_ctx *ctx)
6974{
6975	struct btrfs_root *root = inode->root;
6976	const u64 ino = btrfs_ino(inode);
6977	struct btrfs_path *path;
6978	struct btrfs_key search_key;
6979	int ret;
6980
6981	/*
6982	 * For a single hard link case, go through a fast path that does not
6983	 * need to iterate the fs/subvolume tree.
6984	 */
6985	if (inode->vfs_inode.i_nlink < 2)
6986		return log_new_ancestors_fast(trans, inode, parent, ctx);
6987
6988	path = btrfs_alloc_path();
6989	if (!path)
6990		return -ENOMEM;
6991
6992	search_key.objectid = ino;
6993	search_key.type = BTRFS_INODE_REF_KEY;
6994	search_key.offset = 0;
6995again:
6996	ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
6997	if (ret < 0)
6998		goto out;
6999	if (ret == 0)
7000		path->slots[0]++;
7001
7002	while (true) {
7003		struct extent_buffer *leaf = path->nodes[0];
7004		int slot = path->slots[0];
7005		struct btrfs_key found_key;
7006
7007		if (slot >= btrfs_header_nritems(leaf)) {
7008			ret = btrfs_next_leaf(root, path);
7009			if (ret < 0)
7010				goto out;
7011			else if (ret > 0)
7012				break;
7013			continue;
7014		}
7015
7016		btrfs_item_key_to_cpu(leaf, &found_key, slot);
7017		if (found_key.objectid != ino ||
7018		    found_key.type > BTRFS_INODE_EXTREF_KEY)
7019			break;
7020
7021		/*
7022		 * Don't deal with extended references because they are rare
7023		 * cases and too complex to deal with (we would need to keep
7024		 * track of which subitem we are processing for each item in
7025		 * this loop, etc). So just return some error to fallback to
7026		 * a transaction commit.
7027		 */
7028		if (found_key.type == BTRFS_INODE_EXTREF_KEY) {
7029			ret = -EMLINK;
7030			goto out;
7031		}
7032
7033		/*
7034		 * Logging ancestors needs to do more searches on the fs/subvol
7035		 * tree, so it releases the path as needed to avoid deadlocks.
7036		 * Keep track of the last inode ref key and resume from that key
7037		 * after logging all new ancestors for the current hard link.
7038		 */
7039		memcpy(&search_key, &found_key, sizeof(search_key));
7040
7041		ret = log_new_ancestors(trans, root, path, ctx);
7042		if (ret)
7043			goto out;
7044		btrfs_release_path(path);
7045		goto again;
7046	}
7047	ret = 0;
7048out:
7049	btrfs_free_path(path);
7050	return ret;
7051}
7052
7053/*
7054 * helper function around btrfs_log_inode to make sure newly created
7055 * parent directories also end up in the log.  A minimal inode and backref
7056 * only logging is done of any parent directories that are older than
7057 * the last committed transaction
7058 */
7059static int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
7060				  struct btrfs_inode *inode,
7061				  struct dentry *parent,
7062				  int inode_only,
7063				  struct btrfs_log_ctx *ctx)
7064{
7065	struct btrfs_root *root = inode->root;
7066	struct btrfs_fs_info *fs_info = root->fs_info;
7067	int ret = 0;
7068	bool log_dentries = false;
7069
7070	if (btrfs_test_opt(fs_info, NOTREELOG)) {
7071		ret = BTRFS_LOG_FORCE_COMMIT;
7072		goto end_no_trans;
7073	}
7074
7075	if (btrfs_root_refs(&root->root_item) == 0) {
7076		ret = BTRFS_LOG_FORCE_COMMIT;
7077		goto end_no_trans;
7078	}
7079
7080	/*
7081	 * If we're logging an inode from a subvolume created in the current
7082	 * transaction we must force a commit since the root is not persisted.
7083	 */
7084	if (btrfs_root_generation(&root->root_item) == trans->transid) {
7085		ret = BTRFS_LOG_FORCE_COMMIT;
7086		goto end_no_trans;
7087	}
7088
7089	/*
7090	 * Skip already logged inodes or inodes corresponding to tmpfiles
7091	 * (since logging them is pointless, a link count of 0 means they
7092	 * will never be accessible).
7093	 */
7094	if ((btrfs_inode_in_log(inode, trans->transid) &&
7095	     list_empty(&ctx->ordered_extents)) ||
7096	    inode->vfs_inode.i_nlink == 0) {
7097		ret = BTRFS_NO_LOG_SYNC;
7098		goto end_no_trans;
7099	}
7100
7101	ret = start_log_trans(trans, root, ctx);
7102	if (ret)
7103		goto end_no_trans;
7104
7105	ret = btrfs_log_inode(trans, inode, inode_only, ctx);
7106	if (ret)
7107		goto end_trans;
7108
7109	/*
7110	 * for regular files, if its inode is already on disk, we don't
7111	 * have to worry about the parents at all.  This is because
7112	 * we can use the last_unlink_trans field to record renames
7113	 * and other fun in this file.
7114	 */
7115	if (S_ISREG(inode->vfs_inode.i_mode) &&
7116	    inode->generation < trans->transid &&
7117	    inode->last_unlink_trans < trans->transid) {
7118		ret = 0;
7119		goto end_trans;
7120	}
7121
7122	if (S_ISDIR(inode->vfs_inode.i_mode) && ctx->log_new_dentries)
7123		log_dentries = true;
7124
7125	/*
7126	 * On unlink we must make sure all our current and old parent directory
7127	 * inodes are fully logged. This is to prevent leaving dangling
7128	 * directory index entries in directories that were our parents but are
7129	 * not anymore. Not doing this results in old parent directory being
7130	 * impossible to delete after log replay (rmdir will always fail with
7131	 * error -ENOTEMPTY).
7132	 *
7133	 * Example 1:
7134	 *
7135	 * mkdir testdir
7136	 * touch testdir/foo
7137	 * ln testdir/foo testdir/bar
7138	 * sync
7139	 * unlink testdir/bar
7140	 * xfs_io -c fsync testdir/foo
7141	 * <power failure>
7142	 * mount fs, triggers log replay
7143	 *
7144	 * If we don't log the parent directory (testdir), after log replay the
7145	 * directory still has an entry pointing to the file inode using the bar
7146	 * name, but a matching BTRFS_INODE_[REF|EXTREF]_KEY does not exist and
7147	 * the file inode has a link count of 1.
7148	 *
7149	 * Example 2:
7150	 *
7151	 * mkdir testdir
7152	 * touch foo
7153	 * ln foo testdir/foo2
7154	 * ln foo testdir/foo3
7155	 * sync
7156	 * unlink testdir/foo3
7157	 * xfs_io -c fsync foo
7158	 * <power failure>
7159	 * mount fs, triggers log replay
7160	 *
7161	 * Similar as the first example, after log replay the parent directory
7162	 * testdir still has an entry pointing to the inode file with name foo3
7163	 * but the file inode does not have a matching BTRFS_INODE_REF_KEY item
7164	 * and has a link count of 2.
7165	 */
7166	if (inode->last_unlink_trans >= trans->transid) {
7167		ret = btrfs_log_all_parents(trans, inode, ctx);
7168		if (ret)
7169			goto end_trans;
7170	}
7171
7172	ret = log_all_new_ancestors(trans, inode, parent, ctx);
7173	if (ret)
7174		goto end_trans;
7175
7176	if (log_dentries)
7177		ret = log_new_dir_dentries(trans, inode, ctx);
7178	else
7179		ret = 0;
7180end_trans:
7181	if (ret < 0) {
7182		btrfs_set_log_full_commit(trans);
7183		ret = BTRFS_LOG_FORCE_COMMIT;
7184	}
7185
7186	if (ret)
7187		btrfs_remove_log_ctx(root, ctx);
7188	btrfs_end_log_trans(root);
7189end_no_trans:
7190	return ret;
7191}
7192
7193/*
7194 * it is not safe to log dentry if the chunk root has added new
7195 * chunks.  This returns 0 if the dentry was logged, and 1 otherwise.
7196 * If this returns 1, you must commit the transaction to safely get your
7197 * data on disk.
7198 */
7199int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
7200			  struct dentry *dentry,
7201			  struct btrfs_log_ctx *ctx)
7202{
7203	struct dentry *parent = dget_parent(dentry);
7204	int ret;
7205
7206	ret = btrfs_log_inode_parent(trans, BTRFS_I(d_inode(dentry)), parent,
7207				     LOG_INODE_ALL, ctx);
7208	dput(parent);
7209
7210	return ret;
7211}
7212
7213/*
7214 * should be called during mount to recover any replay any log trees
7215 * from the FS
7216 */
7217int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
7218{
7219	int ret;
7220	struct btrfs_path *path;
7221	struct btrfs_trans_handle *trans;
7222	struct btrfs_key key;
7223	struct btrfs_key found_key;
7224	struct btrfs_root *log;
7225	struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
7226	struct walk_control wc = {
7227		.process_func = process_one_buffer,
7228		.stage = LOG_WALK_PIN_ONLY,
7229	};
7230
7231	path = btrfs_alloc_path();
7232	if (!path)
7233		return -ENOMEM;
7234
7235	set_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
7236
7237	trans = btrfs_start_transaction(fs_info->tree_root, 0);
7238	if (IS_ERR(trans)) {
7239		ret = PTR_ERR(trans);
7240		goto error;
7241	}
7242
7243	wc.trans = trans;
7244	wc.pin = 1;
7245
7246	ret = walk_log_tree(trans, log_root_tree, &wc);
7247	if (ret) {
7248		btrfs_abort_transaction(trans, ret);
7249		goto error;
7250	}
7251
7252again:
7253	key.objectid = BTRFS_TREE_LOG_OBJECTID;
7254	key.offset = (u64)-1;
7255	key.type = BTRFS_ROOT_ITEM_KEY;
7256
7257	while (1) {
7258		ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
7259
7260		if (ret < 0) {
7261			btrfs_abort_transaction(trans, ret);
7262			goto error;
7263		}
7264		if (ret > 0) {
7265			if (path->slots[0] == 0)
7266				break;
7267			path->slots[0]--;
7268		}
7269		btrfs_item_key_to_cpu(path->nodes[0], &found_key,
7270				      path->slots[0]);
7271		btrfs_release_path(path);
7272		if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
7273			break;
7274
7275		log = btrfs_read_tree_root(log_root_tree, &found_key);
7276		if (IS_ERR(log)) {
7277			ret = PTR_ERR(log);
7278			btrfs_abort_transaction(trans, ret);
7279			goto error;
7280		}
7281
7282		wc.replay_dest = btrfs_get_fs_root(fs_info, found_key.offset,
7283						   true);
7284		if (IS_ERR(wc.replay_dest)) {
7285			ret = PTR_ERR(wc.replay_dest);
7286
7287			/*
7288			 * We didn't find the subvol, likely because it was
7289			 * deleted.  This is ok, simply skip this log and go to
7290			 * the next one.
7291			 *
7292			 * We need to exclude the root because we can't have
7293			 * other log replays overwriting this log as we'll read
7294			 * it back in a few more times.  This will keep our
7295			 * block from being modified, and we'll just bail for
7296			 * each subsequent pass.
7297			 */
7298			if (ret == -ENOENT)
7299				ret = btrfs_pin_extent_for_log_replay(trans, log->node);
7300			btrfs_put_root(log);
7301
7302			if (!ret)
7303				goto next;
7304			btrfs_abort_transaction(trans, ret);
7305			goto error;
7306		}
7307
7308		wc.replay_dest->log_root = log;
7309		ret = btrfs_record_root_in_trans(trans, wc.replay_dest);
7310		if (ret)
7311			/* The loop needs to continue due to the root refs */
7312			btrfs_abort_transaction(trans, ret);
7313		else
7314			ret = walk_log_tree(trans, log, &wc);
7315
7316		if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
7317			ret = fixup_inode_link_counts(trans, wc.replay_dest,
7318						      path);
7319			if (ret)
7320				btrfs_abort_transaction(trans, ret);
7321		}
7322
7323		if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
7324			struct btrfs_root *root = wc.replay_dest;
7325
7326			btrfs_release_path(path);
7327
7328			/*
7329			 * We have just replayed everything, and the highest
7330			 * objectid of fs roots probably has changed in case
7331			 * some inode_item's got replayed.
7332			 *
7333			 * root->objectid_mutex is not acquired as log replay
7334			 * could only happen during mount.
7335			 */
7336			ret = btrfs_init_root_free_objectid(root);
7337			if (ret)
7338				btrfs_abort_transaction(trans, ret);
7339		}
7340
7341		wc.replay_dest->log_root = NULL;
7342		btrfs_put_root(wc.replay_dest);
7343		btrfs_put_root(log);
7344
7345		if (ret)
7346			goto error;
7347next:
7348		if (found_key.offset == 0)
7349			break;
7350		key.offset = found_key.offset - 1;
7351	}
7352	btrfs_release_path(path);
7353
7354	/* step one is to pin it all, step two is to replay just inodes */
7355	if (wc.pin) {
7356		wc.pin = 0;
7357		wc.process_func = replay_one_buffer;
7358		wc.stage = LOG_WALK_REPLAY_INODES;
7359		goto again;
7360	}
7361	/* step three is to replay everything */
7362	if (wc.stage < LOG_WALK_REPLAY_ALL) {
7363		wc.stage++;
7364		goto again;
7365	}
7366
7367	btrfs_free_path(path);
7368
7369	/* step 4: commit the transaction, which also unpins the blocks */
7370	ret = btrfs_commit_transaction(trans);
7371	if (ret)
7372		return ret;
7373
7374	log_root_tree->log_root = NULL;
7375	clear_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
7376	btrfs_put_root(log_root_tree);
7377
7378	return 0;
7379error:
7380	if (wc.trans)
7381		btrfs_end_transaction(wc.trans);
7382	clear_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
7383	btrfs_free_path(path);
7384	return ret;
7385}
7386
7387/*
7388 * there are some corner cases where we want to force a full
7389 * commit instead of allowing a directory to be logged.
7390 *
7391 * They revolve around files there were unlinked from the directory, and
7392 * this function updates the parent directory so that a full commit is
7393 * properly done if it is fsync'd later after the unlinks are done.
7394 *
7395 * Must be called before the unlink operations (updates to the subvolume tree,
7396 * inodes, etc) are done.
7397 */
7398void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
7399			     struct btrfs_inode *dir, struct btrfs_inode *inode,
7400			     bool for_rename)
7401{
7402	/*
7403	 * when we're logging a file, if it hasn't been renamed
7404	 * or unlinked, and its inode is fully committed on disk,
7405	 * we don't have to worry about walking up the directory chain
7406	 * to log its parents.
7407	 *
7408	 * So, we use the last_unlink_trans field to put this transid
7409	 * into the file.  When the file is logged we check it and
7410	 * don't log the parents if the file is fully on disk.
7411	 */
7412	mutex_lock(&inode->log_mutex);
7413	inode->last_unlink_trans = trans->transid;
7414	mutex_unlock(&inode->log_mutex);
7415
7416	if (!for_rename)
7417		return;
7418
7419	/*
7420	 * If this directory was already logged, any new names will be logged
7421	 * with btrfs_log_new_name() and old names will be deleted from the log
7422	 * tree with btrfs_del_dir_entries_in_log() or with
7423	 * btrfs_del_inode_ref_in_log().
7424	 */
7425	if (inode_logged(trans, dir, NULL) == 1)
7426		return;
7427
7428	/*
7429	 * If the inode we're about to unlink was logged before, the log will be
7430	 * properly updated with the new name with btrfs_log_new_name() and the
7431	 * old name removed with btrfs_del_dir_entries_in_log() or with
7432	 * btrfs_del_inode_ref_in_log().
7433	 */
7434	if (inode_logged(trans, inode, NULL) == 1)
7435		return;
7436
7437	/*
7438	 * when renaming files across directories, if the directory
7439	 * there we're unlinking from gets fsync'd later on, there's
7440	 * no way to find the destination directory later and fsync it
7441	 * properly.  So, we have to be conservative and force commits
7442	 * so the new name gets discovered.
7443	 */
7444	mutex_lock(&dir->log_mutex);
7445	dir->last_unlink_trans = trans->transid;
7446	mutex_unlock(&dir->log_mutex);
7447}
7448
7449/*
7450 * Make sure that if someone attempts to fsync the parent directory of a deleted
7451 * snapshot, it ends up triggering a transaction commit. This is to guarantee
7452 * that after replaying the log tree of the parent directory's root we will not
7453 * see the snapshot anymore and at log replay time we will not see any log tree
7454 * corresponding to the deleted snapshot's root, which could lead to replaying
7455 * it after replaying the log tree of the parent directory (which would replay
7456 * the snapshot delete operation).
7457 *
7458 * Must be called before the actual snapshot destroy operation (updates to the
7459 * parent root and tree of tree roots trees, etc) are done.
7460 */
7461void btrfs_record_snapshot_destroy(struct btrfs_trans_handle *trans,
7462				   struct btrfs_inode *dir)
7463{
7464	mutex_lock(&dir->log_mutex);
7465	dir->last_unlink_trans = trans->transid;
7466	mutex_unlock(&dir->log_mutex);
7467}
7468
7469/*
7470 * Call this when creating a subvolume in a directory.
7471 * Because we don't commit a transaction when creating a subvolume, we can't
7472 * allow the directory pointing to the subvolume to be logged with an entry that
7473 * points to an unpersisted root if we are still in the transaction used to
7474 * create the subvolume, so make any attempt to log the directory to result in a
7475 * full log sync.
7476 * Also we don't need to worry with renames, since btrfs_rename() marks the log
7477 * for full commit when renaming a subvolume.
7478 */
7479void btrfs_record_new_subvolume(const struct btrfs_trans_handle *trans,
7480				struct btrfs_inode *dir)
7481{
7482	mutex_lock(&dir->log_mutex);
7483	dir->last_unlink_trans = trans->transid;
7484	mutex_unlock(&dir->log_mutex);
7485}
7486
7487/*
7488 * Update the log after adding a new name for an inode.
7489 *
7490 * @trans:              Transaction handle.
7491 * @old_dentry:         The dentry associated with the old name and the old
7492 *                      parent directory.
7493 * @old_dir:            The inode of the previous parent directory for the case
7494 *                      of a rename. For a link operation, it must be NULL.
7495 * @old_dir_index:      The index number associated with the old name, meaningful
7496 *                      only for rename operations (when @old_dir is not NULL).
7497 *                      Ignored for link operations.
7498 * @parent:             The dentry associated with the directory under which the
7499 *                      new name is located.
7500 *
7501 * Call this after adding a new name for an inode, as a result of a link or
7502 * rename operation, and it will properly update the log to reflect the new name.
7503 */
7504void btrfs_log_new_name(struct btrfs_trans_handle *trans,
7505			struct dentry *old_dentry, struct btrfs_inode *old_dir,
7506			u64 old_dir_index, struct dentry *parent)
7507{
7508	struct btrfs_inode *inode = BTRFS_I(d_inode(old_dentry));
7509	struct btrfs_root *root = inode->root;
7510	struct btrfs_log_ctx ctx;
7511	bool log_pinned = false;
7512	int ret;
7513
7514	/*
7515	 * this will force the logging code to walk the dentry chain
7516	 * up for the file
7517	 */
7518	if (!S_ISDIR(inode->vfs_inode.i_mode))
7519		inode->last_unlink_trans = trans->transid;
7520
7521	/*
7522	 * if this inode hasn't been logged and directory we're renaming it
7523	 * from hasn't been logged, we don't need to log it
7524	 */
7525	ret = inode_logged(trans, inode, NULL);
7526	if (ret < 0) {
7527		goto out;
7528	} else if (ret == 0) {
7529		if (!old_dir)
7530			return;
7531		/*
7532		 * If the inode was not logged and we are doing a rename (old_dir is not
7533		 * NULL), check if old_dir was logged - if it was not we can return and
7534		 * do nothing.
7535		 */
7536		ret = inode_logged(trans, old_dir, NULL);
7537		if (ret < 0)
7538			goto out;
7539		else if (ret == 0)
7540			return;
7541	}
7542	ret = 0;
7543
7544	/*
7545	 * If we are doing a rename (old_dir is not NULL) from a directory that
7546	 * was previously logged, make sure that on log replay we get the old
7547	 * dir entry deleted. This is needed because we will also log the new
7548	 * name of the renamed inode, so we need to make sure that after log
7549	 * replay we don't end up with both the new and old dir entries existing.
7550	 */
7551	if (old_dir && old_dir->logged_trans == trans->transid) {
7552		struct btrfs_root *log = old_dir->root->log_root;
7553		struct btrfs_path *path;
7554		struct fscrypt_name fname;
7555
7556		ASSERT(old_dir_index >= BTRFS_DIR_START_INDEX);
7557
7558		ret = fscrypt_setup_filename(&old_dir->vfs_inode,
7559					     &old_dentry->d_name, 0, &fname);
7560		if (ret)
7561			goto out;
7562		/*
7563		 * We have two inodes to update in the log, the old directory and
7564		 * the inode that got renamed, so we must pin the log to prevent
7565		 * anyone from syncing the log until we have updated both inodes
7566		 * in the log.
7567		 */
7568		ret = join_running_log_trans(root);
7569		/*
7570		 * At least one of the inodes was logged before, so this should
7571		 * not fail, but if it does, it's not serious, just bail out and
7572		 * mark the log for a full commit.
7573		 */
7574		if (WARN_ON_ONCE(ret < 0)) {
7575			fscrypt_free_filename(&fname);
7576			goto out;
7577		}
7578
7579		log_pinned = true;
7580
7581		path = btrfs_alloc_path();
7582		if (!path) {
7583			ret = -ENOMEM;
7584			fscrypt_free_filename(&fname);
7585			goto out;
7586		}
7587
7588		/*
7589		 * Other concurrent task might be logging the old directory,
7590		 * as it can be triggered when logging other inode that had or
7591		 * still has a dentry in the old directory. We lock the old
7592		 * directory's log_mutex to ensure the deletion of the old
7593		 * name is persisted, because during directory logging we
7594		 * delete all BTRFS_DIR_LOG_INDEX_KEY keys and the deletion of
7595		 * the old name's dir index item is in the delayed items, so
7596		 * it could be missed by an in progress directory logging.
7597		 */
7598		mutex_lock(&old_dir->log_mutex);
7599		ret = del_logged_dentry(trans, log, path, btrfs_ino(old_dir),
7600					&fname.disk_name, old_dir_index);
7601		if (ret > 0) {
7602			/*
7603			 * The dentry does not exist in the log, so record its
7604			 * deletion.
7605			 */
7606			btrfs_release_path(path);
7607			ret = insert_dir_log_key(trans, log, path,
7608						 btrfs_ino(old_dir),
7609						 old_dir_index, old_dir_index);
7610		}
7611		mutex_unlock(&old_dir->log_mutex);
7612
7613		btrfs_free_path(path);
7614		fscrypt_free_filename(&fname);
7615		if (ret < 0)
7616			goto out;
7617	}
7618
7619	btrfs_init_log_ctx(&ctx, inode);
7620	ctx.logging_new_name = true;
7621	btrfs_init_log_ctx_scratch_eb(&ctx);
7622	/*
7623	 * We don't care about the return value. If we fail to log the new name
7624	 * then we know the next attempt to sync the log will fallback to a full
7625	 * transaction commit (due to a call to btrfs_set_log_full_commit()), so
7626	 * we don't need to worry about getting a log committed that has an
7627	 * inconsistent state after a rename operation.
7628	 */
7629	btrfs_log_inode_parent(trans, inode, parent, LOG_INODE_EXISTS, &ctx);
7630	free_extent_buffer(ctx.scratch_eb);
7631	ASSERT(list_empty(&ctx.conflict_inodes));
7632out:
7633	/*
7634	 * If an error happened mark the log for a full commit because it's not
7635	 * consistent and up to date or we couldn't find out if one of the
7636	 * inodes was logged before in this transaction. Do it before unpinning
7637	 * the log, to avoid any races with someone else trying to commit it.
7638	 */
7639	if (ret < 0)
7640		btrfs_set_log_full_commit(trans);
7641	if (log_pinned)
7642		btrfs_end_log_trans(root);
7643}
7644