1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2012 Alexander Block.  All rights reserved.
   4 */
   5
   6#include <linux/bsearch.h>
   7#include <linux/fs.h>
   8#include <linux/file.h>
   9#include <linux/sort.h>
  10#include <linux/mount.h>
  11#include <linux/xattr.h>
  12#include <linux/posix_acl_xattr.h>
  13#include <linux/radix-tree.h>
  14#include <linux/vmalloc.h>
  15#include <linux/string.h>
  16#include <linux/compat.h>
  17#include <linux/crc32c.h>
  18#include <linux/fsverity.h>
  19
  20#include "send.h"
  21#include "ctree.h"
  22#include "backref.h"
  23#include "locking.h"
  24#include "disk-io.h"
  25#include "btrfs_inode.h"
  26#include "transaction.h"
  27#include "compression.h"
 
  28#include "print-tree.h"
  29#include "accessors.h"
  30#include "dir-item.h"
  31#include "file-item.h"
  32#include "ioctl.h"
  33#include "verity.h"
  34#include "lru_cache.h"
  35
  36/*
  37 * Maximum number of references an extent can have in order for us to attempt to
  38 * issue clone operations instead of write operations. This currently exists to
  39 * avoid hitting limitations of the backreference walking code (taking a lot of
  40 * time and using too much memory for extents with large number of references).
  41 */
  42#define SEND_MAX_EXTENT_REFS	1024
  43
  44/*
  45 * A fs_path is a helper to dynamically build path names with unknown size.
  46 * It reallocates the internal buffer on demand.
  47 * It allows fast adding of path elements on the right side (normal path) and
  48 * fast adding to the left side (reversed path). A reversed path can also be
  49 * unreversed if needed.
  50 */
  51struct fs_path {
  52	union {
  53		struct {
  54			char *start;
  55			char *end;
  56
  57			char *buf;
  58			unsigned short buf_len:15;
  59			unsigned short reversed:1;
  60			char inline_buf[];
  61		};
  62		/*
  63		 * Average path length does not exceed 200 bytes, we'll have
  64		 * better packing in the slab and higher chance to satisfy
  65		 * a allocation later during send.
  66		 */
  67		char pad[256];
  68	};
  69};
  70#define FS_PATH_INLINE_SIZE \
  71	(sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
  72
  73
  74/* reused for each extent */
  75struct clone_root {
  76	struct btrfs_root *root;
  77	u64 ino;
  78	u64 offset;
  79	u64 num_bytes;
  80	bool found_ref;
  81};
  82
  83#define SEND_MAX_NAME_CACHE_SIZE			256
  84
  85/*
  86 * Limit the root_ids array of struct backref_cache_entry to 17 elements.
  87 * This makes the size of a cache entry to be exactly 192 bytes on x86_64, which
  88 * can be satisfied from the kmalloc-192 slab, without wasting any space.
  89 * The most common case is to have a single root for cloning, which corresponds
  90 * to the send root. Having the user specify more than 16 clone roots is not
  91 * common, and in such rare cases we simply don't use caching if the number of
  92 * cloning roots that lead down to a leaf is more than 17.
  93 */
  94#define SEND_MAX_BACKREF_CACHE_ROOTS			17
  95
  96/*
  97 * Max number of entries in the cache.
  98 * With SEND_MAX_BACKREF_CACHE_ROOTS as 17, the size in bytes, excluding
  99 * maple tree's internal nodes, is 24K.
 100 */
 101#define SEND_MAX_BACKREF_CACHE_SIZE 128
 102
 103/*
 104 * A backref cache entry maps a leaf to a list of IDs of roots from which the
 105 * leaf is accessible and we can use for clone operations.
 106 * With SEND_MAX_BACKREF_CACHE_ROOTS as 12, each cache entry is 128 bytes (on
 107 * x86_64).
 108 */
 109struct backref_cache_entry {
 110	struct btrfs_lru_cache_entry entry;
 111	u64 root_ids[SEND_MAX_BACKREF_CACHE_ROOTS];
 112	/* Number of valid elements in the root_ids array. */
 113	int num_roots;
 114};
 115
 116/* See the comment at lru_cache.h about struct btrfs_lru_cache_entry. */
 117static_assert(offsetof(struct backref_cache_entry, entry) == 0);
 118
 119/*
 120 * Max number of entries in the cache that stores directories that were already
 121 * created. The cache uses raw struct btrfs_lru_cache_entry entries, so it uses
 122 * at most 4096 bytes - sizeof(struct btrfs_lru_cache_entry) is 48 bytes, but
 123 * the kmalloc-64 slab is used, so we get 4096 bytes (64 bytes * 64).
 124 */
 125#define SEND_MAX_DIR_CREATED_CACHE_SIZE			64
 126
 127/*
 128 * Max number of entries in the cache that stores directories that were already
 129 * created. The cache uses raw struct btrfs_lru_cache_entry entries, so it uses
 130 * at most 4096 bytes - sizeof(struct btrfs_lru_cache_entry) is 48 bytes, but
 131 * the kmalloc-64 slab is used, so we get 4096 bytes (64 bytes * 64).
 132 */
 133#define SEND_MAX_DIR_UTIMES_CACHE_SIZE			64
 134
 135struct send_ctx {
 136	struct file *send_filp;
 137	loff_t send_off;
 138	char *send_buf;
 139	u32 send_size;
 140	u32 send_max_size;
 141	/*
 142	 * Whether BTRFS_SEND_A_DATA attribute was already added to current
 143	 * command (since protocol v2, data must be the last attribute).
 144	 */
 145	bool put_data;
 146	struct page **send_buf_pages;
 147	u64 flags;	/* 'flags' member of btrfs_ioctl_send_args is u64 */
 148	/* Protocol version compatibility requested */
 149	u32 proto;
 150
 151	struct btrfs_root *send_root;
 152	struct btrfs_root *parent_root;
 153	struct clone_root *clone_roots;
 154	int clone_roots_cnt;
 155
 156	/* current state of the compare_tree call */
 157	struct btrfs_path *left_path;
 158	struct btrfs_path *right_path;
 159	struct btrfs_key *cmp_key;
 160
 161	/*
 162	 * Keep track of the generation of the last transaction that was used
 163	 * for relocating a block group. This is periodically checked in order
 164	 * to detect if a relocation happened since the last check, so that we
 165	 * don't operate on stale extent buffers for nodes (level >= 1) or on
 166	 * stale disk_bytenr values of file extent items.
 167	 */
 168	u64 last_reloc_trans;
 169
 170	/*
 171	 * infos of the currently processed inode. In case of deleted inodes,
 172	 * these are the values from the deleted inode.
 173	 */
 174	u64 cur_ino;
 175	u64 cur_inode_gen;
 176	u64 cur_inode_size;
 177	u64 cur_inode_mode;
 178	u64 cur_inode_rdev;
 179	u64 cur_inode_last_extent;
 180	u64 cur_inode_next_write_offset;
 181	bool cur_inode_new;
 182	bool cur_inode_new_gen;
 183	bool cur_inode_deleted;
 184	bool ignore_cur_inode;
 185	bool cur_inode_needs_verity;
 186	void *verity_descriptor;
 187
 188	u64 send_progress;
 189
 190	struct list_head new_refs;
 191	struct list_head deleted_refs;
 192
 193	struct btrfs_lru_cache name_cache;
 194
 195	/*
 196	 * The inode we are currently processing. It's not NULL only when we
 197	 * need to issue write commands for data extents from this inode.
 198	 */
 199	struct inode *cur_inode;
 200	struct file_ra_state ra;
 201	u64 page_cache_clear_start;
 202	bool clean_page_cache;
 203
 204	/*
 205	 * We process inodes by their increasing order, so if before an
 206	 * incremental send we reverse the parent/child relationship of
 207	 * directories such that a directory with a lower inode number was
 208	 * the parent of a directory with a higher inode number, and the one
 209	 * becoming the new parent got renamed too, we can't rename/move the
 210	 * directory with lower inode number when we finish processing it - we
 211	 * must process the directory with higher inode number first, then
 212	 * rename/move it and then rename/move the directory with lower inode
 213	 * number. Example follows.
 214	 *
 215	 * Tree state when the first send was performed:
 216	 *
 217	 * .
 218	 * |-- a                   (ino 257)
 219	 *     |-- b               (ino 258)
 220	 *         |
 221	 *         |
 222	 *         |-- c           (ino 259)
 223	 *         |   |-- d       (ino 260)
 224	 *         |
 225	 *         |-- c2          (ino 261)
 226	 *
 227	 * Tree state when the second (incremental) send is performed:
 228	 *
 229	 * .
 230	 * |-- a                   (ino 257)
 231	 *     |-- b               (ino 258)
 232	 *         |-- c2          (ino 261)
 233	 *             |-- d2      (ino 260)
 234	 *                 |-- cc  (ino 259)
 235	 *
 236	 * The sequence of steps that lead to the second state was:
 237	 *
 238	 * mv /a/b/c/d /a/b/c2/d2
 239	 * mv /a/b/c /a/b/c2/d2/cc
 240	 *
 241	 * "c" has lower inode number, but we can't move it (2nd mv operation)
 242	 * before we move "d", which has higher inode number.
 243	 *
 244	 * So we just memorize which move/rename operations must be performed
 245	 * later when their respective parent is processed and moved/renamed.
 246	 */
 247
 248	/* Indexed by parent directory inode number. */
 249	struct rb_root pending_dir_moves;
 250
 251	/*
 252	 * Reverse index, indexed by the inode number of a directory that
 253	 * is waiting for the move/rename of its immediate parent before its
 254	 * own move/rename can be performed.
 255	 */
 256	struct rb_root waiting_dir_moves;
 257
 258	/*
 259	 * A directory that is going to be rm'ed might have a child directory
 260	 * which is in the pending directory moves index above. In this case,
 261	 * the directory can only be removed after the move/rename of its child
 262	 * is performed. Example:
 263	 *
 264	 * Parent snapshot:
 265	 *
 266	 * .                        (ino 256)
 267	 * |-- a/                   (ino 257)
 268	 *     |-- b/               (ino 258)
 269	 *         |-- c/           (ino 259)
 270	 *         |   |-- x/       (ino 260)
 271	 *         |
 272	 *         |-- y/           (ino 261)
 273	 *
 274	 * Send snapshot:
 275	 *
 276	 * .                        (ino 256)
 277	 * |-- a/                   (ino 257)
 278	 *     |-- b/               (ino 258)
 279	 *         |-- YY/          (ino 261)
 280	 *              |-- x/      (ino 260)
 281	 *
 282	 * Sequence of steps that lead to the send snapshot:
 283	 * rm -f /a/b/c/foo.txt
 284	 * mv /a/b/y /a/b/YY
 285	 * mv /a/b/c/x /a/b/YY
 286	 * rmdir /a/b/c
 287	 *
 288	 * When the child is processed, its move/rename is delayed until its
 289	 * parent is processed (as explained above), but all other operations
 290	 * like update utimes, chown, chgrp, etc, are performed and the paths
 291	 * that it uses for those operations must use the orphanized name of
 292	 * its parent (the directory we're going to rm later), so we need to
 293	 * memorize that name.
 294	 *
 295	 * Indexed by the inode number of the directory to be deleted.
 296	 */
 297	struct rb_root orphan_dirs;
 298
 299	struct rb_root rbtree_new_refs;
 300	struct rb_root rbtree_deleted_refs;
 301
 302	struct btrfs_lru_cache backref_cache;
 303	u64 backref_cache_last_reloc_trans;
 304
 305	struct btrfs_lru_cache dir_created_cache;
 306	struct btrfs_lru_cache dir_utimes_cache;
 307};
 308
 309struct pending_dir_move {
 310	struct rb_node node;
 311	struct list_head list;
 312	u64 parent_ino;
 313	u64 ino;
 314	u64 gen;
 315	struct list_head update_refs;
 316};
 317
 318struct waiting_dir_move {
 319	struct rb_node node;
 320	u64 ino;
 321	/*
 322	 * There might be some directory that could not be removed because it
 323	 * was waiting for this directory inode to be moved first. Therefore
 324	 * after this directory is moved, we can try to rmdir the ino rmdir_ino.
 325	 */
 326	u64 rmdir_ino;
 327	u64 rmdir_gen;
 328	bool orphanized;
 329};
 330
 331struct orphan_dir_info {
 332	struct rb_node node;
 333	u64 ino;
 334	u64 gen;
 335	u64 last_dir_index_offset;
 336	u64 dir_high_seq_ino;
 337};
 338
 339struct name_cache_entry {
 340	/*
 341	 * The key in the entry is an inode number, and the generation matches
 342	 * the inode's generation.
 343	 */
 344	struct btrfs_lru_cache_entry entry;
 345	u64 parent_ino;
 346	u64 parent_gen;
 347	int ret;
 348	int need_later_update;
 349	int name_len;
 350	char name[];
 351};
 352
 353/* See the comment at lru_cache.h about struct btrfs_lru_cache_entry. */
 354static_assert(offsetof(struct name_cache_entry, entry) == 0);
 355
 356#define ADVANCE							1
 357#define ADVANCE_ONLY_NEXT					-1
 358
 359enum btrfs_compare_tree_result {
 360	BTRFS_COMPARE_TREE_NEW,
 361	BTRFS_COMPARE_TREE_DELETED,
 362	BTRFS_COMPARE_TREE_CHANGED,
 363	BTRFS_COMPARE_TREE_SAME,
 364};
 365
 366__cold
 367static void inconsistent_snapshot_error(struct send_ctx *sctx,
 368					enum btrfs_compare_tree_result result,
 369					const char *what)
 370{
 371	const char *result_string;
 372
 373	switch (result) {
 374	case BTRFS_COMPARE_TREE_NEW:
 375		result_string = "new";
 376		break;
 377	case BTRFS_COMPARE_TREE_DELETED:
 378		result_string = "deleted";
 379		break;
 380	case BTRFS_COMPARE_TREE_CHANGED:
 381		result_string = "updated";
 382		break;
 383	case BTRFS_COMPARE_TREE_SAME:
 384		ASSERT(0);
 385		result_string = "unchanged";
 386		break;
 387	default:
 388		ASSERT(0);
 389		result_string = "unexpected";
 390	}
 391
 392	btrfs_err(sctx->send_root->fs_info,
 393		  "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
 394		  result_string, what, sctx->cmp_key->objectid,
 395		  sctx->send_root->root_key.objectid,
 396		  (sctx->parent_root ?
 397		   sctx->parent_root->root_key.objectid : 0));
 398}
 399
 400__maybe_unused
 401static bool proto_cmd_ok(const struct send_ctx *sctx, int cmd)
 402{
 403	switch (sctx->proto) {
 404	case 1:	 return cmd <= BTRFS_SEND_C_MAX_V1;
 405	case 2:	 return cmd <= BTRFS_SEND_C_MAX_V2;
 406	case 3:	 return cmd <= BTRFS_SEND_C_MAX_V3;
 407	default: return false;
 408	}
 409}
 410
 411static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
 412
 413static struct waiting_dir_move *
 414get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
 415
 416static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen);
 417
 418static int need_send_hole(struct send_ctx *sctx)
 419{
 420	return (sctx->parent_root && !sctx->cur_inode_new &&
 421		!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
 422		S_ISREG(sctx->cur_inode_mode));
 423}
 424
 425static void fs_path_reset(struct fs_path *p)
 426{
 427	if (p->reversed) {
 428		p->start = p->buf + p->buf_len - 1;
 429		p->end = p->start;
 430		*p->start = 0;
 431	} else {
 432		p->start = p->buf;
 433		p->end = p->start;
 434		*p->start = 0;
 435	}
 436}
 437
 438static struct fs_path *fs_path_alloc(void)
 439{
 440	struct fs_path *p;
 441
 442	p = kmalloc(sizeof(*p), GFP_KERNEL);
 443	if (!p)
 444		return NULL;
 445	p->reversed = 0;
 446	p->buf = p->inline_buf;
 447	p->buf_len = FS_PATH_INLINE_SIZE;
 448	fs_path_reset(p);
 449	return p;
 450}
 451
 452static struct fs_path *fs_path_alloc_reversed(void)
 453{
 454	struct fs_path *p;
 455
 456	p = fs_path_alloc();
 457	if (!p)
 458		return NULL;
 459	p->reversed = 1;
 460	fs_path_reset(p);
 461	return p;
 462}
 463
 464static void fs_path_free(struct fs_path *p)
 465{
 466	if (!p)
 467		return;
 468	if (p->buf != p->inline_buf)
 469		kfree(p->buf);
 470	kfree(p);
 471}
 472
 473static int fs_path_len(struct fs_path *p)
 474{
 475	return p->end - p->start;
 476}
 477
 478static int fs_path_ensure_buf(struct fs_path *p, int len)
 479{
 480	char *tmp_buf;
 481	int path_len;
 482	int old_buf_len;
 483
 484	len++;
 485
 486	if (p->buf_len >= len)
 487		return 0;
 488
 489	if (len > PATH_MAX) {
 490		WARN_ON(1);
 491		return -ENOMEM;
 492	}
 493
 494	path_len = p->end - p->start;
 495	old_buf_len = p->buf_len;
 496
 497	/*
 498	 * Allocate to the next largest kmalloc bucket size, to let
 499	 * the fast path happen most of the time.
 500	 */
 501	len = kmalloc_size_roundup(len);
 502	/*
 503	 * First time the inline_buf does not suffice
 504	 */
 505	if (p->buf == p->inline_buf) {
 506		tmp_buf = kmalloc(len, GFP_KERNEL);
 507		if (tmp_buf)
 508			memcpy(tmp_buf, p->buf, old_buf_len);
 509	} else {
 510		tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
 511	}
 512	if (!tmp_buf)
 513		return -ENOMEM;
 514	p->buf = tmp_buf;
 515	p->buf_len = len;
 516
 517	if (p->reversed) {
 518		tmp_buf = p->buf + old_buf_len - path_len - 1;
 519		p->end = p->buf + p->buf_len - 1;
 520		p->start = p->end - path_len;
 521		memmove(p->start, tmp_buf, path_len + 1);
 522	} else {
 523		p->start = p->buf;
 524		p->end = p->start + path_len;
 525	}
 526	return 0;
 527}
 528
 529static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
 530				   char **prepared)
 531{
 532	int ret;
 533	int new_len;
 534
 535	new_len = p->end - p->start + name_len;
 536	if (p->start != p->end)
 537		new_len++;
 538	ret = fs_path_ensure_buf(p, new_len);
 539	if (ret < 0)
 540		goto out;
 541
 542	if (p->reversed) {
 543		if (p->start != p->end)
 544			*--p->start = '/';
 545		p->start -= name_len;
 546		*prepared = p->start;
 547	} else {
 548		if (p->start != p->end)
 549			*p->end++ = '/';
 550		*prepared = p->end;
 551		p->end += name_len;
 552		*p->end = 0;
 553	}
 554
 555out:
 556	return ret;
 557}
 558
 559static int fs_path_add(struct fs_path *p, const char *name, int name_len)
 560{
 561	int ret;
 562	char *prepared;
 563
 564	ret = fs_path_prepare_for_add(p, name_len, &prepared);
 565	if (ret < 0)
 566		goto out;
 567	memcpy(prepared, name, name_len);
 568
 569out:
 570	return ret;
 571}
 572
 573static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
 574{
 575	int ret;
 576	char *prepared;
 577
 578	ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
 579	if (ret < 0)
 580		goto out;
 581	memcpy(prepared, p2->start, p2->end - p2->start);
 582
 583out:
 584	return ret;
 585}
 586
 587static int fs_path_add_from_extent_buffer(struct fs_path *p,
 588					  struct extent_buffer *eb,
 589					  unsigned long off, int len)
 590{
 591	int ret;
 592	char *prepared;
 593
 594	ret = fs_path_prepare_for_add(p, len, &prepared);
 595	if (ret < 0)
 596		goto out;
 597
 598	read_extent_buffer(eb, prepared, off, len);
 599
 600out:
 601	return ret;
 602}
 603
 604static int fs_path_copy(struct fs_path *p, struct fs_path *from)
 605{
 606	p->reversed = from->reversed;
 607	fs_path_reset(p);
 608
 609	return fs_path_add_path(p, from);
 610}
 611
 612static void fs_path_unreverse(struct fs_path *p)
 613{
 614	char *tmp;
 615	int len;
 616
 617	if (!p->reversed)
 618		return;
 619
 620	tmp = p->start;
 621	len = p->end - p->start;
 622	p->start = p->buf;
 623	p->end = p->start + len;
 624	memmove(p->start, tmp, len + 1);
 625	p->reversed = 0;
 626}
 627
 628static struct btrfs_path *alloc_path_for_send(void)
 629{
 630	struct btrfs_path *path;
 631
 632	path = btrfs_alloc_path();
 633	if (!path)
 634		return NULL;
 635	path->search_commit_root = 1;
 636	path->skip_locking = 1;
 637	path->need_commit_sem = 1;
 638	return path;
 639}
 640
 641static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
 642{
 643	int ret;
 644	u32 pos = 0;
 645
 646	while (pos < len) {
 647		ret = kernel_write(filp, buf + pos, len - pos, off);
 648		if (ret < 0)
 649			return ret;
 650		if (ret == 0)
 651			return -EIO;
 652		pos += ret;
 653	}
 654
 655	return 0;
 656}
 657
 658static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
 659{
 660	struct btrfs_tlv_header *hdr;
 661	int total_len = sizeof(*hdr) + len;
 662	int left = sctx->send_max_size - sctx->send_size;
 663
 664	if (WARN_ON_ONCE(sctx->put_data))
 665		return -EINVAL;
 666
 667	if (unlikely(left < total_len))
 668		return -EOVERFLOW;
 669
 670	hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
 671	put_unaligned_le16(attr, &hdr->tlv_type);
 672	put_unaligned_le16(len, &hdr->tlv_len);
 673	memcpy(hdr + 1, data, len);
 674	sctx->send_size += total_len;
 675
 676	return 0;
 677}
 678
 679#define TLV_PUT_DEFINE_INT(bits) \
 680	static int tlv_put_u##bits(struct send_ctx *sctx,	 	\
 681			u##bits attr, u##bits value)			\
 682	{								\
 683		__le##bits __tmp = cpu_to_le##bits(value);		\
 684		return tlv_put(sctx, attr, &__tmp, sizeof(__tmp));	\
 685	}
 686
 687TLV_PUT_DEFINE_INT(8)
 688TLV_PUT_DEFINE_INT(32)
 689TLV_PUT_DEFINE_INT(64)
 690
 691static int tlv_put_string(struct send_ctx *sctx, u16 attr,
 692			  const char *str, int len)
 693{
 694	if (len == -1)
 695		len = strlen(str);
 696	return tlv_put(sctx, attr, str, len);
 697}
 698
 699static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
 700			const u8 *uuid)
 701{
 702	return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
 703}
 704
 705static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
 706				  struct extent_buffer *eb,
 707				  struct btrfs_timespec *ts)
 708{
 709	struct btrfs_timespec bts;
 710	read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
 711	return tlv_put(sctx, attr, &bts, sizeof(bts));
 712}
 713
 714
 715#define TLV_PUT(sctx, attrtype, data, attrlen) \
 716	do { \
 717		ret = tlv_put(sctx, attrtype, data, attrlen); \
 718		if (ret < 0) \
 719			goto tlv_put_failure; \
 720	} while (0)
 721
 722#define TLV_PUT_INT(sctx, attrtype, bits, value) \
 723	do { \
 724		ret = tlv_put_u##bits(sctx, attrtype, value); \
 725		if (ret < 0) \
 726			goto tlv_put_failure; \
 727	} while (0)
 728
 729#define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
 730#define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
 731#define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
 732#define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
 733#define TLV_PUT_STRING(sctx, attrtype, str, len) \
 734	do { \
 735		ret = tlv_put_string(sctx, attrtype, str, len); \
 736		if (ret < 0) \
 737			goto tlv_put_failure; \
 738	} while (0)
 739#define TLV_PUT_PATH(sctx, attrtype, p) \
 740	do { \
 741		ret = tlv_put_string(sctx, attrtype, p->start, \
 742			p->end - p->start); \
 743		if (ret < 0) \
 744			goto tlv_put_failure; \
 745	} while(0)
 746#define TLV_PUT_UUID(sctx, attrtype, uuid) \
 747	do { \
 748		ret = tlv_put_uuid(sctx, attrtype, uuid); \
 749		if (ret < 0) \
 750			goto tlv_put_failure; \
 751	} while (0)
 752#define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
 753	do { \
 754		ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
 755		if (ret < 0) \
 756			goto tlv_put_failure; \
 757	} while (0)
 758
 759static int send_header(struct send_ctx *sctx)
 760{
 761	struct btrfs_stream_header hdr;
 762
 763	strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
 764	hdr.version = cpu_to_le32(sctx->proto);
 765	return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
 766					&sctx->send_off);
 767}
 768
 769/*
 770 * For each command/item we want to send to userspace, we call this function.
 771 */
 772static int begin_cmd(struct send_ctx *sctx, int cmd)
 773{
 774	struct btrfs_cmd_header *hdr;
 775
 776	if (WARN_ON(!sctx->send_buf))
 777		return -EINVAL;
 778
 779	if (unlikely(sctx->send_size != 0)) {
 780		btrfs_err(sctx->send_root->fs_info,
 781			  "send: command header buffer not empty cmd %d offset %llu",
 782			  cmd, sctx->send_off);
 783		return -EINVAL;
 784	}
 785
 786	sctx->send_size += sizeof(*hdr);
 787	hdr = (struct btrfs_cmd_header *)sctx->send_buf;
 788	put_unaligned_le16(cmd, &hdr->cmd);
 789
 790	return 0;
 791}
 792
 793static int send_cmd(struct send_ctx *sctx)
 794{
 795	int ret;
 796	struct btrfs_cmd_header *hdr;
 797	u32 crc;
 798
 799	hdr = (struct btrfs_cmd_header *)sctx->send_buf;
 800	put_unaligned_le32(sctx->send_size - sizeof(*hdr), &hdr->len);
 801	put_unaligned_le32(0, &hdr->crc);
 802
 803	crc = crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
 804	put_unaligned_le32(crc, &hdr->crc);
 805
 806	ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
 807					&sctx->send_off);
 808
 809	sctx->send_size = 0;
 810	sctx->put_data = false;
 811
 812	return ret;
 813}
 814
 815/*
 816 * Sends a move instruction to user space
 817 */
 818static int send_rename(struct send_ctx *sctx,
 819		     struct fs_path *from, struct fs_path *to)
 820{
 821	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
 822	int ret;
 823
 824	btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
 825
 826	ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
 827	if (ret < 0)
 828		goto out;
 829
 830	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
 831	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
 832
 833	ret = send_cmd(sctx);
 834
 835tlv_put_failure:
 836out:
 837	return ret;
 838}
 839
 840/*
 841 * Sends a link instruction to user space
 842 */
 843static int send_link(struct send_ctx *sctx,
 844		     struct fs_path *path, struct fs_path *lnk)
 845{
 846	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
 847	int ret;
 848
 849	btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
 850
 851	ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
 852	if (ret < 0)
 853		goto out;
 854
 855	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
 856	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
 857
 858	ret = send_cmd(sctx);
 859
 860tlv_put_failure:
 861out:
 862	return ret;
 863}
 864
 865/*
 866 * Sends an unlink instruction to user space
 867 */
 868static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
 869{
 870	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
 871	int ret;
 872
 873	btrfs_debug(fs_info, "send_unlink %s", path->start);
 874
 875	ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
 876	if (ret < 0)
 877		goto out;
 878
 879	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
 880
 881	ret = send_cmd(sctx);
 882
 883tlv_put_failure:
 884out:
 885	return ret;
 886}
 887
 888/*
 889 * Sends a rmdir instruction to user space
 890 */
 891static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
 892{
 893	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
 894	int ret;
 895
 896	btrfs_debug(fs_info, "send_rmdir %s", path->start);
 897
 898	ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
 899	if (ret < 0)
 900		goto out;
 901
 902	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
 903
 904	ret = send_cmd(sctx);
 905
 906tlv_put_failure:
 907out:
 908	return ret;
 909}
 910
 911struct btrfs_inode_info {
 912	u64 size;
 913	u64 gen;
 914	u64 mode;
 915	u64 uid;
 916	u64 gid;
 917	u64 rdev;
 918	u64 fileattr;
 919	u64 nlink;
 920};
 921
 922/*
 923 * Helper function to retrieve some fields from an inode item.
 924 */
 925static int get_inode_info(struct btrfs_root *root, u64 ino,
 926			  struct btrfs_inode_info *info)
 927{
 928	int ret;
 929	struct btrfs_path *path;
 930	struct btrfs_inode_item *ii;
 931	struct btrfs_key key;
 932
 933	path = alloc_path_for_send();
 934	if (!path)
 935		return -ENOMEM;
 936
 937	key.objectid = ino;
 938	key.type = BTRFS_INODE_ITEM_KEY;
 939	key.offset = 0;
 940	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 941	if (ret) {
 942		if (ret > 0)
 943			ret = -ENOENT;
 944		goto out;
 945	}
 946
 947	if (!info)
 948		goto out;
 949
 950	ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
 951			struct btrfs_inode_item);
 952	info->size = btrfs_inode_size(path->nodes[0], ii);
 953	info->gen = btrfs_inode_generation(path->nodes[0], ii);
 954	info->mode = btrfs_inode_mode(path->nodes[0], ii);
 955	info->uid = btrfs_inode_uid(path->nodes[0], ii);
 956	info->gid = btrfs_inode_gid(path->nodes[0], ii);
 957	info->rdev = btrfs_inode_rdev(path->nodes[0], ii);
 958	info->nlink = btrfs_inode_nlink(path->nodes[0], ii);
 959	/*
 960	 * Transfer the unchanged u64 value of btrfs_inode_item::flags, that's
 961	 * otherwise logically split to 32/32 parts.
 962	 */
 963	info->fileattr = btrfs_inode_flags(path->nodes[0], ii);
 964
 965out:
 966	btrfs_free_path(path);
 967	return ret;
 968}
 969
 970static int get_inode_gen(struct btrfs_root *root, u64 ino, u64 *gen)
 971{
 972	int ret;
 973	struct btrfs_inode_info info = { 0 };
 974
 975	ASSERT(gen);
 976
 977	ret = get_inode_info(root, ino, &info);
 978	*gen = info.gen;
 979	return ret;
 980}
 981
 982typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
 983				   struct fs_path *p,
 984				   void *ctx);
 985
 986/*
 987 * Helper function to iterate the entries in ONE btrfs_inode_ref or
 988 * btrfs_inode_extref.
 989 * The iterate callback may return a non zero value to stop iteration. This can
 990 * be a negative value for error codes or 1 to simply stop it.
 991 *
 992 * path must point to the INODE_REF or INODE_EXTREF when called.
 993 */
 994static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
 995			     struct btrfs_key *found_key, int resolve,
 996			     iterate_inode_ref_t iterate, void *ctx)
 997{
 998	struct extent_buffer *eb = path->nodes[0];
 999	struct btrfs_inode_ref *iref;
1000	struct btrfs_inode_extref *extref;
1001	struct btrfs_path *tmp_path;
1002	struct fs_path *p;
1003	u32 cur = 0;
1004	u32 total;
1005	int slot = path->slots[0];
1006	u32 name_len;
1007	char *start;
1008	int ret = 0;
1009	int num = 0;
1010	int index;
1011	u64 dir;
1012	unsigned long name_off;
1013	unsigned long elem_size;
1014	unsigned long ptr;
1015
1016	p = fs_path_alloc_reversed();
1017	if (!p)
1018		return -ENOMEM;
1019
1020	tmp_path = alloc_path_for_send();
1021	if (!tmp_path) {
1022		fs_path_free(p);
1023		return -ENOMEM;
1024	}
1025
1026
1027	if (found_key->type == BTRFS_INODE_REF_KEY) {
1028		ptr = (unsigned long)btrfs_item_ptr(eb, slot,
1029						    struct btrfs_inode_ref);
1030		total = btrfs_item_size(eb, slot);
1031		elem_size = sizeof(*iref);
1032	} else {
1033		ptr = btrfs_item_ptr_offset(eb, slot);
1034		total = btrfs_item_size(eb, slot);
1035		elem_size = sizeof(*extref);
1036	}
1037
1038	while (cur < total) {
1039		fs_path_reset(p);
1040
1041		if (found_key->type == BTRFS_INODE_REF_KEY) {
1042			iref = (struct btrfs_inode_ref *)(ptr + cur);
1043			name_len = btrfs_inode_ref_name_len(eb, iref);
1044			name_off = (unsigned long)(iref + 1);
1045			index = btrfs_inode_ref_index(eb, iref);
1046			dir = found_key->offset;
1047		} else {
1048			extref = (struct btrfs_inode_extref *)(ptr + cur);
1049			name_len = btrfs_inode_extref_name_len(eb, extref);
1050			name_off = (unsigned long)&extref->name;
1051			index = btrfs_inode_extref_index(eb, extref);
1052			dir = btrfs_inode_extref_parent(eb, extref);
1053		}
1054
1055		if (resolve) {
1056			start = btrfs_ref_to_path(root, tmp_path, name_len,
1057						  name_off, eb, dir,
1058						  p->buf, p->buf_len);
1059			if (IS_ERR(start)) {
1060				ret = PTR_ERR(start);
1061				goto out;
1062			}
1063			if (start < p->buf) {
1064				/* overflow , try again with larger buffer */
1065				ret = fs_path_ensure_buf(p,
1066						p->buf_len + p->buf - start);
1067				if (ret < 0)
1068					goto out;
1069				start = btrfs_ref_to_path(root, tmp_path,
1070							  name_len, name_off,
1071							  eb, dir,
1072							  p->buf, p->buf_len);
1073				if (IS_ERR(start)) {
1074					ret = PTR_ERR(start);
1075					goto out;
1076				}
1077				if (unlikely(start < p->buf)) {
1078					btrfs_err(root->fs_info,
1079			"send: path ref buffer underflow for key (%llu %u %llu)",
1080						  found_key->objectid,
1081						  found_key->type,
1082						  found_key->offset);
1083					ret = -EINVAL;
1084					goto out;
1085				}
1086			}
1087			p->start = start;
1088		} else {
1089			ret = fs_path_add_from_extent_buffer(p, eb, name_off,
1090							     name_len);
1091			if (ret < 0)
1092				goto out;
1093		}
1094
1095		cur += elem_size + name_len;
1096		ret = iterate(num, dir, index, p, ctx);
1097		if (ret)
1098			goto out;
1099		num++;
1100	}
1101
1102out:
1103	btrfs_free_path(tmp_path);
1104	fs_path_free(p);
1105	return ret;
1106}
1107
1108typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
1109				  const char *name, int name_len,
1110				  const char *data, int data_len,
1111				  void *ctx);
1112
1113/*
1114 * Helper function to iterate the entries in ONE btrfs_dir_item.
1115 * The iterate callback may return a non zero value to stop iteration. This can
1116 * be a negative value for error codes or 1 to simply stop it.
1117 *
1118 * path must point to the dir item when called.
1119 */
1120static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
1121			    iterate_dir_item_t iterate, void *ctx)
1122{
1123	int ret = 0;
1124	struct extent_buffer *eb;
1125	struct btrfs_dir_item *di;
1126	struct btrfs_key di_key;
1127	char *buf = NULL;
1128	int buf_len;
1129	u32 name_len;
1130	u32 data_len;
1131	u32 cur;
1132	u32 len;
1133	u32 total;
1134	int slot;
1135	int num;
1136
1137	/*
1138	 * Start with a small buffer (1 page). If later we end up needing more
1139	 * space, which can happen for xattrs on a fs with a leaf size greater
1140	 * then the page size, attempt to increase the buffer. Typically xattr
1141	 * values are small.
1142	 */
1143	buf_len = PATH_MAX;
1144	buf = kmalloc(buf_len, GFP_KERNEL);
1145	if (!buf) {
1146		ret = -ENOMEM;
1147		goto out;
1148	}
1149
1150	eb = path->nodes[0];
1151	slot = path->slots[0];
1152	di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1153	cur = 0;
1154	len = 0;
1155	total = btrfs_item_size(eb, slot);
1156
1157	num = 0;
1158	while (cur < total) {
1159		name_len = btrfs_dir_name_len(eb, di);
1160		data_len = btrfs_dir_data_len(eb, di);
1161		btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1162
1163		if (btrfs_dir_ftype(eb, di) == BTRFS_FT_XATTR) {
1164			if (name_len > XATTR_NAME_MAX) {
1165				ret = -ENAMETOOLONG;
1166				goto out;
1167			}
1168			if (name_len + data_len >
1169					BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1170				ret = -E2BIG;
1171				goto out;
1172			}
1173		} else {
1174			/*
1175			 * Path too long
1176			 */
1177			if (name_len + data_len > PATH_MAX) {
1178				ret = -ENAMETOOLONG;
1179				goto out;
1180			}
1181		}
1182
1183		if (name_len + data_len > buf_len) {
1184			buf_len = name_len + data_len;
1185			if (is_vmalloc_addr(buf)) {
1186				vfree(buf);
1187				buf = NULL;
1188			} else {
1189				char *tmp = krealloc(buf, buf_len,
1190						GFP_KERNEL | __GFP_NOWARN);
1191
1192				if (!tmp)
1193					kfree(buf);
1194				buf = tmp;
1195			}
1196			if (!buf) {
1197				buf = kvmalloc(buf_len, GFP_KERNEL);
1198				if (!buf) {
1199					ret = -ENOMEM;
1200					goto out;
1201				}
1202			}
1203		}
1204
1205		read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1206				name_len + data_len);
1207
1208		len = sizeof(*di) + name_len + data_len;
1209		di = (struct btrfs_dir_item *)((char *)di + len);
1210		cur += len;
1211
1212		ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1213			      data_len, ctx);
1214		if (ret < 0)
1215			goto out;
1216		if (ret) {
1217			ret = 0;
1218			goto out;
1219		}
1220
1221		num++;
1222	}
1223
1224out:
1225	kvfree(buf);
1226	return ret;
1227}
1228
1229static int __copy_first_ref(int num, u64 dir, int index,
1230			    struct fs_path *p, void *ctx)
1231{
1232	int ret;
1233	struct fs_path *pt = ctx;
1234
1235	ret = fs_path_copy(pt, p);
1236	if (ret < 0)
1237		return ret;
1238
1239	/* we want the first only */
1240	return 1;
1241}
1242
1243/*
1244 * Retrieve the first path of an inode. If an inode has more then one
1245 * ref/hardlink, this is ignored.
1246 */
1247static int get_inode_path(struct btrfs_root *root,
1248			  u64 ino, struct fs_path *path)
1249{
1250	int ret;
1251	struct btrfs_key key, found_key;
1252	struct btrfs_path *p;
1253
1254	p = alloc_path_for_send();
1255	if (!p)
1256		return -ENOMEM;
1257
1258	fs_path_reset(path);
1259
1260	key.objectid = ino;
1261	key.type = BTRFS_INODE_REF_KEY;
1262	key.offset = 0;
1263
1264	ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1265	if (ret < 0)
1266		goto out;
1267	if (ret) {
1268		ret = 1;
1269		goto out;
1270	}
1271	btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1272	if (found_key.objectid != ino ||
1273	    (found_key.type != BTRFS_INODE_REF_KEY &&
1274	     found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1275		ret = -ENOENT;
1276		goto out;
1277	}
1278
1279	ret = iterate_inode_ref(root, p, &found_key, 1,
1280				__copy_first_ref, path);
1281	if (ret < 0)
1282		goto out;
1283	ret = 0;
1284
1285out:
1286	btrfs_free_path(p);
1287	return ret;
1288}
1289
1290struct backref_ctx {
1291	struct send_ctx *sctx;
1292
1293	/* number of total found references */
1294	u64 found;
1295
1296	/*
1297	 * used for clones found in send_root. clones found behind cur_objectid
1298	 * and cur_offset are not considered as allowed clones.
1299	 */
1300	u64 cur_objectid;
1301	u64 cur_offset;
1302
1303	/* may be truncated in case it's the last extent in a file */
1304	u64 extent_len;
1305
1306	/* The bytenr the file extent item we are processing refers to. */
1307	u64 bytenr;
1308	/* The owner (root id) of the data backref for the current extent. */
1309	u64 backref_owner;
1310	/* The offset of the data backref for the current extent. */
1311	u64 backref_offset;
1312};
1313
1314static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1315{
1316	u64 root = (u64)(uintptr_t)key;
1317	const struct clone_root *cr = elt;
1318
1319	if (root < cr->root->root_key.objectid)
1320		return -1;
1321	if (root > cr->root->root_key.objectid)
1322		return 1;
1323	return 0;
1324}
1325
1326static int __clone_root_cmp_sort(const void *e1, const void *e2)
1327{
1328	const struct clone_root *cr1 = e1;
1329	const struct clone_root *cr2 = e2;
1330
1331	if (cr1->root->root_key.objectid < cr2->root->root_key.objectid)
1332		return -1;
1333	if (cr1->root->root_key.objectid > cr2->root->root_key.objectid)
1334		return 1;
1335	return 0;
1336}
1337
1338/*
1339 * Called for every backref that is found for the current extent.
1340 * Results are collected in sctx->clone_roots->ino/offset.
1341 */
1342static int iterate_backrefs(u64 ino, u64 offset, u64 num_bytes, u64 root_id,
1343			    void *ctx_)
1344{
1345	struct backref_ctx *bctx = ctx_;
1346	struct clone_root *clone_root;
1347
1348	/* First check if the root is in the list of accepted clone sources */
1349	clone_root = bsearch((void *)(uintptr_t)root_id, bctx->sctx->clone_roots,
1350			     bctx->sctx->clone_roots_cnt,
1351			     sizeof(struct clone_root),
1352			     __clone_root_cmp_bsearch);
1353	if (!clone_root)
1354		return 0;
1355
1356	/* This is our own reference, bail out as we can't clone from it. */
1357	if (clone_root->root == bctx->sctx->send_root &&
1358	    ino == bctx->cur_objectid &&
1359	    offset == bctx->cur_offset)
1360		return 0;
1361
1362	/*
1363	 * Make sure we don't consider clones from send_root that are
1364	 * behind the current inode/offset.
1365	 */
1366	if (clone_root->root == bctx->sctx->send_root) {
1367		/*
1368		 * If the source inode was not yet processed we can't issue a
1369		 * clone operation, as the source extent does not exist yet at
1370		 * the destination of the stream.
1371		 */
1372		if (ino > bctx->cur_objectid)
1373			return 0;
1374		/*
1375		 * We clone from the inode currently being sent as long as the
1376		 * source extent is already processed, otherwise we could try
1377		 * to clone from an extent that does not exist yet at the
1378		 * destination of the stream.
1379		 */
1380		if (ino == bctx->cur_objectid &&
1381		    offset + bctx->extent_len >
1382		    bctx->sctx->cur_inode_next_write_offset)
1383			return 0;
1384	}
1385
1386	bctx->found++;
1387	clone_root->found_ref = true;
1388
1389	/*
1390	 * If the given backref refers to a file extent item with a larger
1391	 * number of bytes than what we found before, use the new one so that
1392	 * we clone more optimally and end up doing less writes and getting
1393	 * less exclusive, non-shared extents at the destination.
1394	 */
1395	if (num_bytes > clone_root->num_bytes) {
1396		clone_root->ino = ino;
1397		clone_root->offset = offset;
1398		clone_root->num_bytes = num_bytes;
1399
1400		/*
1401		 * Found a perfect candidate, so there's no need to continue
1402		 * backref walking.
1403		 */
1404		if (num_bytes >= bctx->extent_len)
1405			return BTRFS_ITERATE_EXTENT_INODES_STOP;
1406	}
1407
1408	return 0;
1409}
1410
1411static bool lookup_backref_cache(u64 leaf_bytenr, void *ctx,
1412				 const u64 **root_ids_ret, int *root_count_ret)
1413{
1414	struct backref_ctx *bctx = ctx;
1415	struct send_ctx *sctx = bctx->sctx;
1416	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1417	const u64 key = leaf_bytenr >> fs_info->sectorsize_bits;
1418	struct btrfs_lru_cache_entry *raw_entry;
1419	struct backref_cache_entry *entry;
1420
1421	if (sctx->backref_cache.size == 0)
1422		return false;
1423
1424	/*
1425	 * If relocation happened since we first filled the cache, then we must
1426	 * empty the cache and can not use it, because even though we operate on
1427	 * read-only roots, their leaves and nodes may have been reallocated and
1428	 * now be used for different nodes/leaves of the same tree or some other
1429	 * tree.
1430	 *
1431	 * We are called from iterate_extent_inodes() while either holding a
1432	 * transaction handle or holding fs_info->commit_root_sem, so no need
1433	 * to take any lock here.
1434	 */
1435	if (fs_info->last_reloc_trans > sctx->backref_cache_last_reloc_trans) {
1436		btrfs_lru_cache_clear(&sctx->backref_cache);
1437		return false;
1438	}
1439
1440	raw_entry = btrfs_lru_cache_lookup(&sctx->backref_cache, key, 0);
1441	if (!raw_entry)
1442		return false;
1443
1444	entry = container_of(raw_entry, struct backref_cache_entry, entry);
1445	*root_ids_ret = entry->root_ids;
1446	*root_count_ret = entry->num_roots;
1447
1448	return true;
1449}
1450
1451static void store_backref_cache(u64 leaf_bytenr, const struct ulist *root_ids,
1452				void *ctx)
1453{
1454	struct backref_ctx *bctx = ctx;
1455	struct send_ctx *sctx = bctx->sctx;
1456	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1457	struct backref_cache_entry *new_entry;
1458	struct ulist_iterator uiter;
1459	struct ulist_node *node;
1460	int ret;
1461
1462	/*
1463	 * We're called while holding a transaction handle or while holding
1464	 * fs_info->commit_root_sem (at iterate_extent_inodes()), so must do a
1465	 * NOFS allocation.
1466	 */
1467	new_entry = kmalloc(sizeof(struct backref_cache_entry), GFP_NOFS);
1468	/* No worries, cache is optional. */
1469	if (!new_entry)
1470		return;
1471
1472	new_entry->entry.key = leaf_bytenr >> fs_info->sectorsize_bits;
1473	new_entry->entry.gen = 0;
1474	new_entry->num_roots = 0;
1475	ULIST_ITER_INIT(&uiter);
1476	while ((node = ulist_next(root_ids, &uiter)) != NULL) {
1477		const u64 root_id = node->val;
1478		struct clone_root *root;
1479
1480		root = bsearch((void *)(uintptr_t)root_id, sctx->clone_roots,
1481			       sctx->clone_roots_cnt, sizeof(struct clone_root),
1482			       __clone_root_cmp_bsearch);
1483		if (!root)
1484			continue;
1485
1486		/* Too many roots, just exit, no worries as caching is optional. */
1487		if (new_entry->num_roots >= SEND_MAX_BACKREF_CACHE_ROOTS) {
1488			kfree(new_entry);
1489			return;
1490		}
1491
1492		new_entry->root_ids[new_entry->num_roots] = root_id;
1493		new_entry->num_roots++;
1494	}
1495
1496	/*
1497	 * We may have not added any roots to the new cache entry, which means
1498	 * none of the roots is part of the list of roots from which we are
1499	 * allowed to clone. Cache the new entry as it's still useful to avoid
1500	 * backref walking to determine which roots have a path to the leaf.
1501	 *
1502	 * Also use GFP_NOFS because we're called while holding a transaction
1503	 * handle or while holding fs_info->commit_root_sem.
1504	 */
1505	ret = btrfs_lru_cache_store(&sctx->backref_cache, &new_entry->entry,
1506				    GFP_NOFS);
1507	ASSERT(ret == 0 || ret == -ENOMEM);
1508	if (ret) {
1509		/* Caching is optional, no worries. */
1510		kfree(new_entry);
1511		return;
1512	}
1513
1514	/*
1515	 * We are called from iterate_extent_inodes() while either holding a
1516	 * transaction handle or holding fs_info->commit_root_sem, so no need
1517	 * to take any lock here.
1518	 */
1519	if (sctx->backref_cache.size == 1)
1520		sctx->backref_cache_last_reloc_trans = fs_info->last_reloc_trans;
1521}
1522
1523static int check_extent_item(u64 bytenr, const struct btrfs_extent_item *ei,
1524			     const struct extent_buffer *leaf, void *ctx)
1525{
1526	const u64 refs = btrfs_extent_refs(leaf, ei);
1527	const struct backref_ctx *bctx = ctx;
1528	const struct send_ctx *sctx = bctx->sctx;
1529
1530	if (bytenr == bctx->bytenr) {
1531		const u64 flags = btrfs_extent_flags(leaf, ei);
1532
1533		if (WARN_ON(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
1534			return -EUCLEAN;
1535
1536		/*
1537		 * If we have only one reference and only the send root as a
1538		 * clone source - meaning no clone roots were given in the
1539		 * struct btrfs_ioctl_send_args passed to the send ioctl - then
1540		 * it's our reference and there's no point in doing backref
1541		 * walking which is expensive, so exit early.
1542		 */
1543		if (refs == 1 && sctx->clone_roots_cnt == 1)
1544			return -ENOENT;
1545	}
1546
1547	/*
1548	 * Backreference walking (iterate_extent_inodes() below) is currently
1549	 * too expensive when an extent has a large number of references, both
1550	 * in time spent and used memory. So for now just fallback to write
1551	 * operations instead of clone operations when an extent has more than
1552	 * a certain amount of references.
1553	 */
1554	if (refs > SEND_MAX_EXTENT_REFS)
1555		return -ENOENT;
1556
1557	return 0;
1558}
1559
1560static bool skip_self_data_ref(u64 root, u64 ino, u64 offset, void *ctx)
1561{
1562	const struct backref_ctx *bctx = ctx;
1563
1564	if (ino == bctx->cur_objectid &&
1565	    root == bctx->backref_owner &&
1566	    offset == bctx->backref_offset)
1567		return true;
1568
1569	return false;
1570}
1571
1572/*
1573 * Given an inode, offset and extent item, it finds a good clone for a clone
1574 * instruction. Returns -ENOENT when none could be found. The function makes
1575 * sure that the returned clone is usable at the point where sending is at the
1576 * moment. This means, that no clones are accepted which lie behind the current
1577 * inode+offset.
1578 *
1579 * path must point to the extent item when called.
1580 */
1581static int find_extent_clone(struct send_ctx *sctx,
1582			     struct btrfs_path *path,
1583			     u64 ino, u64 data_offset,
1584			     u64 ino_size,
1585			     struct clone_root **found)
1586{
1587	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1588	int ret;
1589	int extent_type;
1590	u64 logical;
1591	u64 disk_byte;
1592	u64 num_bytes;
1593	struct btrfs_file_extent_item *fi;
1594	struct extent_buffer *eb = path->nodes[0];
1595	struct backref_ctx backref_ctx = { 0 };
1596	struct btrfs_backref_walk_ctx backref_walk_ctx = { 0 };
1597	struct clone_root *cur_clone_root;
1598	int compressed;
1599	u32 i;
1600
1601	/*
1602	 * With fallocate we can get prealloc extents beyond the inode's i_size,
1603	 * so we don't do anything here because clone operations can not clone
1604	 * to a range beyond i_size without increasing the i_size of the
1605	 * destination inode.
1606	 */
1607	if (data_offset >= ino_size)
1608		return 0;
1609
1610	fi = btrfs_item_ptr(eb, path->slots[0], struct btrfs_file_extent_item);
1611	extent_type = btrfs_file_extent_type(eb, fi);
1612	if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1613		return -ENOENT;
1614
1615	disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1616	if (disk_byte == 0)
1617		return -ENOENT;
1618
1619	compressed = btrfs_file_extent_compression(eb, fi);
1620	num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1621	logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1622
1623	/*
1624	 * Setup the clone roots.
1625	 */
1626	for (i = 0; i < sctx->clone_roots_cnt; i++) {
1627		cur_clone_root = sctx->clone_roots + i;
1628		cur_clone_root->ino = (u64)-1;
1629		cur_clone_root->offset = 0;
1630		cur_clone_root->num_bytes = 0;
1631		cur_clone_root->found_ref = false;
1632	}
1633
1634	backref_ctx.sctx = sctx;
1635	backref_ctx.cur_objectid = ino;
1636	backref_ctx.cur_offset = data_offset;
1637	backref_ctx.bytenr = disk_byte;
1638	/*
1639	 * Use the header owner and not the send root's id, because in case of a
1640	 * snapshot we can have shared subtrees.
1641	 */
1642	backref_ctx.backref_owner = btrfs_header_owner(eb);
1643	backref_ctx.backref_offset = data_offset - btrfs_file_extent_offset(eb, fi);
1644
1645	/*
1646	 * The last extent of a file may be too large due to page alignment.
1647	 * We need to adjust extent_len in this case so that the checks in
1648	 * iterate_backrefs() work.
1649	 */
1650	if (data_offset + num_bytes >= ino_size)
1651		backref_ctx.extent_len = ino_size - data_offset;
1652	else
1653		backref_ctx.extent_len = num_bytes;
1654
1655	/*
1656	 * Now collect all backrefs.
1657	 */
1658	backref_walk_ctx.bytenr = disk_byte;
1659	if (compressed == BTRFS_COMPRESS_NONE)
1660		backref_walk_ctx.extent_item_pos = btrfs_file_extent_offset(eb, fi);
1661	backref_walk_ctx.fs_info = fs_info;
1662	backref_walk_ctx.cache_lookup = lookup_backref_cache;
1663	backref_walk_ctx.cache_store = store_backref_cache;
1664	backref_walk_ctx.indirect_ref_iterator = iterate_backrefs;
1665	backref_walk_ctx.check_extent_item = check_extent_item;
1666	backref_walk_ctx.user_ctx = &backref_ctx;
1667
1668	/*
1669	 * If have a single clone root, then it's the send root and we can tell
1670	 * the backref walking code to skip our own backref and not resolve it,
1671	 * since we can not use it for cloning - the source and destination
1672	 * ranges can't overlap and in case the leaf is shared through a subtree
1673	 * due to snapshots, we can't use those other roots since they are not
1674	 * in the list of clone roots.
1675	 */
1676	if (sctx->clone_roots_cnt == 1)
1677		backref_walk_ctx.skip_data_ref = skip_self_data_ref;
1678
1679	ret = iterate_extent_inodes(&backref_walk_ctx, true, iterate_backrefs,
1680				    &backref_ctx);
1681	if (ret < 0)
1682		return ret;
1683
1684	down_read(&fs_info->commit_root_sem);
1685	if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
1686		/*
1687		 * A transaction commit for a transaction in which block group
1688		 * relocation was done just happened.
1689		 * The disk_bytenr of the file extent item we processed is
1690		 * possibly stale, referring to the extent's location before
1691		 * relocation. So act as if we haven't found any clone sources
1692		 * and fallback to write commands, which will read the correct
1693		 * data from the new extent location. Otherwise we will fail
1694		 * below because we haven't found our own back reference or we
1695		 * could be getting incorrect sources in case the old extent
1696		 * was already reallocated after the relocation.
1697		 */
1698		up_read(&fs_info->commit_root_sem);
1699		return -ENOENT;
1700	}
1701	up_read(&fs_info->commit_root_sem);
1702
1703	btrfs_debug(fs_info,
1704		    "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1705		    data_offset, ino, num_bytes, logical);
1706
1707	if (!backref_ctx.found) {
1708		btrfs_debug(fs_info, "no clones found");
1709		return -ENOENT;
1710	}
1711
1712	cur_clone_root = NULL;
1713	for (i = 0; i < sctx->clone_roots_cnt; i++) {
1714		struct clone_root *clone_root = &sctx->clone_roots[i];
1715
1716		if (!clone_root->found_ref)
1717			continue;
1718
1719		/*
1720		 * Choose the root from which we can clone more bytes, to
1721		 * minimize write operations and therefore have more extent
1722		 * sharing at the destination (the same as in the source).
1723		 */
1724		if (!cur_clone_root ||
1725		    clone_root->num_bytes > cur_clone_root->num_bytes) {
1726			cur_clone_root = clone_root;
1727
1728			/*
1729			 * We found an optimal clone candidate (any inode from
1730			 * any root is fine), so we're done.
1731			 */
1732			if (clone_root->num_bytes >= backref_ctx.extent_len)
1733				break;
1734		}
1735	}
1736
1737	if (cur_clone_root) {
1738		*found = cur_clone_root;
1739		ret = 0;
1740	} else {
1741		ret = -ENOENT;
1742	}
1743
1744	return ret;
1745}
1746
1747static int read_symlink(struct btrfs_root *root,
1748			u64 ino,
1749			struct fs_path *dest)
1750{
1751	int ret;
1752	struct btrfs_path *path;
1753	struct btrfs_key key;
1754	struct btrfs_file_extent_item *ei;
1755	u8 type;
1756	u8 compression;
1757	unsigned long off;
1758	int len;
1759
1760	path = alloc_path_for_send();
1761	if (!path)
1762		return -ENOMEM;
1763
1764	key.objectid = ino;
1765	key.type = BTRFS_EXTENT_DATA_KEY;
1766	key.offset = 0;
1767	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1768	if (ret < 0)
1769		goto out;
1770	if (ret) {
1771		/*
1772		 * An empty symlink inode. Can happen in rare error paths when
1773		 * creating a symlink (transaction committed before the inode
1774		 * eviction handler removed the symlink inode items and a crash
1775		 * happened in between or the subvol was snapshoted in between).
1776		 * Print an informative message to dmesg/syslog so that the user
1777		 * can delete the symlink.
1778		 */
1779		btrfs_err(root->fs_info,
1780			  "Found empty symlink inode %llu at root %llu",
1781			  ino, root->root_key.objectid);
1782		ret = -EIO;
1783		goto out;
1784	}
1785
1786	ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1787			struct btrfs_file_extent_item);
1788	type = btrfs_file_extent_type(path->nodes[0], ei);
1789	if (unlikely(type != BTRFS_FILE_EXTENT_INLINE)) {
1790		ret = -EUCLEAN;
1791		btrfs_crit(root->fs_info,
1792"send: found symlink extent that is not inline, ino %llu root %llu extent type %d",
1793			   ino, btrfs_root_id(root), type);
1794		goto out;
1795	}
1796	compression = btrfs_file_extent_compression(path->nodes[0], ei);
1797	if (unlikely(compression != BTRFS_COMPRESS_NONE)) {
1798		ret = -EUCLEAN;
1799		btrfs_crit(root->fs_info,
1800"send: found symlink extent with compression, ino %llu root %llu compression type %d",
1801			   ino, btrfs_root_id(root), compression);
1802		goto out;
1803	}
1804
1805	off = btrfs_file_extent_inline_start(ei);
1806	len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
1807
1808	ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1809
1810out:
1811	btrfs_free_path(path);
1812	return ret;
1813}
1814
1815/*
1816 * Helper function to generate a file name that is unique in the root of
1817 * send_root and parent_root. This is used to generate names for orphan inodes.
1818 */
1819static int gen_unique_name(struct send_ctx *sctx,
1820			   u64 ino, u64 gen,
1821			   struct fs_path *dest)
1822{
1823	int ret = 0;
1824	struct btrfs_path *path;
1825	struct btrfs_dir_item *di;
1826	char tmp[64];
1827	int len;
1828	u64 idx = 0;
1829
1830	path = alloc_path_for_send();
1831	if (!path)
1832		return -ENOMEM;
1833
1834	while (1) {
1835		struct fscrypt_str tmp_name;
1836
1837		len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1838				ino, gen, idx);
1839		ASSERT(len < sizeof(tmp));
1840		tmp_name.name = tmp;
1841		tmp_name.len = strlen(tmp);
1842
1843		di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1844				path, BTRFS_FIRST_FREE_OBJECTID,
1845				&tmp_name, 0);
1846		btrfs_release_path(path);
1847		if (IS_ERR(di)) {
1848			ret = PTR_ERR(di);
1849			goto out;
1850		}
1851		if (di) {
1852			/* not unique, try again */
1853			idx++;
1854			continue;
1855		}
1856
1857		if (!sctx->parent_root) {
1858			/* unique */
1859			ret = 0;
1860			break;
1861		}
1862
1863		di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1864				path, BTRFS_FIRST_FREE_OBJECTID,
1865				&tmp_name, 0);
1866		btrfs_release_path(path);
1867		if (IS_ERR(di)) {
1868			ret = PTR_ERR(di);
1869			goto out;
1870		}
1871		if (di) {
1872			/* not unique, try again */
1873			idx++;
1874			continue;
1875		}
1876		/* unique */
1877		break;
1878	}
1879
1880	ret = fs_path_add(dest, tmp, strlen(tmp));
1881
1882out:
1883	btrfs_free_path(path);
1884	return ret;
1885}
1886
1887enum inode_state {
1888	inode_state_no_change,
1889	inode_state_will_create,
1890	inode_state_did_create,
1891	inode_state_will_delete,
1892	inode_state_did_delete,
1893};
1894
1895static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen,
1896			       u64 *send_gen, u64 *parent_gen)
1897{
1898	int ret;
1899	int left_ret;
1900	int right_ret;
1901	u64 left_gen;
1902	u64 right_gen = 0;
1903	struct btrfs_inode_info info;
1904
1905	ret = get_inode_info(sctx->send_root, ino, &info);
1906	if (ret < 0 && ret != -ENOENT)
1907		goto out;
1908	left_ret = (info.nlink == 0) ? -ENOENT : ret;
1909	left_gen = info.gen;
1910	if (send_gen)
1911		*send_gen = ((left_ret == -ENOENT) ? 0 : info.gen);
1912
1913	if (!sctx->parent_root) {
1914		right_ret = -ENOENT;
1915	} else {
1916		ret = get_inode_info(sctx->parent_root, ino, &info);
1917		if (ret < 0 && ret != -ENOENT)
1918			goto out;
1919		right_ret = (info.nlink == 0) ? -ENOENT : ret;
1920		right_gen = info.gen;
1921		if (parent_gen)
1922			*parent_gen = ((right_ret == -ENOENT) ? 0 : info.gen);
1923	}
1924
1925	if (!left_ret && !right_ret) {
1926		if (left_gen == gen && right_gen == gen) {
1927			ret = inode_state_no_change;
1928		} else if (left_gen == gen) {
1929			if (ino < sctx->send_progress)
1930				ret = inode_state_did_create;
1931			else
1932				ret = inode_state_will_create;
1933		} else if (right_gen == gen) {
1934			if (ino < sctx->send_progress)
1935				ret = inode_state_did_delete;
1936			else
1937				ret = inode_state_will_delete;
1938		} else  {
1939			ret = -ENOENT;
1940		}
1941	} else if (!left_ret) {
1942		if (left_gen == gen) {
1943			if (ino < sctx->send_progress)
1944				ret = inode_state_did_create;
1945			else
1946				ret = inode_state_will_create;
1947		} else {
1948			ret = -ENOENT;
1949		}
1950	} else if (!right_ret) {
1951		if (right_gen == gen) {
1952			if (ino < sctx->send_progress)
1953				ret = inode_state_did_delete;
1954			else
1955				ret = inode_state_will_delete;
1956		} else {
1957			ret = -ENOENT;
1958		}
1959	} else {
1960		ret = -ENOENT;
1961	}
1962
1963out:
1964	return ret;
1965}
1966
1967static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen,
1968			     u64 *send_gen, u64 *parent_gen)
1969{
1970	int ret;
1971
1972	if (ino == BTRFS_FIRST_FREE_OBJECTID)
1973		return 1;
1974
1975	ret = get_cur_inode_state(sctx, ino, gen, send_gen, parent_gen);
1976	if (ret < 0)
1977		goto out;
1978
1979	if (ret == inode_state_no_change ||
1980	    ret == inode_state_did_create ||
1981	    ret == inode_state_will_delete)
1982		ret = 1;
1983	else
1984		ret = 0;
1985
1986out:
1987	return ret;
1988}
1989
1990/*
1991 * Helper function to lookup a dir item in a dir.
1992 */
1993static int lookup_dir_item_inode(struct btrfs_root *root,
1994				 u64 dir, const char *name, int name_len,
1995				 u64 *found_inode)
1996{
1997	int ret = 0;
1998	struct btrfs_dir_item *di;
1999	struct btrfs_key key;
2000	struct btrfs_path *path;
2001	struct fscrypt_str name_str = FSTR_INIT((char *)name, name_len);
2002
2003	path = alloc_path_for_send();
2004	if (!path)
2005		return -ENOMEM;
2006
2007	di = btrfs_lookup_dir_item(NULL, root, path, dir, &name_str, 0);
2008	if (IS_ERR_OR_NULL(di)) {
2009		ret = di ? PTR_ERR(di) : -ENOENT;
2010		goto out;
2011	}
2012	btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
2013	if (key.type == BTRFS_ROOT_ITEM_KEY) {
2014		ret = -ENOENT;
2015		goto out;
2016	}
2017	*found_inode = key.objectid;
2018
2019out:
2020	btrfs_free_path(path);
2021	return ret;
2022}
2023
2024/*
2025 * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
2026 * generation of the parent dir and the name of the dir entry.
2027 */
2028static int get_first_ref(struct btrfs_root *root, u64 ino,
2029			 u64 *dir, u64 *dir_gen, struct fs_path *name)
2030{
2031	int ret;
2032	struct btrfs_key key;
2033	struct btrfs_key found_key;
2034	struct btrfs_path *path;
2035	int len;
2036	u64 parent_dir;
2037
2038	path = alloc_path_for_send();
2039	if (!path)
2040		return -ENOMEM;
2041
2042	key.objectid = ino;
2043	key.type = BTRFS_INODE_REF_KEY;
2044	key.offset = 0;
2045
2046	ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
2047	if (ret < 0)
2048		goto out;
2049	if (!ret)
2050		btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2051				path->slots[0]);
2052	if (ret || found_key.objectid != ino ||
2053	    (found_key.type != BTRFS_INODE_REF_KEY &&
2054	     found_key.type != BTRFS_INODE_EXTREF_KEY)) {
2055		ret = -ENOENT;
2056		goto out;
2057	}
2058
2059	if (found_key.type == BTRFS_INODE_REF_KEY) {
2060		struct btrfs_inode_ref *iref;
2061		iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
2062				      struct btrfs_inode_ref);
2063		len = btrfs_inode_ref_name_len(path->nodes[0], iref);
2064		ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
2065						     (unsigned long)(iref + 1),
2066						     len);
2067		parent_dir = found_key.offset;
2068	} else {
2069		struct btrfs_inode_extref *extref;
2070		extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
2071					struct btrfs_inode_extref);
2072		len = btrfs_inode_extref_name_len(path->nodes[0], extref);
2073		ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
2074					(unsigned long)&extref->name, len);
2075		parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
2076	}
2077	if (ret < 0)
2078		goto out;
2079	btrfs_release_path(path);
2080
2081	if (dir_gen) {
2082		ret = get_inode_gen(root, parent_dir, dir_gen);
2083		if (ret < 0)
2084			goto out;
2085	}
2086
2087	*dir = parent_dir;
2088
2089out:
2090	btrfs_free_path(path);
2091	return ret;
2092}
2093
2094static int is_first_ref(struct btrfs_root *root,
2095			u64 ino, u64 dir,
2096			const char *name, int name_len)
2097{
2098	int ret;
2099	struct fs_path *tmp_name;
2100	u64 tmp_dir;
2101
2102	tmp_name = fs_path_alloc();
2103	if (!tmp_name)
2104		return -ENOMEM;
2105
2106	ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
2107	if (ret < 0)
2108		goto out;
2109
2110	if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
2111		ret = 0;
2112		goto out;
2113	}
2114
2115	ret = !memcmp(tmp_name->start, name, name_len);
2116
2117out:
2118	fs_path_free(tmp_name);
2119	return ret;
2120}
2121
2122/*
2123 * Used by process_recorded_refs to determine if a new ref would overwrite an
2124 * already existing ref. In case it detects an overwrite, it returns the
2125 * inode/gen in who_ino/who_gen.
2126 * When an overwrite is detected, process_recorded_refs does proper orphanizing
2127 * to make sure later references to the overwritten inode are possible.
2128 * Orphanizing is however only required for the first ref of an inode.
2129 * process_recorded_refs does an additional is_first_ref check to see if
2130 * orphanizing is really required.
2131 */
2132static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2133			      const char *name, int name_len,
2134			      u64 *who_ino, u64 *who_gen, u64 *who_mode)
2135{
2136	int ret;
2137	u64 parent_root_dir_gen;
2138	u64 other_inode = 0;
2139	struct btrfs_inode_info info;
2140
2141	if (!sctx->parent_root)
2142		return 0;
2143
2144	ret = is_inode_existent(sctx, dir, dir_gen, NULL, &parent_root_dir_gen);
2145	if (ret <= 0)
2146		return 0;
2147
2148	/*
2149	 * If we have a parent root we need to verify that the parent dir was
2150	 * not deleted and then re-created, if it was then we have no overwrite
2151	 * and we can just unlink this entry.
2152	 *
2153	 * @parent_root_dir_gen was set to 0 if the inode does not exist in the
2154	 * parent root.
2155	 */
2156	if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID &&
2157	    parent_root_dir_gen != dir_gen)
2158		return 0;
2159
2160	ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
2161				    &other_inode);
2162	if (ret == -ENOENT)
2163		return 0;
2164	else if (ret < 0)
2165		return ret;
2166
2167	/*
2168	 * Check if the overwritten ref was already processed. If yes, the ref
2169	 * was already unlinked/moved, so we can safely assume that we will not
2170	 * overwrite anything at this point in time.
2171	 */
2172	if (other_inode > sctx->send_progress ||
2173	    is_waiting_for_move(sctx, other_inode)) {
2174		ret = get_inode_info(sctx->parent_root, other_inode, &info);
2175		if (ret < 0)
2176			return ret;
2177
2178		*who_ino = other_inode;
2179		*who_gen = info.gen;
2180		*who_mode = info.mode;
2181		return 1;
2182	}
2183
2184	return 0;
2185}
2186
2187/*
2188 * Checks if the ref was overwritten by an already processed inode. This is
2189 * used by __get_cur_name_and_parent to find out if the ref was orphanized and
2190 * thus the orphan name needs be used.
2191 * process_recorded_refs also uses it to avoid unlinking of refs that were
2192 * overwritten.
2193 */
2194static int did_overwrite_ref(struct send_ctx *sctx,
2195			    u64 dir, u64 dir_gen,
2196			    u64 ino, u64 ino_gen,
2197			    const char *name, int name_len)
2198{
2199	int ret;
2200	u64 ow_inode;
2201	u64 ow_gen = 0;
2202	u64 send_root_dir_gen;
2203
2204	if (!sctx->parent_root)
2205		return 0;
2206
2207	ret = is_inode_existent(sctx, dir, dir_gen, &send_root_dir_gen, NULL);
2208	if (ret <= 0)
2209		return ret;
2210
2211	/*
2212	 * @send_root_dir_gen was set to 0 if the inode does not exist in the
2213	 * send root.
2214	 */
2215	if (dir != BTRFS_FIRST_FREE_OBJECTID && send_root_dir_gen != dir_gen)
2216		return 0;
2217
2218	/* check if the ref was overwritten by another ref */
2219	ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
2220				    &ow_inode);
2221	if (ret == -ENOENT) {
2222		/* was never and will never be overwritten */
2223		return 0;
2224	} else if (ret < 0) {
2225		return ret;
2226	}
2227
2228	if (ow_inode == ino) {
2229		ret = get_inode_gen(sctx->send_root, ow_inode, &ow_gen);
2230		if (ret < 0)
2231			return ret;
2232
2233		/* It's the same inode, so no overwrite happened. */
2234		if (ow_gen == ino_gen)
2235			return 0;
2236	}
2237
2238	/*
2239	 * We know that it is or will be overwritten. Check this now.
2240	 * The current inode being processed might have been the one that caused
2241	 * inode 'ino' to be orphanized, therefore check if ow_inode matches
2242	 * the current inode being processed.
2243	 */
2244	if (ow_inode < sctx->send_progress)
2245		return 1;
2246
2247	if (ino != sctx->cur_ino && ow_inode == sctx->cur_ino) {
2248		if (ow_gen == 0) {
2249			ret = get_inode_gen(sctx->send_root, ow_inode, &ow_gen);
2250			if (ret < 0)
2251				return ret;
2252		}
2253		if (ow_gen == sctx->cur_inode_gen)
2254			return 1;
2255	}
2256
2257	return 0;
2258}
2259
2260/*
2261 * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
2262 * that got overwritten. This is used by process_recorded_refs to determine
2263 * if it has to use the path as returned by get_cur_path or the orphan name.
2264 */
2265static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
2266{
2267	int ret = 0;
2268	struct fs_path *name = NULL;
2269	u64 dir;
2270	u64 dir_gen;
2271
2272	if (!sctx->parent_root)
2273		goto out;
2274
2275	name = fs_path_alloc();
2276	if (!name)
2277		return -ENOMEM;
2278
2279	ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
2280	if (ret < 0)
2281		goto out;
2282
2283	ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
2284			name->start, fs_path_len(name));
2285
2286out:
2287	fs_path_free(name);
2288	return ret;
2289}
2290
2291static inline struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2292							 u64 ino, u64 gen)
2293{
2294	struct btrfs_lru_cache_entry *entry;
2295
2296	entry = btrfs_lru_cache_lookup(&sctx->name_cache, ino, gen);
2297	if (!entry)
2298		return NULL;
2299
2300	return container_of(entry, struct name_cache_entry, entry);
2301}
2302
2303/*
2304 * Used by get_cur_path for each ref up to the root.
2305 * Returns 0 if it succeeded.
2306 * Returns 1 if the inode is not existent or got overwritten. In that case, the
2307 * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2308 * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2309 * Returns <0 in case of error.
2310 */
2311static int __get_cur_name_and_parent(struct send_ctx *sctx,
2312				     u64 ino, u64 gen,
2313				     u64 *parent_ino,
2314				     u64 *parent_gen,
2315				     struct fs_path *dest)
2316{
2317	int ret;
2318	int nce_ret;
2319	struct name_cache_entry *nce;
2320
2321	/*
2322	 * First check if we already did a call to this function with the same
2323	 * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2324	 * return the cached result.
2325	 */
2326	nce = name_cache_search(sctx, ino, gen);
2327	if (nce) {
2328		if (ino < sctx->send_progress && nce->need_later_update) {
2329			btrfs_lru_cache_remove(&sctx->name_cache, &nce->entry);
2330			nce = NULL;
2331		} else {
2332			*parent_ino = nce->parent_ino;
2333			*parent_gen = nce->parent_gen;
2334			ret = fs_path_add(dest, nce->name, nce->name_len);
2335			if (ret < 0)
2336				goto out;
2337			ret = nce->ret;
2338			goto out;
2339		}
2340	}
2341
2342	/*
2343	 * If the inode is not existent yet, add the orphan name and return 1.
2344	 * This should only happen for the parent dir that we determine in
2345	 * record_new_ref_if_needed().
2346	 */
2347	ret = is_inode_existent(sctx, ino, gen, NULL, NULL);
2348	if (ret < 0)
2349		goto out;
2350
2351	if (!ret) {
2352		ret = gen_unique_name(sctx, ino, gen, dest);
2353		if (ret < 0)
2354			goto out;
2355		ret = 1;
2356		goto out_cache;
2357	}
2358
2359	/*
2360	 * Depending on whether the inode was already processed or not, use
2361	 * send_root or parent_root for ref lookup.
2362	 */
2363	if (ino < sctx->send_progress)
2364		ret = get_first_ref(sctx->send_root, ino,
2365				    parent_ino, parent_gen, dest);
2366	else
2367		ret = get_first_ref(sctx->parent_root, ino,
2368				    parent_ino, parent_gen, dest);
2369	if (ret < 0)
2370		goto out;
2371
2372	/*
2373	 * Check if the ref was overwritten by an inode's ref that was processed
2374	 * earlier. If yes, treat as orphan and return 1.
2375	 */
2376	ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2377			dest->start, dest->end - dest->start);
2378	if (ret < 0)
2379		goto out;
2380	if (ret) {
2381		fs_path_reset(dest);
2382		ret = gen_unique_name(sctx, ino, gen, dest);
2383		if (ret < 0)
2384			goto out;
2385		ret = 1;
2386	}
2387
2388out_cache:
2389	/*
2390	 * Store the result of the lookup in the name cache.
2391	 */
2392	nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2393	if (!nce) {
2394		ret = -ENOMEM;
2395		goto out;
2396	}
2397
2398	nce->entry.key = ino;
2399	nce->entry.gen = gen;
2400	nce->parent_ino = *parent_ino;
2401	nce->parent_gen = *parent_gen;
2402	nce->name_len = fs_path_len(dest);
2403	nce->ret = ret;
2404	strcpy(nce->name, dest->start);
2405
2406	if (ino < sctx->send_progress)
2407		nce->need_later_update = 0;
2408	else
2409		nce->need_later_update = 1;
2410
2411	nce_ret = btrfs_lru_cache_store(&sctx->name_cache, &nce->entry, GFP_KERNEL);
2412	if (nce_ret < 0) {
2413		kfree(nce);
2414		ret = nce_ret;
2415	}
2416
2417out:
2418	return ret;
2419}
2420
2421/*
2422 * Magic happens here. This function returns the first ref to an inode as it
2423 * would look like while receiving the stream at this point in time.
2424 * We walk the path up to the root. For every inode in between, we check if it
2425 * was already processed/sent. If yes, we continue with the parent as found
2426 * in send_root. If not, we continue with the parent as found in parent_root.
2427 * If we encounter an inode that was deleted at this point in time, we use the
2428 * inodes "orphan" name instead of the real name and stop. Same with new inodes
2429 * that were not created yet and overwritten inodes/refs.
2430 *
2431 * When do we have orphan inodes:
2432 * 1. When an inode is freshly created and thus no valid refs are available yet
2433 * 2. When a directory lost all it's refs (deleted) but still has dir items
2434 *    inside which were not processed yet (pending for move/delete). If anyone
2435 *    tried to get the path to the dir items, it would get a path inside that
2436 *    orphan directory.
2437 * 3. When an inode is moved around or gets new links, it may overwrite the ref
2438 *    of an unprocessed inode. If in that case the first ref would be
2439 *    overwritten, the overwritten inode gets "orphanized". Later when we
2440 *    process this overwritten inode, it is restored at a new place by moving
2441 *    the orphan inode.
2442 *
2443 * sctx->send_progress tells this function at which point in time receiving
2444 * would be.
2445 */
2446static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2447			struct fs_path *dest)
2448{
2449	int ret = 0;
2450	struct fs_path *name = NULL;
2451	u64 parent_inode = 0;
2452	u64 parent_gen = 0;
2453	int stop = 0;
2454
2455	name = fs_path_alloc();
2456	if (!name) {
2457		ret = -ENOMEM;
2458		goto out;
2459	}
2460
2461	dest->reversed = 1;
2462	fs_path_reset(dest);
2463
2464	while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2465		struct waiting_dir_move *wdm;
2466
2467		fs_path_reset(name);
2468
2469		if (is_waiting_for_rm(sctx, ino, gen)) {
2470			ret = gen_unique_name(sctx, ino, gen, name);
2471			if (ret < 0)
2472				goto out;
2473			ret = fs_path_add_path(dest, name);
2474			break;
2475		}
2476
2477		wdm = get_waiting_dir_move(sctx, ino);
2478		if (wdm && wdm->orphanized) {
2479			ret = gen_unique_name(sctx, ino, gen, name);
2480			stop = 1;
2481		} else if (wdm) {
2482			ret = get_first_ref(sctx->parent_root, ino,
2483					    &parent_inode, &parent_gen, name);
2484		} else {
2485			ret = __get_cur_name_and_parent(sctx, ino, gen,
2486							&parent_inode,
2487							&parent_gen, name);
2488			if (ret)
2489				stop = 1;
2490		}
2491
2492		if (ret < 0)
2493			goto out;
2494
2495		ret = fs_path_add_path(dest, name);
2496		if (ret < 0)
2497			goto out;
2498
2499		ino = parent_inode;
2500		gen = parent_gen;
2501	}
2502
2503out:
2504	fs_path_free(name);
2505	if (!ret)
2506		fs_path_unreverse(dest);
2507	return ret;
2508}
2509
2510/*
2511 * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2512 */
2513static int send_subvol_begin(struct send_ctx *sctx)
2514{
2515	int ret;
2516	struct btrfs_root *send_root = sctx->send_root;
2517	struct btrfs_root *parent_root = sctx->parent_root;
2518	struct btrfs_path *path;
2519	struct btrfs_key key;
2520	struct btrfs_root_ref *ref;
2521	struct extent_buffer *leaf;
2522	char *name = NULL;
2523	int namelen;
2524
2525	path = btrfs_alloc_path();
2526	if (!path)
2527		return -ENOMEM;
2528
2529	name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2530	if (!name) {
2531		btrfs_free_path(path);
2532		return -ENOMEM;
2533	}
2534
2535	key.objectid = send_root->root_key.objectid;
2536	key.type = BTRFS_ROOT_BACKREF_KEY;
2537	key.offset = 0;
2538
2539	ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2540				&key, path, 1, 0);
2541	if (ret < 0)
2542		goto out;
2543	if (ret) {
2544		ret = -ENOENT;
2545		goto out;
2546	}
2547
2548	leaf = path->nodes[0];
2549	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2550	if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2551	    key.objectid != send_root->root_key.objectid) {
2552		ret = -ENOENT;
2553		goto out;
2554	}
2555	ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2556	namelen = btrfs_root_ref_name_len(leaf, ref);
2557	read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2558	btrfs_release_path(path);
2559
2560	if (parent_root) {
2561		ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2562		if (ret < 0)
2563			goto out;
2564	} else {
2565		ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2566		if (ret < 0)
2567			goto out;
2568	}
2569
2570	TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2571
2572	if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2573		TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2574			    sctx->send_root->root_item.received_uuid);
2575	else
2576		TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2577			    sctx->send_root->root_item.uuid);
2578
2579	TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2580		    btrfs_root_ctransid(&sctx->send_root->root_item));
2581	if (parent_root) {
2582		if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2583			TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2584				     parent_root->root_item.received_uuid);
2585		else
2586			TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2587				     parent_root->root_item.uuid);
2588		TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2589			    btrfs_root_ctransid(&sctx->parent_root->root_item));
2590	}
2591
2592	ret = send_cmd(sctx);
2593
2594tlv_put_failure:
2595out:
2596	btrfs_free_path(path);
2597	kfree(name);
2598	return ret;
2599}
2600
2601static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2602{
2603	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2604	int ret = 0;
2605	struct fs_path *p;
2606
2607	btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2608
2609	p = fs_path_alloc();
2610	if (!p)
2611		return -ENOMEM;
2612
2613	ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2614	if (ret < 0)
2615		goto out;
2616
2617	ret = get_cur_path(sctx, ino, gen, p);
2618	if (ret < 0)
2619		goto out;
2620	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2621	TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2622
2623	ret = send_cmd(sctx);
2624
2625tlv_put_failure:
2626out:
2627	fs_path_free(p);
2628	return ret;
2629}
2630
2631static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2632{
2633	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2634	int ret = 0;
2635	struct fs_path *p;
2636
2637	btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2638
2639	p = fs_path_alloc();
2640	if (!p)
2641		return -ENOMEM;
2642
2643	ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2644	if (ret < 0)
2645		goto out;
2646
2647	ret = get_cur_path(sctx, ino, gen, p);
2648	if (ret < 0)
2649		goto out;
2650	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2651	TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2652
2653	ret = send_cmd(sctx);
2654
2655tlv_put_failure:
2656out:
2657	fs_path_free(p);
2658	return ret;
2659}
2660
2661static int send_fileattr(struct send_ctx *sctx, u64 ino, u64 gen, u64 fileattr)
2662{
2663	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2664	int ret = 0;
2665	struct fs_path *p;
2666
2667	if (sctx->proto < 2)
2668		return 0;
2669
2670	btrfs_debug(fs_info, "send_fileattr %llu fileattr=%llu", ino, fileattr);
2671
2672	p = fs_path_alloc();
2673	if (!p)
2674		return -ENOMEM;
2675
2676	ret = begin_cmd(sctx, BTRFS_SEND_C_FILEATTR);
2677	if (ret < 0)
2678		goto out;
2679
2680	ret = get_cur_path(sctx, ino, gen, p);
2681	if (ret < 0)
2682		goto out;
2683	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2684	TLV_PUT_U64(sctx, BTRFS_SEND_A_FILEATTR, fileattr);
2685
2686	ret = send_cmd(sctx);
2687
2688tlv_put_failure:
2689out:
2690	fs_path_free(p);
2691	return ret;
2692}
2693
2694static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2695{
2696	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2697	int ret = 0;
2698	struct fs_path *p;
2699
2700	btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2701		    ino, uid, gid);
2702
2703	p = fs_path_alloc();
2704	if (!p)
2705		return -ENOMEM;
2706
2707	ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2708	if (ret < 0)
2709		goto out;
2710
2711	ret = get_cur_path(sctx, ino, gen, p);
2712	if (ret < 0)
2713		goto out;
2714	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2715	TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2716	TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2717
2718	ret = send_cmd(sctx);
2719
2720tlv_put_failure:
2721out:
2722	fs_path_free(p);
2723	return ret;
2724}
2725
2726static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2727{
2728	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2729	int ret = 0;
2730	struct fs_path *p = NULL;
2731	struct btrfs_inode_item *ii;
2732	struct btrfs_path *path = NULL;
2733	struct extent_buffer *eb;
2734	struct btrfs_key key;
2735	int slot;
2736
2737	btrfs_debug(fs_info, "send_utimes %llu", ino);
2738
2739	p = fs_path_alloc();
2740	if (!p)
2741		return -ENOMEM;
2742
2743	path = alloc_path_for_send();
2744	if (!path) {
2745		ret = -ENOMEM;
2746		goto out;
2747	}
2748
2749	key.objectid = ino;
2750	key.type = BTRFS_INODE_ITEM_KEY;
2751	key.offset = 0;
2752	ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2753	if (ret > 0)
2754		ret = -ENOENT;
2755	if (ret < 0)
2756		goto out;
2757
2758	eb = path->nodes[0];
2759	slot = path->slots[0];
2760	ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2761
2762	ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2763	if (ret < 0)
2764		goto out;
2765
2766	ret = get_cur_path(sctx, ino, gen, p);
2767	if (ret < 0)
2768		goto out;
2769	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2770	TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2771	TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2772	TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2773	if (sctx->proto >= 2)
2774		TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_OTIME, eb, &ii->otime);
2775
2776	ret = send_cmd(sctx);
2777
2778tlv_put_failure:
2779out:
2780	fs_path_free(p);
2781	btrfs_free_path(path);
2782	return ret;
2783}
2784
2785/*
2786 * If the cache is full, we can't remove entries from it and do a call to
2787 * send_utimes() for each respective inode, because we might be finishing
2788 * processing an inode that is a directory and it just got renamed, and existing
2789 * entries in the cache may refer to inodes that have the directory in their
2790 * full path - in which case we would generate outdated paths (pre-rename)
2791 * for the inodes that the cache entries point to. Instead of prunning the
2792 * cache when inserting, do it after we finish processing each inode at
2793 * finish_inode_if_needed().
2794 */
2795static int cache_dir_utimes(struct send_ctx *sctx, u64 dir, u64 gen)
2796{
2797	struct btrfs_lru_cache_entry *entry;
2798	int ret;
2799
2800	entry = btrfs_lru_cache_lookup(&sctx->dir_utimes_cache, dir, gen);
2801	if (entry != NULL)
2802		return 0;
2803
2804	/* Caching is optional, don't fail if we can't allocate memory. */
2805	entry = kmalloc(sizeof(*entry), GFP_KERNEL);
2806	if (!entry)
2807		return send_utimes(sctx, dir, gen);
2808
2809	entry->key = dir;
2810	entry->gen = gen;
2811
2812	ret = btrfs_lru_cache_store(&sctx->dir_utimes_cache, entry, GFP_KERNEL);
2813	ASSERT(ret != -EEXIST);
2814	if (ret) {
2815		kfree(entry);
2816		return send_utimes(sctx, dir, gen);
2817	}
2818
2819	return 0;
2820}
2821
2822static int trim_dir_utimes_cache(struct send_ctx *sctx)
2823{
2824	while (sctx->dir_utimes_cache.size > SEND_MAX_DIR_UTIMES_CACHE_SIZE) {
 
2825		struct btrfs_lru_cache_entry *lru;
2826		int ret;
2827
2828		lru = btrfs_lru_cache_lru_entry(&sctx->dir_utimes_cache);
2829		ASSERT(lru != NULL);
2830
2831		ret = send_utimes(sctx, lru->key, lru->gen);
2832		if (ret)
2833			return ret;
2834
2835		btrfs_lru_cache_remove(&sctx->dir_utimes_cache, lru);
2836	}
2837
2838	return 0;
2839}
2840
2841/*
2842 * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2843 * a valid path yet because we did not process the refs yet. So, the inode
2844 * is created as orphan.
2845 */
2846static int send_create_inode(struct send_ctx *sctx, u64 ino)
2847{
2848	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2849	int ret = 0;
2850	struct fs_path *p;
2851	int cmd;
2852	struct btrfs_inode_info info;
2853	u64 gen;
2854	u64 mode;
2855	u64 rdev;
2856
2857	btrfs_debug(fs_info, "send_create_inode %llu", ino);
2858
2859	p = fs_path_alloc();
2860	if (!p)
2861		return -ENOMEM;
2862
2863	if (ino != sctx->cur_ino) {
2864		ret = get_inode_info(sctx->send_root, ino, &info);
2865		if (ret < 0)
2866			goto out;
2867		gen = info.gen;
2868		mode = info.mode;
2869		rdev = info.rdev;
2870	} else {
2871		gen = sctx->cur_inode_gen;
2872		mode = sctx->cur_inode_mode;
2873		rdev = sctx->cur_inode_rdev;
2874	}
2875
2876	if (S_ISREG(mode)) {
2877		cmd = BTRFS_SEND_C_MKFILE;
2878	} else if (S_ISDIR(mode)) {
2879		cmd = BTRFS_SEND_C_MKDIR;
2880	} else if (S_ISLNK(mode)) {
2881		cmd = BTRFS_SEND_C_SYMLINK;
2882	} else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2883		cmd = BTRFS_SEND_C_MKNOD;
2884	} else if (S_ISFIFO(mode)) {
2885		cmd = BTRFS_SEND_C_MKFIFO;
2886	} else if (S_ISSOCK(mode)) {
2887		cmd = BTRFS_SEND_C_MKSOCK;
2888	} else {
2889		btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2890				(int)(mode & S_IFMT));
2891		ret = -EOPNOTSUPP;
2892		goto out;
2893	}
2894
2895	ret = begin_cmd(sctx, cmd);
2896	if (ret < 0)
2897		goto out;
2898
2899	ret = gen_unique_name(sctx, ino, gen, p);
2900	if (ret < 0)
2901		goto out;
2902
2903	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2904	TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2905
2906	if (S_ISLNK(mode)) {
2907		fs_path_reset(p);
2908		ret = read_symlink(sctx->send_root, ino, p);
2909		if (ret < 0)
2910			goto out;
2911		TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2912	} else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2913		   S_ISFIFO(mode) || S_ISSOCK(mode)) {
2914		TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2915		TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2916	}
2917
2918	ret = send_cmd(sctx);
2919	if (ret < 0)
2920		goto out;
2921
2922
2923tlv_put_failure:
2924out:
2925	fs_path_free(p);
2926	return ret;
2927}
2928
2929static void cache_dir_created(struct send_ctx *sctx, u64 dir)
2930{
2931	struct btrfs_lru_cache_entry *entry;
2932	int ret;
2933
2934	/* Caching is optional, ignore any failures. */
2935	entry = kmalloc(sizeof(*entry), GFP_KERNEL);
2936	if (!entry)
2937		return;
2938
2939	entry->key = dir;
2940	entry->gen = 0;
2941	ret = btrfs_lru_cache_store(&sctx->dir_created_cache, entry, GFP_KERNEL);
2942	if (ret < 0)
2943		kfree(entry);
2944}
2945
2946/*
2947 * We need some special handling for inodes that get processed before the parent
2948 * directory got created. See process_recorded_refs for details.
2949 * This function does the check if we already created the dir out of order.
2950 */
2951static int did_create_dir(struct send_ctx *sctx, u64 dir)
2952{
2953	int ret = 0;
2954	int iter_ret = 0;
2955	struct btrfs_path *path = NULL;
2956	struct btrfs_key key;
2957	struct btrfs_key found_key;
2958	struct btrfs_key di_key;
2959	struct btrfs_dir_item *di;
2960
2961	if (btrfs_lru_cache_lookup(&sctx->dir_created_cache, dir, 0))
2962		return 1;
2963
2964	path = alloc_path_for_send();
2965	if (!path)
2966		return -ENOMEM;
2967
2968	key.objectid = dir;
2969	key.type = BTRFS_DIR_INDEX_KEY;
2970	key.offset = 0;
2971
2972	btrfs_for_each_slot(sctx->send_root, &key, &found_key, path, iter_ret) {
2973		struct extent_buffer *eb = path->nodes[0];
2974
2975		if (found_key.objectid != key.objectid ||
2976		    found_key.type != key.type) {
2977			ret = 0;
2978			break;
2979		}
2980
2981		di = btrfs_item_ptr(eb, path->slots[0], struct btrfs_dir_item);
2982		btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2983
2984		if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2985		    di_key.objectid < sctx->send_progress) {
2986			ret = 1;
2987			cache_dir_created(sctx, dir);
2988			break;
2989		}
2990	}
2991	/* Catch error found during iteration */
2992	if (iter_ret < 0)
2993		ret = iter_ret;
2994
2995	btrfs_free_path(path);
2996	return ret;
2997}
2998
2999/*
3000 * Only creates the inode if it is:
3001 * 1. Not a directory
3002 * 2. Or a directory which was not created already due to out of order
3003 *    directories. See did_create_dir and process_recorded_refs for details.
3004 */
3005static int send_create_inode_if_needed(struct send_ctx *sctx)
3006{
3007	int ret;
3008
3009	if (S_ISDIR(sctx->cur_inode_mode)) {
3010		ret = did_create_dir(sctx, sctx->cur_ino);
3011		if (ret < 0)
3012			return ret;
3013		else if (ret > 0)
3014			return 0;
3015	}
3016
3017	ret = send_create_inode(sctx, sctx->cur_ino);
3018
3019	if (ret == 0 && S_ISDIR(sctx->cur_inode_mode))
3020		cache_dir_created(sctx, sctx->cur_ino);
3021
3022	return ret;
3023}
3024
3025struct recorded_ref {
3026	struct list_head list;
3027	char *name;
3028	struct fs_path *full_path;
3029	u64 dir;
3030	u64 dir_gen;
3031	int name_len;
3032	struct rb_node node;
3033	struct rb_root *root;
3034};
3035
3036static struct recorded_ref *recorded_ref_alloc(void)
3037{
3038	struct recorded_ref *ref;
3039
3040	ref = kzalloc(sizeof(*ref), GFP_KERNEL);
3041	if (!ref)
3042		return NULL;
3043	RB_CLEAR_NODE(&ref->node);
3044	INIT_LIST_HEAD(&ref->list);
3045	return ref;
3046}
3047
3048static void recorded_ref_free(struct recorded_ref *ref)
3049{
3050	if (!ref)
3051		return;
3052	if (!RB_EMPTY_NODE(&ref->node))
3053		rb_erase(&ref->node, ref->root);
3054	list_del(&ref->list);
3055	fs_path_free(ref->full_path);
3056	kfree(ref);
3057}
3058
3059static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
3060{
3061	ref->full_path = path;
3062	ref->name = (char *)kbasename(ref->full_path->start);
3063	ref->name_len = ref->full_path->end - ref->name;
3064}
3065
3066static int dup_ref(struct recorded_ref *ref, struct list_head *list)
3067{
3068	struct recorded_ref *new;
3069
3070	new = recorded_ref_alloc();
3071	if (!new)
3072		return -ENOMEM;
3073
3074	new->dir = ref->dir;
3075	new->dir_gen = ref->dir_gen;
3076	list_add_tail(&new->list, list);
3077	return 0;
3078}
3079
3080static void __free_recorded_refs(struct list_head *head)
3081{
3082	struct recorded_ref *cur;
3083
3084	while (!list_empty(head)) {
3085		cur = list_entry(head->next, struct recorded_ref, list);
3086		recorded_ref_free(cur);
3087	}
3088}
3089
3090static void free_recorded_refs(struct send_ctx *sctx)
3091{
3092	__free_recorded_refs(&sctx->new_refs);
3093	__free_recorded_refs(&sctx->deleted_refs);
3094}
3095
3096/*
3097 * Renames/moves a file/dir to its orphan name. Used when the first
3098 * ref of an unprocessed inode gets overwritten and for all non empty
3099 * directories.
3100 */
3101static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
3102			  struct fs_path *path)
3103{
3104	int ret;
3105	struct fs_path *orphan;
3106
3107	orphan = fs_path_alloc();
3108	if (!orphan)
3109		return -ENOMEM;
3110
3111	ret = gen_unique_name(sctx, ino, gen, orphan);
3112	if (ret < 0)
3113		goto out;
3114
3115	ret = send_rename(sctx, path, orphan);
3116
3117out:
3118	fs_path_free(orphan);
3119	return ret;
3120}
3121
3122static struct orphan_dir_info *add_orphan_dir_info(struct send_ctx *sctx,
3123						   u64 dir_ino, u64 dir_gen)
3124{
3125	struct rb_node **p = &sctx->orphan_dirs.rb_node;
3126	struct rb_node *parent = NULL;
3127	struct orphan_dir_info *entry, *odi;
3128
3129	while (*p) {
3130		parent = *p;
3131		entry = rb_entry(parent, struct orphan_dir_info, node);
3132		if (dir_ino < entry->ino)
3133			p = &(*p)->rb_left;
3134		else if (dir_ino > entry->ino)
3135			p = &(*p)->rb_right;
3136		else if (dir_gen < entry->gen)
3137			p = &(*p)->rb_left;
3138		else if (dir_gen > entry->gen)
3139			p = &(*p)->rb_right;
3140		else
3141			return entry;
3142	}
3143
3144	odi = kmalloc(sizeof(*odi), GFP_KERNEL);
3145	if (!odi)
3146		return ERR_PTR(-ENOMEM);
3147	odi->ino = dir_ino;
3148	odi->gen = dir_gen;
3149	odi->last_dir_index_offset = 0;
3150	odi->dir_high_seq_ino = 0;
3151
3152	rb_link_node(&odi->node, parent, p);
3153	rb_insert_color(&odi->node, &sctx->orphan_dirs);
3154	return odi;
3155}
3156
3157static struct orphan_dir_info *get_orphan_dir_info(struct send_ctx *sctx,
3158						   u64 dir_ino, u64 gen)
3159{
3160	struct rb_node *n = sctx->orphan_dirs.rb_node;
3161	struct orphan_dir_info *entry;
3162
3163	while (n) {
3164		entry = rb_entry(n, struct orphan_dir_info, node);
3165		if (dir_ino < entry->ino)
3166			n = n->rb_left;
3167		else if (dir_ino > entry->ino)
3168			n = n->rb_right;
3169		else if (gen < entry->gen)
3170			n = n->rb_left;
3171		else if (gen > entry->gen)
3172			n = n->rb_right;
3173		else
3174			return entry;
3175	}
3176	return NULL;
3177}
3178
3179static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen)
3180{
3181	struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino, gen);
3182
3183	return odi != NULL;
3184}
3185
3186static void free_orphan_dir_info(struct send_ctx *sctx,
3187				 struct orphan_dir_info *odi)
3188{
3189	if (!odi)
3190		return;
3191	rb_erase(&odi->node, &sctx->orphan_dirs);
3192	kfree(odi);
3193}
3194
3195/*
3196 * Returns 1 if a directory can be removed at this point in time.
3197 * We check this by iterating all dir items and checking if the inode behind
3198 * the dir item was already processed.
3199 */
3200static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen)
3201{
3202	int ret = 0;
3203	int iter_ret = 0;
3204	struct btrfs_root *root = sctx->parent_root;
3205	struct btrfs_path *path;
3206	struct btrfs_key key;
3207	struct btrfs_key found_key;
3208	struct btrfs_key loc;
3209	struct btrfs_dir_item *di;
3210	struct orphan_dir_info *odi = NULL;
3211	u64 dir_high_seq_ino = 0;
3212	u64 last_dir_index_offset = 0;
3213
3214	/*
3215	 * Don't try to rmdir the top/root subvolume dir.
3216	 */
3217	if (dir == BTRFS_FIRST_FREE_OBJECTID)
3218		return 0;
3219
3220	odi = get_orphan_dir_info(sctx, dir, dir_gen);
3221	if (odi && sctx->cur_ino < odi->dir_high_seq_ino)
3222		return 0;
3223
3224	path = alloc_path_for_send();
3225	if (!path)
3226		return -ENOMEM;
3227
3228	if (!odi) {
3229		/*
3230		 * Find the inode number associated with the last dir index
3231		 * entry. This is very likely the inode with the highest number
3232		 * of all inodes that have an entry in the directory. We can
3233		 * then use it to avoid future calls to can_rmdir(), when
3234		 * processing inodes with a lower number, from having to search
3235		 * the parent root b+tree for dir index keys.
3236		 */
3237		key.objectid = dir;
3238		key.type = BTRFS_DIR_INDEX_KEY;
3239		key.offset = (u64)-1;
3240
3241		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3242		if (ret < 0) {
3243			goto out;
3244		} else if (ret > 0) {
3245			/* Can't happen, the root is never empty. */
3246			ASSERT(path->slots[0] > 0);
3247			if (WARN_ON(path->slots[0] == 0)) {
3248				ret = -EUCLEAN;
3249				goto out;
3250			}
3251			path->slots[0]--;
3252		}
3253
3254		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3255		if (key.objectid != dir || key.type != BTRFS_DIR_INDEX_KEY) {
3256			/* No index keys, dir can be removed. */
3257			ret = 1;
3258			goto out;
3259		}
3260
3261		di = btrfs_item_ptr(path->nodes[0], path->slots[0],
3262				    struct btrfs_dir_item);
3263		btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
3264		dir_high_seq_ino = loc.objectid;
3265		if (sctx->cur_ino < dir_high_seq_ino) {
3266			ret = 0;
3267			goto out;
3268		}
3269
3270		btrfs_release_path(path);
3271	}
3272
3273	key.objectid = dir;
3274	key.type = BTRFS_DIR_INDEX_KEY;
3275	key.offset = (odi ? odi->last_dir_index_offset : 0);
3276
3277	btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
3278		struct waiting_dir_move *dm;
3279
3280		if (found_key.objectid != key.objectid ||
3281		    found_key.type != key.type)
3282			break;
3283
3284		di = btrfs_item_ptr(path->nodes[0], path->slots[0],
3285				struct btrfs_dir_item);
3286		btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
3287
3288		dir_high_seq_ino = max(dir_high_seq_ino, loc.objectid);
3289		last_dir_index_offset = found_key.offset;
3290
3291		dm = get_waiting_dir_move(sctx, loc.objectid);
3292		if (dm) {
3293			dm->rmdir_ino = dir;
3294			dm->rmdir_gen = dir_gen;
3295			ret = 0;
3296			goto out;
3297		}
3298
3299		if (loc.objectid > sctx->cur_ino) {
3300			ret = 0;
3301			goto out;
3302		}
3303	}
3304	if (iter_ret < 0) {
3305		ret = iter_ret;
3306		goto out;
3307	}
3308	free_orphan_dir_info(sctx, odi);
3309
3310	ret = 1;
3311
3312out:
3313	btrfs_free_path(path);
3314
3315	if (ret)
3316		return ret;
3317
3318	if (!odi) {
3319		odi = add_orphan_dir_info(sctx, dir, dir_gen);
3320		if (IS_ERR(odi))
3321			return PTR_ERR(odi);
3322
3323		odi->gen = dir_gen;
3324	}
3325
3326	odi->last_dir_index_offset = last_dir_index_offset;
3327	odi->dir_high_seq_ino = max(odi->dir_high_seq_ino, dir_high_seq_ino);
3328
3329	return 0;
3330}
3331
3332static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3333{
3334	struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3335
3336	return entry != NULL;
3337}
3338
3339static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3340{
3341	struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3342	struct rb_node *parent = NULL;
3343	struct waiting_dir_move *entry, *dm;
3344
3345	dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3346	if (!dm)
3347		return -ENOMEM;
3348	dm->ino = ino;
3349	dm->rmdir_ino = 0;
3350	dm->rmdir_gen = 0;
3351	dm->orphanized = orphanized;
3352
3353	while (*p) {
3354		parent = *p;
3355		entry = rb_entry(parent, struct waiting_dir_move, node);
3356		if (ino < entry->ino) {
3357			p = &(*p)->rb_left;
3358		} else if (ino > entry->ino) {
3359			p = &(*p)->rb_right;
3360		} else {
3361			kfree(dm);
3362			return -EEXIST;
3363		}
3364	}
3365
3366	rb_link_node(&dm->node, parent, p);
3367	rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3368	return 0;
3369}
3370
3371static struct waiting_dir_move *
3372get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3373{
3374	struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3375	struct waiting_dir_move *entry;
3376
3377	while (n) {
3378		entry = rb_entry(n, struct waiting_dir_move, node);
3379		if (ino < entry->ino)
3380			n = n->rb_left;
3381		else if (ino > entry->ino)
3382			n = n->rb_right;
3383		else
3384			return entry;
3385	}
3386	return NULL;
3387}
3388
3389static void free_waiting_dir_move(struct send_ctx *sctx,
3390				  struct waiting_dir_move *dm)
3391{
3392	if (!dm)
3393		return;
3394	rb_erase(&dm->node, &sctx->waiting_dir_moves);
3395	kfree(dm);
3396}
3397
3398static int add_pending_dir_move(struct send_ctx *sctx,
3399				u64 ino,
3400				u64 ino_gen,
3401				u64 parent_ino,
3402				struct list_head *new_refs,
3403				struct list_head *deleted_refs,
3404				const bool is_orphan)
3405{
3406	struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3407	struct rb_node *parent = NULL;
3408	struct pending_dir_move *entry = NULL, *pm;
3409	struct recorded_ref *cur;
3410	int exists = 0;
3411	int ret;
3412
3413	pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3414	if (!pm)
3415		return -ENOMEM;
3416	pm->parent_ino = parent_ino;
3417	pm->ino = ino;
3418	pm->gen = ino_gen;
3419	INIT_LIST_HEAD(&pm->list);
3420	INIT_LIST_HEAD(&pm->update_refs);
3421	RB_CLEAR_NODE(&pm->node);
3422
3423	while (*p) {
3424		parent = *p;
3425		entry = rb_entry(parent, struct pending_dir_move, node);
3426		if (parent_ino < entry->parent_ino) {
3427			p = &(*p)->rb_left;
3428		} else if (parent_ino > entry->parent_ino) {
3429			p = &(*p)->rb_right;
3430		} else {
3431			exists = 1;
3432			break;
3433		}
3434	}
3435
3436	list_for_each_entry(cur, deleted_refs, list) {
3437		ret = dup_ref(cur, &pm->update_refs);
3438		if (ret < 0)
3439			goto out;
3440	}
3441	list_for_each_entry(cur, new_refs, list) {
3442		ret = dup_ref(cur, &pm->update_refs);
3443		if (ret < 0)
3444			goto out;
3445	}
3446
3447	ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3448	if (ret)
3449		goto out;
3450
3451	if (exists) {
3452		list_add_tail(&pm->list, &entry->list);
3453	} else {
3454		rb_link_node(&pm->node, parent, p);
3455		rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3456	}
3457	ret = 0;
3458out:
3459	if (ret) {
3460		__free_recorded_refs(&pm->update_refs);
3461		kfree(pm);
3462	}
3463	return ret;
3464}
3465
3466static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3467						      u64 parent_ino)
3468{
3469	struct rb_node *n = sctx->pending_dir_moves.rb_node;
3470	struct pending_dir_move *entry;
3471
3472	while (n) {
3473		entry = rb_entry(n, struct pending_dir_move, node);
3474		if (parent_ino < entry->parent_ino)
3475			n = n->rb_left;
3476		else if (parent_ino > entry->parent_ino)
3477			n = n->rb_right;
3478		else
3479			return entry;
3480	}
3481	return NULL;
3482}
3483
3484static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3485		     u64 ino, u64 gen, u64 *ancestor_ino)
3486{
3487	int ret = 0;
3488	u64 parent_inode = 0;
3489	u64 parent_gen = 0;
3490	u64 start_ino = ino;
3491
3492	*ancestor_ino = 0;
3493	while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3494		fs_path_reset(name);
3495
3496		if (is_waiting_for_rm(sctx, ino, gen))
3497			break;
3498		if (is_waiting_for_move(sctx, ino)) {
3499			if (*ancestor_ino == 0)
3500				*ancestor_ino = ino;
3501			ret = get_first_ref(sctx->parent_root, ino,
3502					    &parent_inode, &parent_gen, name);
3503		} else {
3504			ret = __get_cur_name_and_parent(sctx, ino, gen,
3505							&parent_inode,
3506							&parent_gen, name);
3507			if (ret > 0) {
3508				ret = 0;
3509				break;
3510			}
3511		}
3512		if (ret < 0)
3513			break;
3514		if (parent_inode == start_ino) {
3515			ret = 1;
3516			if (*ancestor_ino == 0)
3517				*ancestor_ino = ino;
3518			break;
3519		}
3520		ino = parent_inode;
3521		gen = parent_gen;
3522	}
3523	return ret;
3524}
3525
3526static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3527{
3528	struct fs_path *from_path = NULL;
3529	struct fs_path *to_path = NULL;
3530	struct fs_path *name = NULL;
3531	u64 orig_progress = sctx->send_progress;
3532	struct recorded_ref *cur;
3533	u64 parent_ino, parent_gen;
3534	struct waiting_dir_move *dm = NULL;
3535	u64 rmdir_ino = 0;
3536	u64 rmdir_gen;
3537	u64 ancestor;
3538	bool is_orphan;
3539	int ret;
3540
3541	name = fs_path_alloc();
3542	from_path = fs_path_alloc();
3543	if (!name || !from_path) {
3544		ret = -ENOMEM;
3545		goto out;
3546	}
3547
3548	dm = get_waiting_dir_move(sctx, pm->ino);
3549	ASSERT(dm);
3550	rmdir_ino = dm->rmdir_ino;
3551	rmdir_gen = dm->rmdir_gen;
3552	is_orphan = dm->orphanized;
3553	free_waiting_dir_move(sctx, dm);
3554
3555	if (is_orphan) {
3556		ret = gen_unique_name(sctx, pm->ino,
3557				      pm->gen, from_path);
3558	} else {
3559		ret = get_first_ref(sctx->parent_root, pm->ino,
3560				    &parent_ino, &parent_gen, name);
3561		if (ret < 0)
3562			goto out;
3563		ret = get_cur_path(sctx, parent_ino, parent_gen,
3564				   from_path);
3565		if (ret < 0)
3566			goto out;
3567		ret = fs_path_add_path(from_path, name);
3568	}
3569	if (ret < 0)
3570		goto out;
3571
3572	sctx->send_progress = sctx->cur_ino + 1;
3573	ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3574	if (ret < 0)
3575		goto out;
3576	if (ret) {
3577		LIST_HEAD(deleted_refs);
3578		ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3579		ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3580					   &pm->update_refs, &deleted_refs,
3581					   is_orphan);
3582		if (ret < 0)
3583			goto out;
3584		if (rmdir_ino) {
3585			dm = get_waiting_dir_move(sctx, pm->ino);
3586			ASSERT(dm);
3587			dm->rmdir_ino = rmdir_ino;
3588			dm->rmdir_gen = rmdir_gen;
3589		}
3590		goto out;
3591	}
3592	fs_path_reset(name);
3593	to_path = name;
3594	name = NULL;
3595	ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3596	if (ret < 0)
3597		goto out;
3598
3599	ret = send_rename(sctx, from_path, to_path);
3600	if (ret < 0)
3601		goto out;
3602
3603	if (rmdir_ino) {
3604		struct orphan_dir_info *odi;
3605		u64 gen;
3606
3607		odi = get_orphan_dir_info(sctx, rmdir_ino, rmdir_gen);
3608		if (!odi) {
3609			/* already deleted */
3610			goto finish;
3611		}
3612		gen = odi->gen;
3613
3614		ret = can_rmdir(sctx, rmdir_ino, gen);
3615		if (ret < 0)
3616			goto out;
3617		if (!ret)
3618			goto finish;
3619
3620		name = fs_path_alloc();
3621		if (!name) {
3622			ret = -ENOMEM;
3623			goto out;
3624		}
3625		ret = get_cur_path(sctx, rmdir_ino, gen, name);
3626		if (ret < 0)
3627			goto out;
3628		ret = send_rmdir(sctx, name);
3629		if (ret < 0)
3630			goto out;
3631	}
3632
3633finish:
3634	ret = cache_dir_utimes(sctx, pm->ino, pm->gen);
3635	if (ret < 0)
3636		goto out;
3637
3638	/*
3639	 * After rename/move, need to update the utimes of both new parent(s)
3640	 * and old parent(s).
3641	 */
3642	list_for_each_entry(cur, &pm->update_refs, list) {
3643		/*
3644		 * The parent inode might have been deleted in the send snapshot
3645		 */
3646		ret = get_inode_info(sctx->send_root, cur->dir, NULL);
3647		if (ret == -ENOENT) {
3648			ret = 0;
3649			continue;
3650		}
3651		if (ret < 0)
3652			goto out;
3653
3654		ret = cache_dir_utimes(sctx, cur->dir, cur->dir_gen);
3655		if (ret < 0)
3656			goto out;
3657	}
3658
3659out:
3660	fs_path_free(name);
3661	fs_path_free(from_path);
3662	fs_path_free(to_path);
3663	sctx->send_progress = orig_progress;
3664
3665	return ret;
3666}
3667
3668static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3669{
3670	if (!list_empty(&m->list))
3671		list_del(&m->list);
3672	if (!RB_EMPTY_NODE(&m->node))
3673		rb_erase(&m->node, &sctx->pending_dir_moves);
3674	__free_recorded_refs(&m->update_refs);
3675	kfree(m);
3676}
3677
3678static void tail_append_pending_moves(struct send_ctx *sctx,
3679				      struct pending_dir_move *moves,
3680				      struct list_head *stack)
3681{
3682	if (list_empty(&moves->list)) {
3683		list_add_tail(&moves->list, stack);
3684	} else {
3685		LIST_HEAD(list);
3686		list_splice_init(&moves->list, &list);
3687		list_add_tail(&moves->list, stack);
3688		list_splice_tail(&list, stack);
3689	}
3690	if (!RB_EMPTY_NODE(&moves->node)) {
3691		rb_erase(&moves->node, &sctx->pending_dir_moves);
3692		RB_CLEAR_NODE(&moves->node);
3693	}
3694}
3695
3696static int apply_children_dir_moves(struct send_ctx *sctx)
3697{
3698	struct pending_dir_move *pm;
3699	LIST_HEAD(stack);
3700	u64 parent_ino = sctx->cur_ino;
3701	int ret = 0;
3702
3703	pm = get_pending_dir_moves(sctx, parent_ino);
3704	if (!pm)
3705		return 0;
3706
3707	tail_append_pending_moves(sctx, pm, &stack);
3708
3709	while (!list_empty(&stack)) {
3710		pm = list_first_entry(&stack, struct pending_dir_move, list);
3711		parent_ino = pm->ino;
3712		ret = apply_dir_move(sctx, pm);
3713		free_pending_move(sctx, pm);
3714		if (ret)
3715			goto out;
3716		pm = get_pending_dir_moves(sctx, parent_ino);
3717		if (pm)
3718			tail_append_pending_moves(sctx, pm, &stack);
3719	}
3720	return 0;
3721
3722out:
3723	while (!list_empty(&stack)) {
3724		pm = list_first_entry(&stack, struct pending_dir_move, list);
3725		free_pending_move(sctx, pm);
3726	}
3727	return ret;
3728}
3729
3730/*
3731 * We might need to delay a directory rename even when no ancestor directory
3732 * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3733 * renamed. This happens when we rename a directory to the old name (the name
3734 * in the parent root) of some other unrelated directory that got its rename
3735 * delayed due to some ancestor with higher number that got renamed.
3736 *
3737 * Example:
3738 *
3739 * Parent snapshot:
3740 * .                                       (ino 256)
3741 * |---- a/                                (ino 257)
3742 * |     |---- file                        (ino 260)
3743 * |
3744 * |---- b/                                (ino 258)
3745 * |---- c/                                (ino 259)
3746 *
3747 * Send snapshot:
3748 * .                                       (ino 256)
3749 * |---- a/                                (ino 258)
3750 * |---- x/                                (ino 259)
3751 *       |---- y/                          (ino 257)
3752 *             |----- file                 (ino 260)
3753 *
3754 * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3755 * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3756 * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3757 * must issue is:
3758 *
3759 * 1 - rename 259 from 'c' to 'x'
3760 * 2 - rename 257 from 'a' to 'x/y'
3761 * 3 - rename 258 from 'b' to 'a'
3762 *
3763 * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3764 * be done right away and < 0 on error.
3765 */
3766static int wait_for_dest_dir_move(struct send_ctx *sctx,
3767				  struct recorded_ref *parent_ref,
3768				  const bool is_orphan)
3769{
3770	struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3771	struct btrfs_path *path;
3772	struct btrfs_key key;
3773	struct btrfs_key di_key;
3774	struct btrfs_dir_item *di;
3775	u64 left_gen;
3776	u64 right_gen;
3777	int ret = 0;
3778	struct waiting_dir_move *wdm;
3779
3780	if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3781		return 0;
3782
3783	path = alloc_path_for_send();
3784	if (!path)
3785		return -ENOMEM;
3786
3787	key.objectid = parent_ref->dir;
3788	key.type = BTRFS_DIR_ITEM_KEY;
3789	key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3790
3791	ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3792	if (ret < 0) {
3793		goto out;
3794	} else if (ret > 0) {
3795		ret = 0;
3796		goto out;
3797	}
3798
3799	di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3800				       parent_ref->name_len);
3801	if (!di) {
3802		ret = 0;
3803		goto out;
3804	}
3805	/*
3806	 * di_key.objectid has the number of the inode that has a dentry in the
3807	 * parent directory with the same name that sctx->cur_ino is being
3808	 * renamed to. We need to check if that inode is in the send root as
3809	 * well and if it is currently marked as an inode with a pending rename,
3810	 * if it is, we need to delay the rename of sctx->cur_ino as well, so
3811	 * that it happens after that other inode is renamed.
3812	 */
3813	btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3814	if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3815		ret = 0;
3816		goto out;
3817	}
3818
3819	ret = get_inode_gen(sctx->parent_root, di_key.objectid, &left_gen);
3820	if (ret < 0)
3821		goto out;
3822	ret = get_inode_gen(sctx->send_root, di_key.objectid, &right_gen);
3823	if (ret < 0) {
3824		if (ret == -ENOENT)
3825			ret = 0;
3826		goto out;
3827	}
3828
3829	/* Different inode, no need to delay the rename of sctx->cur_ino */
3830	if (right_gen != left_gen) {
3831		ret = 0;
3832		goto out;
3833	}
3834
3835	wdm = get_waiting_dir_move(sctx, di_key.objectid);
3836	if (wdm && !wdm->orphanized) {
3837		ret = add_pending_dir_move(sctx,
3838					   sctx->cur_ino,
3839					   sctx->cur_inode_gen,
3840					   di_key.objectid,
3841					   &sctx->new_refs,
3842					   &sctx->deleted_refs,
3843					   is_orphan);
3844		if (!ret)
3845			ret = 1;
3846	}
3847out:
3848	btrfs_free_path(path);
3849	return ret;
3850}
3851
3852/*
3853 * Check if inode ino2, or any of its ancestors, is inode ino1.
3854 * Return 1 if true, 0 if false and < 0 on error.
3855 */
3856static int check_ino_in_path(struct btrfs_root *root,
3857			     const u64 ino1,
3858			     const u64 ino1_gen,
3859			     const u64 ino2,
3860			     const u64 ino2_gen,
3861			     struct fs_path *fs_path)
3862{
3863	u64 ino = ino2;
3864
3865	if (ino1 == ino2)
3866		return ino1_gen == ino2_gen;
3867
3868	while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3869		u64 parent;
3870		u64 parent_gen;
3871		int ret;
3872
3873		fs_path_reset(fs_path);
3874		ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3875		if (ret < 0)
3876			return ret;
3877		if (parent == ino1)
3878			return parent_gen == ino1_gen;
3879		ino = parent;
3880	}
3881	return 0;
3882}
3883
3884/*
3885 * Check if inode ino1 is an ancestor of inode ino2 in the given root for any
3886 * possible path (in case ino2 is not a directory and has multiple hard links).
3887 * Return 1 if true, 0 if false and < 0 on error.
3888 */
3889static int is_ancestor(struct btrfs_root *root,
3890		       const u64 ino1,
3891		       const u64 ino1_gen,
3892		       const u64 ino2,
3893		       struct fs_path *fs_path)
3894{
3895	bool free_fs_path = false;
3896	int ret = 0;
3897	int iter_ret = 0;
3898	struct btrfs_path *path = NULL;
3899	struct btrfs_key key;
3900
3901	if (!fs_path) {
3902		fs_path = fs_path_alloc();
3903		if (!fs_path)
3904			return -ENOMEM;
3905		free_fs_path = true;
3906	}
3907
3908	path = alloc_path_for_send();
3909	if (!path) {
3910		ret = -ENOMEM;
3911		goto out;
3912	}
3913
3914	key.objectid = ino2;
3915	key.type = BTRFS_INODE_REF_KEY;
3916	key.offset = 0;
3917
3918	btrfs_for_each_slot(root, &key, &key, path, iter_ret) {
3919		struct extent_buffer *leaf = path->nodes[0];
3920		int slot = path->slots[0];
3921		u32 cur_offset = 0;
3922		u32 item_size;
3923
3924		if (key.objectid != ino2)
3925			break;
3926		if (key.type != BTRFS_INODE_REF_KEY &&
3927		    key.type != BTRFS_INODE_EXTREF_KEY)
3928			break;
3929
3930		item_size = btrfs_item_size(leaf, slot);
3931		while (cur_offset < item_size) {
3932			u64 parent;
3933			u64 parent_gen;
3934
3935			if (key.type == BTRFS_INODE_EXTREF_KEY) {
3936				unsigned long ptr;
3937				struct btrfs_inode_extref *extref;
3938
3939				ptr = btrfs_item_ptr_offset(leaf, slot);
3940				extref = (struct btrfs_inode_extref *)
3941					(ptr + cur_offset);
3942				parent = btrfs_inode_extref_parent(leaf,
3943								   extref);
3944				cur_offset += sizeof(*extref);
3945				cur_offset += btrfs_inode_extref_name_len(leaf,
3946								  extref);
3947			} else {
3948				parent = key.offset;
3949				cur_offset = item_size;
3950			}
3951
3952			ret = get_inode_gen(root, parent, &parent_gen);
3953			if (ret < 0)
3954				goto out;
3955			ret = check_ino_in_path(root, ino1, ino1_gen,
3956						parent, parent_gen, fs_path);
3957			if (ret)
3958				goto out;
3959		}
3960	}
3961	ret = 0;
3962	if (iter_ret < 0)
3963		ret = iter_ret;
3964
3965out:
3966	btrfs_free_path(path);
3967	if (free_fs_path)
3968		fs_path_free(fs_path);
3969	return ret;
3970}
3971
3972static int wait_for_parent_move(struct send_ctx *sctx,
3973				struct recorded_ref *parent_ref,
3974				const bool is_orphan)
3975{
3976	int ret = 0;
3977	u64 ino = parent_ref->dir;
3978	u64 ino_gen = parent_ref->dir_gen;
3979	u64 parent_ino_before, parent_ino_after;
3980	struct fs_path *path_before = NULL;
3981	struct fs_path *path_after = NULL;
3982	int len1, len2;
3983
3984	path_after = fs_path_alloc();
3985	path_before = fs_path_alloc();
3986	if (!path_after || !path_before) {
3987		ret = -ENOMEM;
3988		goto out;
3989	}
3990
3991	/*
3992	 * Our current directory inode may not yet be renamed/moved because some
3993	 * ancestor (immediate or not) has to be renamed/moved first. So find if
3994	 * such ancestor exists and make sure our own rename/move happens after
3995	 * that ancestor is processed to avoid path build infinite loops (done
3996	 * at get_cur_path()).
3997	 */
3998	while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3999		u64 parent_ino_after_gen;
4000
4001		if (is_waiting_for_move(sctx, ino)) {
4002			/*
4003			 * If the current inode is an ancestor of ino in the
4004			 * parent root, we need to delay the rename of the
4005			 * current inode, otherwise don't delayed the rename
4006			 * because we can end up with a circular dependency
4007			 * of renames, resulting in some directories never
4008			 * getting the respective rename operations issued in
4009			 * the send stream or getting into infinite path build
4010			 * loops.
4011			 */
4012			ret = is_ancestor(sctx->parent_root,
4013					  sctx->cur_ino, sctx->cur_inode_gen,
4014					  ino, path_before);
4015			if (ret)
4016				break;
4017		}
4018
4019		fs_path_reset(path_before);
4020		fs_path_reset(path_after);
4021
4022		ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
4023				    &parent_ino_after_gen, path_after);
4024		if (ret < 0)
4025			goto out;
4026		ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
4027				    NULL, path_before);
4028		if (ret < 0 && ret != -ENOENT) {
4029			goto out;
4030		} else if (ret == -ENOENT) {
4031			ret = 0;
4032			break;
4033		}
4034
4035		len1 = fs_path_len(path_before);
4036		len2 = fs_path_len(path_after);
4037		if (ino > sctx->cur_ino &&
4038		    (parent_ino_before != parent_ino_after || len1 != len2 ||
4039		     memcmp(path_before->start, path_after->start, len1))) {
4040			u64 parent_ino_gen;
4041
4042			ret = get_inode_gen(sctx->parent_root, ino, &parent_ino_gen);
4043			if (ret < 0)
4044				goto out;
4045			if (ino_gen == parent_ino_gen) {
4046				ret = 1;
4047				break;
4048			}
4049		}
4050		ino = parent_ino_after;
4051		ino_gen = parent_ino_after_gen;
4052	}
4053
4054out:
4055	fs_path_free(path_before);
4056	fs_path_free(path_after);
4057
4058	if (ret == 1) {
4059		ret = add_pending_dir_move(sctx,
4060					   sctx->cur_ino,
4061					   sctx->cur_inode_gen,
4062					   ino,
4063					   &sctx->new_refs,
4064					   &sctx->deleted_refs,
4065					   is_orphan);
4066		if (!ret)
4067			ret = 1;
4068	}
4069
4070	return ret;
4071}
4072
4073static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
4074{
4075	int ret;
4076	struct fs_path *new_path;
4077
4078	/*
4079	 * Our reference's name member points to its full_path member string, so
4080	 * we use here a new path.
4081	 */
4082	new_path = fs_path_alloc();
4083	if (!new_path)
4084		return -ENOMEM;
4085
4086	ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path);
4087	if (ret < 0) {
4088		fs_path_free(new_path);
4089		return ret;
4090	}
4091	ret = fs_path_add(new_path, ref->name, ref->name_len);
4092	if (ret < 0) {
4093		fs_path_free(new_path);
4094		return ret;
4095	}
4096
4097	fs_path_free(ref->full_path);
4098	set_ref_path(ref, new_path);
4099
4100	return 0;
4101}
4102
4103/*
4104 * When processing the new references for an inode we may orphanize an existing
4105 * directory inode because its old name conflicts with one of the new references
4106 * of the current inode. Later, when processing another new reference of our
4107 * inode, we might need to orphanize another inode, but the path we have in the
4108 * reference reflects the pre-orphanization name of the directory we previously
4109 * orphanized. For example:
4110 *
4111 * parent snapshot looks like:
4112 *
4113 * .                                     (ino 256)
4114 * |----- f1                             (ino 257)
4115 * |----- f2                             (ino 258)
4116 * |----- d1/                            (ino 259)
4117 *        |----- d2/                     (ino 260)
4118 *
4119 * send snapshot looks like:
4120 *
4121 * .                                     (ino 256)
4122 * |----- d1                             (ino 258)
4123 * |----- f2/                            (ino 259)
4124 *        |----- f2_link/                (ino 260)
4125 *        |       |----- f1              (ino 257)
4126 *        |
4127 *        |----- d2                      (ino 258)
4128 *
4129 * When processing inode 257 we compute the name for inode 259 as "d1", and we
4130 * cache it in the name cache. Later when we start processing inode 258, when
4131 * collecting all its new references we set a full path of "d1/d2" for its new
4132 * reference with name "d2". When we start processing the new references we
4133 * start by processing the new reference with name "d1", and this results in
4134 * orphanizing inode 259, since its old reference causes a conflict. Then we
4135 * move on the next new reference, with name "d2", and we find out we must
4136 * orphanize inode 260, as its old reference conflicts with ours - but for the
4137 * orphanization we use a source path corresponding to the path we stored in the
4138 * new reference, which is "d1/d2" and not "o259-6-0/d2" - this makes the
4139 * receiver fail since the path component "d1/" no longer exists, it was renamed
4140 * to "o259-6-0/" when processing the previous new reference. So in this case we
4141 * must recompute the path in the new reference and use it for the new
4142 * orphanization operation.
4143 */
4144static int refresh_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
4145{
4146	char *name;
4147	int ret;
4148
4149	name = kmemdup(ref->name, ref->name_len, GFP_KERNEL);
4150	if (!name)
4151		return -ENOMEM;
4152
4153	fs_path_reset(ref->full_path);
4154	ret = get_cur_path(sctx, ref->dir, ref->dir_gen, ref->full_path);
4155	if (ret < 0)
4156		goto out;
4157
4158	ret = fs_path_add(ref->full_path, name, ref->name_len);
4159	if (ret < 0)
4160		goto out;
4161
4162	/* Update the reference's base name pointer. */
4163	set_ref_path(ref, ref->full_path);
4164out:
4165	kfree(name);
4166	return ret;
4167}
4168
4169/*
4170 * This does all the move/link/unlink/rmdir magic.
4171 */
4172static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
4173{
4174	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
4175	int ret = 0;
4176	struct recorded_ref *cur;
4177	struct recorded_ref *cur2;
4178	LIST_HEAD(check_dirs);
4179	struct fs_path *valid_path = NULL;
4180	u64 ow_inode = 0;
4181	u64 ow_gen;
4182	u64 ow_mode;
4183	int did_overwrite = 0;
4184	int is_orphan = 0;
4185	u64 last_dir_ino_rm = 0;
4186	bool can_rename = true;
4187	bool orphanized_dir = false;
4188	bool orphanized_ancestor = false;
4189
4190	btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
4191
4192	/*
4193	 * This should never happen as the root dir always has the same ref
4194	 * which is always '..'
4195	 */
4196	if (unlikely(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID)) {
4197		btrfs_err(fs_info,
4198			  "send: unexpected inode %llu in process_recorded_refs()",
4199			  sctx->cur_ino);
4200		ret = -EINVAL;
4201		goto out;
4202	}
4203
4204	valid_path = fs_path_alloc();
4205	if (!valid_path) {
4206		ret = -ENOMEM;
4207		goto out;
4208	}
4209
4210	/*
4211	 * First, check if the first ref of the current inode was overwritten
4212	 * before. If yes, we know that the current inode was already orphanized
4213	 * and thus use the orphan name. If not, we can use get_cur_path to
4214	 * get the path of the first ref as it would like while receiving at
4215	 * this point in time.
4216	 * New inodes are always orphan at the beginning, so force to use the
4217	 * orphan name in this case.
4218	 * The first ref is stored in valid_path and will be updated if it
4219	 * gets moved around.
4220	 */
4221	if (!sctx->cur_inode_new) {
4222		ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
4223				sctx->cur_inode_gen);
4224		if (ret < 0)
4225			goto out;
4226		if (ret)
4227			did_overwrite = 1;
4228	}
4229	if (sctx->cur_inode_new || did_overwrite) {
4230		ret = gen_unique_name(sctx, sctx->cur_ino,
4231				sctx->cur_inode_gen, valid_path);
4232		if (ret < 0)
4233			goto out;
4234		is_orphan = 1;
4235	} else {
4236		ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
4237				valid_path);
4238		if (ret < 0)
4239			goto out;
4240	}
4241
4242	/*
4243	 * Before doing any rename and link operations, do a first pass on the
4244	 * new references to orphanize any unprocessed inodes that may have a
4245	 * reference that conflicts with one of the new references of the current
4246	 * inode. This needs to happen first because a new reference may conflict
4247	 * with the old reference of a parent directory, so we must make sure
4248	 * that the path used for link and rename commands don't use an
4249	 * orphanized name when an ancestor was not yet orphanized.
4250	 *
4251	 * Example:
4252	 *
4253	 * Parent snapshot:
4254	 *
4255	 * .                                                      (ino 256)
4256	 * |----- testdir/                                        (ino 259)
4257	 * |          |----- a                                    (ino 257)
4258	 * |
4259	 * |----- b                                               (ino 258)
4260	 *
4261	 * Send snapshot:
4262	 *
4263	 * .                                                      (ino 256)
4264	 * |----- testdir_2/                                      (ino 259)
4265	 * |          |----- a                                    (ino 260)
4266	 * |
4267	 * |----- testdir                                         (ino 257)
4268	 * |----- b                                               (ino 257)
4269	 * |----- b2                                              (ino 258)
4270	 *
4271	 * Processing the new reference for inode 257 with name "b" may happen
4272	 * before processing the new reference with name "testdir". If so, we
4273	 * must make sure that by the time we send a link command to create the
4274	 * hard link "b", inode 259 was already orphanized, since the generated
4275	 * path in "valid_path" already contains the orphanized name for 259.
4276	 * We are processing inode 257, so only later when processing 259 we do
4277	 * the rename operation to change its temporary (orphanized) name to
4278	 * "testdir_2".
4279	 */
4280	list_for_each_entry(cur, &sctx->new_refs, list) {
4281		ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4282		if (ret < 0)
4283			goto out;
4284		if (ret == inode_state_will_create)
4285			continue;
4286
4287		/*
4288		 * Check if this new ref would overwrite the first ref of another
4289		 * unprocessed inode. If yes, orphanize the overwritten inode.
4290		 * If we find an overwritten ref that is not the first ref,
4291		 * simply unlink it.
4292		 */
4293		ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4294				cur->name, cur->name_len,
4295				&ow_inode, &ow_gen, &ow_mode);
4296		if (ret < 0)
4297			goto out;
4298		if (ret) {
4299			ret = is_first_ref(sctx->parent_root,
4300					   ow_inode, cur->dir, cur->name,
4301					   cur->name_len);
4302			if (ret < 0)
4303				goto out;
4304			if (ret) {
4305				struct name_cache_entry *nce;
4306				struct waiting_dir_move *wdm;
4307
4308				if (orphanized_dir) {
4309					ret = refresh_ref_path(sctx, cur);
4310					if (ret < 0)
4311						goto out;
4312				}
4313
4314				ret = orphanize_inode(sctx, ow_inode, ow_gen,
4315						cur->full_path);
4316				if (ret < 0)
4317					goto out;
4318				if (S_ISDIR(ow_mode))
4319					orphanized_dir = true;
4320
4321				/*
4322				 * If ow_inode has its rename operation delayed
4323				 * make sure that its orphanized name is used in
4324				 * the source path when performing its rename
4325				 * operation.
4326				 */
4327				wdm = get_waiting_dir_move(sctx, ow_inode);
4328				if (wdm)
4329					wdm->orphanized = true;
4330
4331				/*
4332				 * Make sure we clear our orphanized inode's
4333				 * name from the name cache. This is because the
4334				 * inode ow_inode might be an ancestor of some
4335				 * other inode that will be orphanized as well
4336				 * later and has an inode number greater than
4337				 * sctx->send_progress. We need to prevent
4338				 * future name lookups from using the old name
4339				 * and get instead the orphan name.
4340				 */
4341				nce = name_cache_search(sctx, ow_inode, ow_gen);
4342				if (nce)
4343					btrfs_lru_cache_remove(&sctx->name_cache,
4344							       &nce->entry);
4345
4346				/*
4347				 * ow_inode might currently be an ancestor of
4348				 * cur_ino, therefore compute valid_path (the
4349				 * current path of cur_ino) again because it
4350				 * might contain the pre-orphanization name of
4351				 * ow_inode, which is no longer valid.
4352				 */
4353				ret = is_ancestor(sctx->parent_root,
4354						  ow_inode, ow_gen,
4355						  sctx->cur_ino, NULL);
4356				if (ret > 0) {
4357					orphanized_ancestor = true;
4358					fs_path_reset(valid_path);
4359					ret = get_cur_path(sctx, sctx->cur_ino,
4360							   sctx->cur_inode_gen,
4361							   valid_path);
4362				}
4363				if (ret < 0)
4364					goto out;
4365			} else {
4366				/*
4367				 * If we previously orphanized a directory that
4368				 * collided with a new reference that we already
4369				 * processed, recompute the current path because
4370				 * that directory may be part of the path.
4371				 */
4372				if (orphanized_dir) {
4373					ret = refresh_ref_path(sctx, cur);
4374					if (ret < 0)
4375						goto out;
4376				}
4377				ret = send_unlink(sctx, cur->full_path);
4378				if (ret < 0)
4379					goto out;
4380			}
4381		}
4382
4383	}
4384
4385	list_for_each_entry(cur, &sctx->new_refs, list) {
4386		/*
4387		 * We may have refs where the parent directory does not exist
4388		 * yet. This happens if the parent directories inum is higher
4389		 * than the current inum. To handle this case, we create the
4390		 * parent directory out of order. But we need to check if this
4391		 * did already happen before due to other refs in the same dir.
4392		 */
4393		ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4394		if (ret < 0)
4395			goto out;
4396		if (ret == inode_state_will_create) {
4397			ret = 0;
4398			/*
4399			 * First check if any of the current inodes refs did
4400			 * already create the dir.
4401			 */
4402			list_for_each_entry(cur2, &sctx->new_refs, list) {
4403				if (cur == cur2)
4404					break;
4405				if (cur2->dir == cur->dir) {
4406					ret = 1;
4407					break;
4408				}
4409			}
4410
4411			/*
4412			 * If that did not happen, check if a previous inode
4413			 * did already create the dir.
4414			 */
4415			if (!ret)
4416				ret = did_create_dir(sctx, cur->dir);
4417			if (ret < 0)
4418				goto out;
4419			if (!ret) {
4420				ret = send_create_inode(sctx, cur->dir);
4421				if (ret < 0)
4422					goto out;
4423				cache_dir_created(sctx, cur->dir);
4424			}
4425		}
4426
4427		if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
4428			ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
4429			if (ret < 0)
4430				goto out;
4431			if (ret == 1) {
4432				can_rename = false;
4433				*pending_move = 1;
4434			}
4435		}
4436
4437		if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
4438		    can_rename) {
4439			ret = wait_for_parent_move(sctx, cur, is_orphan);
4440			if (ret < 0)
4441				goto out;
4442			if (ret == 1) {
4443				can_rename = false;
4444				*pending_move = 1;
4445			}
4446		}
4447
4448		/*
4449		 * link/move the ref to the new place. If we have an orphan
4450		 * inode, move it and update valid_path. If not, link or move
4451		 * it depending on the inode mode.
4452		 */
4453		if (is_orphan && can_rename) {
4454			ret = send_rename(sctx, valid_path, cur->full_path);
4455			if (ret < 0)
4456				goto out;
4457			is_orphan = 0;
4458			ret = fs_path_copy(valid_path, cur->full_path);
4459			if (ret < 0)
4460				goto out;
4461		} else if (can_rename) {
4462			if (S_ISDIR(sctx->cur_inode_mode)) {
4463				/*
4464				 * Dirs can't be linked, so move it. For moved
4465				 * dirs, we always have one new and one deleted
4466				 * ref. The deleted ref is ignored later.
4467				 */
4468				ret = send_rename(sctx, valid_path,
4469						  cur->full_path);
4470				if (!ret)
4471					ret = fs_path_copy(valid_path,
4472							   cur->full_path);
4473				if (ret < 0)
4474					goto out;
4475			} else {
4476				/*
4477				 * We might have previously orphanized an inode
4478				 * which is an ancestor of our current inode,
4479				 * so our reference's full path, which was
4480				 * computed before any such orphanizations, must
4481				 * be updated.
4482				 */
4483				if (orphanized_dir) {
4484					ret = update_ref_path(sctx, cur);
4485					if (ret < 0)
4486						goto out;
4487				}
4488				ret = send_link(sctx, cur->full_path,
4489						valid_path);
4490				if (ret < 0)
4491					goto out;
4492			}
4493		}
4494		ret = dup_ref(cur, &check_dirs);
4495		if (ret < 0)
4496			goto out;
4497	}
4498
4499	if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
4500		/*
4501		 * Check if we can already rmdir the directory. If not,
4502		 * orphanize it. For every dir item inside that gets deleted
4503		 * later, we do this check again and rmdir it then if possible.
4504		 * See the use of check_dirs for more details.
4505		 */
4506		ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen);
4507		if (ret < 0)
4508			goto out;
4509		if (ret) {
4510			ret = send_rmdir(sctx, valid_path);
4511			if (ret < 0)
4512				goto out;
4513		} else if (!is_orphan) {
4514			ret = orphanize_inode(sctx, sctx->cur_ino,
4515					sctx->cur_inode_gen, valid_path);
4516			if (ret < 0)
4517				goto out;
4518			is_orphan = 1;
4519		}
4520
4521		list_for_each_entry(cur, &sctx->deleted_refs, list) {
4522			ret = dup_ref(cur, &check_dirs);
4523			if (ret < 0)
4524				goto out;
4525		}
4526	} else if (S_ISDIR(sctx->cur_inode_mode) &&
4527		   !list_empty(&sctx->deleted_refs)) {
4528		/*
4529		 * We have a moved dir. Add the old parent to check_dirs
4530		 */
4531		cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
4532				list);
4533		ret = dup_ref(cur, &check_dirs);
4534		if (ret < 0)
4535			goto out;
4536	} else if (!S_ISDIR(sctx->cur_inode_mode)) {
4537		/*
4538		 * We have a non dir inode. Go through all deleted refs and
4539		 * unlink them if they were not already overwritten by other
4540		 * inodes.
4541		 */
4542		list_for_each_entry(cur, &sctx->deleted_refs, list) {
4543			ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4544					sctx->cur_ino, sctx->cur_inode_gen,
4545					cur->name, cur->name_len);
4546			if (ret < 0)
4547				goto out;
4548			if (!ret) {
4549				/*
4550				 * If we orphanized any ancestor before, we need
4551				 * to recompute the full path for deleted names,
4552				 * since any such path was computed before we
4553				 * processed any references and orphanized any
4554				 * ancestor inode.
4555				 */
4556				if (orphanized_ancestor) {
4557					ret = update_ref_path(sctx, cur);
4558					if (ret < 0)
4559						goto out;
4560				}
4561				ret = send_unlink(sctx, cur->full_path);
4562				if (ret < 0)
4563					goto out;
4564			}
4565			ret = dup_ref(cur, &check_dirs);
4566			if (ret < 0)
4567				goto out;
4568		}
4569		/*
4570		 * If the inode is still orphan, unlink the orphan. This may
4571		 * happen when a previous inode did overwrite the first ref
4572		 * of this inode and no new refs were added for the current
4573		 * inode. Unlinking does not mean that the inode is deleted in
4574		 * all cases. There may still be links to this inode in other
4575		 * places.
4576		 */
4577		if (is_orphan) {
4578			ret = send_unlink(sctx, valid_path);
4579			if (ret < 0)
4580				goto out;
4581		}
4582	}
4583
4584	/*
4585	 * We did collect all parent dirs where cur_inode was once located. We
4586	 * now go through all these dirs and check if they are pending for
4587	 * deletion and if it's finally possible to perform the rmdir now.
4588	 * We also update the inode stats of the parent dirs here.
4589	 */
4590	list_for_each_entry(cur, &check_dirs, list) {
4591		/*
4592		 * In case we had refs into dirs that were not processed yet,
4593		 * we don't need to do the utime and rmdir logic for these dirs.
4594		 * The dir will be processed later.
4595		 */
4596		if (cur->dir > sctx->cur_ino)
4597			continue;
4598
4599		ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4600		if (ret < 0)
4601			goto out;
4602
4603		if (ret == inode_state_did_create ||
4604		    ret == inode_state_no_change) {
4605			ret = cache_dir_utimes(sctx, cur->dir, cur->dir_gen);
4606			if (ret < 0)
4607				goto out;
4608		} else if (ret == inode_state_did_delete &&
4609			   cur->dir != last_dir_ino_rm) {
4610			ret = can_rmdir(sctx, cur->dir, cur->dir_gen);
4611			if (ret < 0)
4612				goto out;
4613			if (ret) {
4614				ret = get_cur_path(sctx, cur->dir,
4615						   cur->dir_gen, valid_path);
4616				if (ret < 0)
4617					goto out;
4618				ret = send_rmdir(sctx, valid_path);
4619				if (ret < 0)
4620					goto out;
4621				last_dir_ino_rm = cur->dir;
4622			}
4623		}
4624	}
4625
4626	ret = 0;
4627
4628out:
4629	__free_recorded_refs(&check_dirs);
4630	free_recorded_refs(sctx);
4631	fs_path_free(valid_path);
4632	return ret;
4633}
4634
4635static int rbtree_ref_comp(const void *k, const struct rb_node *node)
4636{
4637	const struct recorded_ref *data = k;
4638	const struct recorded_ref *ref = rb_entry(node, struct recorded_ref, node);
4639	int result;
4640
4641	if (data->dir > ref->dir)
4642		return 1;
4643	if (data->dir < ref->dir)
4644		return -1;
4645	if (data->dir_gen > ref->dir_gen)
4646		return 1;
4647	if (data->dir_gen < ref->dir_gen)
4648		return -1;
4649	if (data->name_len > ref->name_len)
4650		return 1;
4651	if (data->name_len < ref->name_len)
4652		return -1;
4653	result = strcmp(data->name, ref->name);
4654	if (result > 0)
4655		return 1;
4656	if (result < 0)
4657		return -1;
4658	return 0;
4659}
4660
4661static bool rbtree_ref_less(struct rb_node *node, const struct rb_node *parent)
4662{
4663	const struct recorded_ref *entry = rb_entry(node, struct recorded_ref, node);
4664
4665	return rbtree_ref_comp(entry, parent) < 0;
4666}
4667
4668static int record_ref_in_tree(struct rb_root *root, struct list_head *refs,
4669			      struct fs_path *name, u64 dir, u64 dir_gen,
4670			      struct send_ctx *sctx)
4671{
4672	int ret = 0;
4673	struct fs_path *path = NULL;
4674	struct recorded_ref *ref = NULL;
4675
4676	path = fs_path_alloc();
4677	if (!path) {
4678		ret = -ENOMEM;
4679		goto out;
4680	}
4681
4682	ref = recorded_ref_alloc();
4683	if (!ref) {
4684		ret = -ENOMEM;
4685		goto out;
4686	}
4687
4688	ret = get_cur_path(sctx, dir, dir_gen, path);
4689	if (ret < 0)
4690		goto out;
4691	ret = fs_path_add_path(path, name);
4692	if (ret < 0)
4693		goto out;
4694
4695	ref->dir = dir;
4696	ref->dir_gen = dir_gen;
4697	set_ref_path(ref, path);
4698	list_add_tail(&ref->list, refs);
4699	rb_add(&ref->node, root, rbtree_ref_less);
4700	ref->root = root;
4701out:
4702	if (ret) {
4703		if (path && (!ref || !ref->full_path))
4704			fs_path_free(path);
4705		recorded_ref_free(ref);
4706	}
4707	return ret;
4708}
4709
4710static int record_new_ref_if_needed(int num, u64 dir, int index,
4711				    struct fs_path *name, void *ctx)
4712{
4713	int ret = 0;
4714	struct send_ctx *sctx = ctx;
4715	struct rb_node *node = NULL;
4716	struct recorded_ref data;
4717	struct recorded_ref *ref;
4718	u64 dir_gen;
4719
4720	ret = get_inode_gen(sctx->send_root, dir, &dir_gen);
4721	if (ret < 0)
4722		goto out;
4723
4724	data.dir = dir;
4725	data.dir_gen = dir_gen;
4726	set_ref_path(&data, name);
4727	node = rb_find(&data, &sctx->rbtree_deleted_refs, rbtree_ref_comp);
4728	if (node) {
4729		ref = rb_entry(node, struct recorded_ref, node);
4730		recorded_ref_free(ref);
4731	} else {
4732		ret = record_ref_in_tree(&sctx->rbtree_new_refs,
4733					 &sctx->new_refs, name, dir, dir_gen,
4734					 sctx);
4735	}
4736out:
4737	return ret;
4738}
4739
4740static int record_deleted_ref_if_needed(int num, u64 dir, int index,
4741					struct fs_path *name, void *ctx)
4742{
4743	int ret = 0;
4744	struct send_ctx *sctx = ctx;
4745	struct rb_node *node = NULL;
4746	struct recorded_ref data;
4747	struct recorded_ref *ref;
4748	u64 dir_gen;
4749
4750	ret = get_inode_gen(sctx->parent_root, dir, &dir_gen);
4751	if (ret < 0)
4752		goto out;
4753
4754	data.dir = dir;
4755	data.dir_gen = dir_gen;
4756	set_ref_path(&data, name);
4757	node = rb_find(&data, &sctx->rbtree_new_refs, rbtree_ref_comp);
4758	if (node) {
4759		ref = rb_entry(node, struct recorded_ref, node);
4760		recorded_ref_free(ref);
4761	} else {
4762		ret = record_ref_in_tree(&sctx->rbtree_deleted_refs,
4763					 &sctx->deleted_refs, name, dir,
4764					 dir_gen, sctx);
4765	}
4766out:
4767	return ret;
4768}
4769
4770static int record_new_ref(struct send_ctx *sctx)
4771{
4772	int ret;
4773
4774	ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4775				sctx->cmp_key, 0, record_new_ref_if_needed, sctx);
4776	if (ret < 0)
4777		goto out;
4778	ret = 0;
4779
4780out:
4781	return ret;
4782}
4783
4784static int record_deleted_ref(struct send_ctx *sctx)
4785{
4786	int ret;
4787
4788	ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4789				sctx->cmp_key, 0, record_deleted_ref_if_needed,
4790				sctx);
4791	if (ret < 0)
4792		goto out;
4793	ret = 0;
4794
4795out:
4796	return ret;
4797}
4798
4799static int record_changed_ref(struct send_ctx *sctx)
4800{
4801	int ret = 0;
4802
4803	ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4804			sctx->cmp_key, 0, record_new_ref_if_needed, sctx);
4805	if (ret < 0)
4806		goto out;
4807	ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4808			sctx->cmp_key, 0, record_deleted_ref_if_needed, sctx);
4809	if (ret < 0)
4810		goto out;
4811	ret = 0;
4812
4813out:
4814	return ret;
4815}
4816
4817/*
4818 * Record and process all refs at once. Needed when an inode changes the
4819 * generation number, which means that it was deleted and recreated.
4820 */
4821static int process_all_refs(struct send_ctx *sctx,
4822			    enum btrfs_compare_tree_result cmd)
4823{
4824	int ret = 0;
4825	int iter_ret = 0;
4826	struct btrfs_root *root;
4827	struct btrfs_path *path;
4828	struct btrfs_key key;
4829	struct btrfs_key found_key;
4830	iterate_inode_ref_t cb;
4831	int pending_move = 0;
4832
4833	path = alloc_path_for_send();
4834	if (!path)
4835		return -ENOMEM;
4836
4837	if (cmd == BTRFS_COMPARE_TREE_NEW) {
4838		root = sctx->send_root;
4839		cb = record_new_ref_if_needed;
4840	} else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4841		root = sctx->parent_root;
4842		cb = record_deleted_ref_if_needed;
4843	} else {
4844		btrfs_err(sctx->send_root->fs_info,
4845				"Wrong command %d in process_all_refs", cmd);
4846		ret = -EINVAL;
4847		goto out;
4848	}
4849
4850	key.objectid = sctx->cmp_key->objectid;
4851	key.type = BTRFS_INODE_REF_KEY;
4852	key.offset = 0;
4853	btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
4854		if (found_key.objectid != key.objectid ||
4855		    (found_key.type != BTRFS_INODE_REF_KEY &&
4856		     found_key.type != BTRFS_INODE_EXTREF_KEY))
4857			break;
4858
4859		ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4860		if (ret < 0)
4861			goto out;
4862	}
4863	/* Catch error found during iteration */
4864	if (iter_ret < 0) {
4865		ret = iter_ret;
4866		goto out;
4867	}
4868	btrfs_release_path(path);
4869
4870	/*
4871	 * We don't actually care about pending_move as we are simply
4872	 * re-creating this inode and will be rename'ing it into place once we
4873	 * rename the parent directory.
4874	 */
4875	ret = process_recorded_refs(sctx, &pending_move);
4876out:
4877	btrfs_free_path(path);
4878	return ret;
4879}
4880
4881static int send_set_xattr(struct send_ctx *sctx,
4882			  struct fs_path *path,
4883			  const char *name, int name_len,
4884			  const char *data, int data_len)
4885{
4886	int ret = 0;
4887
4888	ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4889	if (ret < 0)
4890		goto out;
4891
4892	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4893	TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4894	TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4895
4896	ret = send_cmd(sctx);
4897
4898tlv_put_failure:
4899out:
4900	return ret;
4901}
4902
4903static int send_remove_xattr(struct send_ctx *sctx,
4904			  struct fs_path *path,
4905			  const char *name, int name_len)
4906{
4907	int ret = 0;
4908
4909	ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4910	if (ret < 0)
4911		goto out;
4912
4913	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4914	TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4915
4916	ret = send_cmd(sctx);
4917
4918tlv_put_failure:
4919out:
4920	return ret;
4921}
4922
4923static int __process_new_xattr(int num, struct btrfs_key *di_key,
4924			       const char *name, int name_len, const char *data,
4925			       int data_len, void *ctx)
4926{
4927	int ret;
4928	struct send_ctx *sctx = ctx;
4929	struct fs_path *p;
4930	struct posix_acl_xattr_header dummy_acl;
4931
4932	/* Capabilities are emitted by finish_inode_if_needed */
4933	if (!strncmp(name, XATTR_NAME_CAPS, name_len))
4934		return 0;
4935
4936	p = fs_path_alloc();
4937	if (!p)
4938		return -ENOMEM;
4939
4940	/*
4941	 * This hack is needed because empty acls are stored as zero byte
4942	 * data in xattrs. Problem with that is, that receiving these zero byte
4943	 * acls will fail later. To fix this, we send a dummy acl list that
4944	 * only contains the version number and no entries.
4945	 */
4946	if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4947	    !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4948		if (data_len == 0) {
4949			dummy_acl.a_version =
4950					cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4951			data = (char *)&dummy_acl;
4952			data_len = sizeof(dummy_acl);
4953		}
4954	}
4955
4956	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4957	if (ret < 0)
4958		goto out;
4959
4960	ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4961
4962out:
4963	fs_path_free(p);
4964	return ret;
4965}
4966
4967static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4968				   const char *name, int name_len,
4969				   const char *data, int data_len, void *ctx)
4970{
4971	int ret;
4972	struct send_ctx *sctx = ctx;
4973	struct fs_path *p;
4974
4975	p = fs_path_alloc();
4976	if (!p)
4977		return -ENOMEM;
4978
4979	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4980	if (ret < 0)
4981		goto out;
4982
4983	ret = send_remove_xattr(sctx, p, name, name_len);
4984
4985out:
4986	fs_path_free(p);
4987	return ret;
4988}
4989
4990static int process_new_xattr(struct send_ctx *sctx)
4991{
4992	int ret = 0;
4993
4994	ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4995			       __process_new_xattr, sctx);
4996
4997	return ret;
4998}
4999
5000static int process_deleted_xattr(struct send_ctx *sctx)
5001{
5002	return iterate_dir_item(sctx->parent_root, sctx->right_path,
5003				__process_deleted_xattr, sctx);
5004}
5005
5006struct find_xattr_ctx {
5007	const char *name;
5008	int name_len;
5009	int found_idx;
5010	char *found_data;
5011	int found_data_len;
5012};
5013
5014static int __find_xattr(int num, struct btrfs_key *di_key, const char *name,
5015			int name_len, const char *data, int data_len, void *vctx)
5016{
5017	struct find_xattr_ctx *ctx = vctx;
5018
5019	if (name_len == ctx->name_len &&
5020	    strncmp(name, ctx->name, name_len) == 0) {
5021		ctx->found_idx = num;
5022		ctx->found_data_len = data_len;
5023		ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
5024		if (!ctx->found_data)
5025			return -ENOMEM;
5026		return 1;
5027	}
5028	return 0;
5029}
5030
5031static int find_xattr(struct btrfs_root *root,
5032		      struct btrfs_path *path,
5033		      struct btrfs_key *key,
5034		      const char *name, int name_len,
5035		      char **data, int *data_len)
5036{
5037	int ret;
5038	struct find_xattr_ctx ctx;
5039
5040	ctx.name = name;
5041	ctx.name_len = name_len;
5042	ctx.found_idx = -1;
5043	ctx.found_data = NULL;
5044	ctx.found_data_len = 0;
5045
5046	ret = iterate_dir_item(root, path, __find_xattr, &ctx);
5047	if (ret < 0)
5048		return ret;
5049
5050	if (ctx.found_idx == -1)
5051		return -ENOENT;
5052	if (data) {
5053		*data = ctx.found_data;
5054		*data_len = ctx.found_data_len;
5055	} else {
5056		kfree(ctx.found_data);
5057	}
5058	return ctx.found_idx;
5059}
5060
5061
5062static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
5063				       const char *name, int name_len,
5064				       const char *data, int data_len,
5065				       void *ctx)
5066{
5067	int ret;
5068	struct send_ctx *sctx = ctx;
5069	char *found_data = NULL;
5070	int found_data_len  = 0;
5071
5072	ret = find_xattr(sctx->parent_root, sctx->right_path,
5073			 sctx->cmp_key, name, name_len, &found_data,
5074			 &found_data_len);
5075	if (ret == -ENOENT) {
5076		ret = __process_new_xattr(num, di_key, name, name_len, data,
5077					  data_len, ctx);
5078	} else if (ret >= 0) {
5079		if (data_len != found_data_len ||
5080		    memcmp(data, found_data, data_len)) {
5081			ret = __process_new_xattr(num, di_key, name, name_len,
5082						  data, data_len, ctx);
5083		} else {
5084			ret = 0;
5085		}
5086	}
5087
5088	kfree(found_data);
5089	return ret;
5090}
5091
5092static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
5093					   const char *name, int name_len,
5094					   const char *data, int data_len,
5095					   void *ctx)
5096{
5097	int ret;
5098	struct send_ctx *sctx = ctx;
5099
5100	ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
5101			 name, name_len, NULL, NULL);
5102	if (ret == -ENOENT)
5103		ret = __process_deleted_xattr(num, di_key, name, name_len, data,
5104					      data_len, ctx);
5105	else if (ret >= 0)
5106		ret = 0;
5107
5108	return ret;
5109}
5110
5111static int process_changed_xattr(struct send_ctx *sctx)
5112{
5113	int ret = 0;
5114
5115	ret = iterate_dir_item(sctx->send_root, sctx->left_path,
5116			__process_changed_new_xattr, sctx);
5117	if (ret < 0)
5118		goto out;
5119	ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
5120			__process_changed_deleted_xattr, sctx);
5121
5122out:
5123	return ret;
5124}
5125
5126static int process_all_new_xattrs(struct send_ctx *sctx)
5127{
5128	int ret = 0;
5129	int iter_ret = 0;
5130	struct btrfs_root *root;
5131	struct btrfs_path *path;
5132	struct btrfs_key key;
5133	struct btrfs_key found_key;
5134
5135	path = alloc_path_for_send();
5136	if (!path)
5137		return -ENOMEM;
5138
5139	root = sctx->send_root;
5140
5141	key.objectid = sctx->cmp_key->objectid;
5142	key.type = BTRFS_XATTR_ITEM_KEY;
5143	key.offset = 0;
5144	btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
5145		if (found_key.objectid != key.objectid ||
5146		    found_key.type != key.type) {
5147			ret = 0;
5148			break;
5149		}
5150
5151		ret = iterate_dir_item(root, path, __process_new_xattr, sctx);
5152		if (ret < 0)
5153			break;
5154	}
5155	/* Catch error found during iteration */
5156	if (iter_ret < 0)
5157		ret = iter_ret;
5158
5159	btrfs_free_path(path);
5160	return ret;
5161}
5162
5163static int send_verity(struct send_ctx *sctx, struct fs_path *path,
5164		       struct fsverity_descriptor *desc)
5165{
5166	int ret;
5167
5168	ret = begin_cmd(sctx, BTRFS_SEND_C_ENABLE_VERITY);
5169	if (ret < 0)
5170		goto out;
5171
5172	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
5173	TLV_PUT_U8(sctx, BTRFS_SEND_A_VERITY_ALGORITHM,
5174			le8_to_cpu(desc->hash_algorithm));
5175	TLV_PUT_U32(sctx, BTRFS_SEND_A_VERITY_BLOCK_SIZE,
5176			1U << le8_to_cpu(desc->log_blocksize));
5177	TLV_PUT(sctx, BTRFS_SEND_A_VERITY_SALT_DATA, desc->salt,
5178			le8_to_cpu(desc->salt_size));
5179	TLV_PUT(sctx, BTRFS_SEND_A_VERITY_SIG_DATA, desc->signature,
5180			le32_to_cpu(desc->sig_size));
5181
5182	ret = send_cmd(sctx);
5183
5184tlv_put_failure:
5185out:
5186	return ret;
5187}
5188
5189static int process_verity(struct send_ctx *sctx)
5190{
5191	int ret = 0;
5192	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5193	struct inode *inode;
5194	struct fs_path *p;
5195
5196	inode = btrfs_iget(fs_info->sb, sctx->cur_ino, sctx->send_root);
5197	if (IS_ERR(inode))
5198		return PTR_ERR(inode);
5199
5200	ret = btrfs_get_verity_descriptor(inode, NULL, 0);
5201	if (ret < 0)
5202		goto iput;
5203
5204	if (ret > FS_VERITY_MAX_DESCRIPTOR_SIZE) {
5205		ret = -EMSGSIZE;
5206		goto iput;
5207	}
5208	if (!sctx->verity_descriptor) {
5209		sctx->verity_descriptor = kvmalloc(FS_VERITY_MAX_DESCRIPTOR_SIZE,
5210						   GFP_KERNEL);
5211		if (!sctx->verity_descriptor) {
5212			ret = -ENOMEM;
5213			goto iput;
5214		}
5215	}
5216
5217	ret = btrfs_get_verity_descriptor(inode, sctx->verity_descriptor, ret);
5218	if (ret < 0)
5219		goto iput;
5220
5221	p = fs_path_alloc();
5222	if (!p) {
5223		ret = -ENOMEM;
5224		goto iput;
5225	}
5226	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5227	if (ret < 0)
5228		goto free_path;
5229
5230	ret = send_verity(sctx, p, sctx->verity_descriptor);
5231	if (ret < 0)
5232		goto free_path;
5233
5234free_path:
5235	fs_path_free(p);
5236iput:
5237	iput(inode);
5238	return ret;
5239}
5240
5241static inline u64 max_send_read_size(const struct send_ctx *sctx)
5242{
5243	return sctx->send_max_size - SZ_16K;
5244}
5245
5246static int put_data_header(struct send_ctx *sctx, u32 len)
5247{
5248	if (WARN_ON_ONCE(sctx->put_data))
5249		return -EINVAL;
5250	sctx->put_data = true;
5251	if (sctx->proto >= 2) {
5252		/*
5253		 * Since v2, the data attribute header doesn't include a length,
5254		 * it is implicitly to the end of the command.
5255		 */
5256		if (sctx->send_max_size - sctx->send_size < sizeof(__le16) + len)
5257			return -EOVERFLOW;
5258		put_unaligned_le16(BTRFS_SEND_A_DATA, sctx->send_buf + sctx->send_size);
5259		sctx->send_size += sizeof(__le16);
5260	} else {
5261		struct btrfs_tlv_header *hdr;
5262
5263		if (sctx->send_max_size - sctx->send_size < sizeof(*hdr) + len)
5264			return -EOVERFLOW;
5265		hdr = (struct btrfs_tlv_header *)(sctx->send_buf + sctx->send_size);
5266		put_unaligned_le16(BTRFS_SEND_A_DATA, &hdr->tlv_type);
5267		put_unaligned_le16(len, &hdr->tlv_len);
5268		sctx->send_size += sizeof(*hdr);
5269	}
5270	return 0;
5271}
5272
5273static int put_file_data(struct send_ctx *sctx, u64 offset, u32 len)
5274{
5275	struct btrfs_root *root = sctx->send_root;
5276	struct btrfs_fs_info *fs_info = root->fs_info;
5277	struct page *page;
5278	pgoff_t index = offset >> PAGE_SHIFT;
5279	pgoff_t last_index;
5280	unsigned pg_offset = offset_in_page(offset);
5281	int ret;
5282
5283	ret = put_data_header(sctx, len);
5284	if (ret)
5285		return ret;
5286
5287	last_index = (offset + len - 1) >> PAGE_SHIFT;
5288
5289	while (index <= last_index) {
5290		unsigned cur_len = min_t(unsigned, len,
5291					 PAGE_SIZE - pg_offset);
5292
5293		page = find_lock_page(sctx->cur_inode->i_mapping, index);
5294		if (!page) {
5295			page_cache_sync_readahead(sctx->cur_inode->i_mapping,
5296						  &sctx->ra, NULL, index,
5297						  last_index + 1 - index);
5298
5299			page = find_or_create_page(sctx->cur_inode->i_mapping,
5300						   index, GFP_KERNEL);
5301			if (!page) {
5302				ret = -ENOMEM;
5303				break;
5304			}
5305		}
5306
5307		if (PageReadahead(page))
5308			page_cache_async_readahead(sctx->cur_inode->i_mapping,
5309						   &sctx->ra, NULL, page_folio(page),
5310						   index, last_index + 1 - index);
5311
5312		if (!PageUptodate(page)) {
5313			btrfs_read_folio(NULL, page_folio(page));
5314			lock_page(page);
5315			if (!PageUptodate(page)) {
5316				unlock_page(page);
5317				btrfs_err(fs_info,
5318			"send: IO error at offset %llu for inode %llu root %llu",
5319					page_offset(page), sctx->cur_ino,
5320					sctx->send_root->root_key.objectid);
5321				put_page(page);
5322				ret = -EIO;
5323				break;
5324			}
5325		}
5326
5327		memcpy_from_page(sctx->send_buf + sctx->send_size, page,
5328				 pg_offset, cur_len);
5329		unlock_page(page);
5330		put_page(page);
5331		index++;
5332		pg_offset = 0;
5333		len -= cur_len;
5334		sctx->send_size += cur_len;
5335	}
5336
5337	return ret;
5338}
5339
5340/*
5341 * Read some bytes from the current inode/file and send a write command to
5342 * user space.
5343 */
5344static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
5345{
5346	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5347	int ret = 0;
5348	struct fs_path *p;
5349
5350	p = fs_path_alloc();
5351	if (!p)
5352		return -ENOMEM;
5353
5354	btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
5355
5356	ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5357	if (ret < 0)
5358		goto out;
5359
5360	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5361	if (ret < 0)
5362		goto out;
5363
5364	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5365	TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5366	ret = put_file_data(sctx, offset, len);
5367	if (ret < 0)
5368		goto out;
5369
5370	ret = send_cmd(sctx);
5371
5372tlv_put_failure:
5373out:
5374	fs_path_free(p);
5375	return ret;
5376}
5377
5378/*
5379 * Send a clone command to user space.
5380 */
5381static int send_clone(struct send_ctx *sctx,
5382		      u64 offset, u32 len,
5383		      struct clone_root *clone_root)
5384{
5385	int ret = 0;
5386	struct fs_path *p;
5387	u64 gen;
5388
5389	btrfs_debug(sctx->send_root->fs_info,
5390		    "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
5391		    offset, len, clone_root->root->root_key.objectid,
5392		    clone_root->ino, clone_root->offset);
5393
5394	p = fs_path_alloc();
5395	if (!p)
5396		return -ENOMEM;
5397
5398	ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
5399	if (ret < 0)
5400		goto out;
5401
5402	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5403	if (ret < 0)
5404		goto out;
5405
5406	TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5407	TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
5408	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5409
5410	if (clone_root->root == sctx->send_root) {
5411		ret = get_inode_gen(sctx->send_root, clone_root->ino, &gen);
5412		if (ret < 0)
5413			goto out;
5414		ret = get_cur_path(sctx, clone_root->ino, gen, p);
5415	} else {
5416		ret = get_inode_path(clone_root->root, clone_root->ino, p);
5417	}
5418	if (ret < 0)
5419		goto out;
5420
5421	/*
5422	 * If the parent we're using has a received_uuid set then use that as
5423	 * our clone source as that is what we will look for when doing a
5424	 * receive.
5425	 *
5426	 * This covers the case that we create a snapshot off of a received
5427	 * subvolume and then use that as the parent and try to receive on a
5428	 * different host.
5429	 */
5430	if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
5431		TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5432			     clone_root->root->root_item.received_uuid);
5433	else
5434		TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5435			     clone_root->root->root_item.uuid);
5436	TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
5437		    btrfs_root_ctransid(&clone_root->root->root_item));
5438	TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
5439	TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
5440			clone_root->offset);
5441
5442	ret = send_cmd(sctx);
5443
5444tlv_put_failure:
5445out:
5446	fs_path_free(p);
5447	return ret;
5448}
5449
5450/*
5451 * Send an update extent command to user space.
5452 */
5453static int send_update_extent(struct send_ctx *sctx,
5454			      u64 offset, u32 len)
5455{
5456	int ret = 0;
5457	struct fs_path *p;
5458
5459	p = fs_path_alloc();
5460	if (!p)
5461		return -ENOMEM;
5462
5463	ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
5464	if (ret < 0)
5465		goto out;
5466
5467	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5468	if (ret < 0)
5469		goto out;
5470
5471	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5472	TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5473	TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
5474
5475	ret = send_cmd(sctx);
5476
5477tlv_put_failure:
5478out:
5479	fs_path_free(p);
5480	return ret;
5481}
5482
5483static int send_hole(struct send_ctx *sctx, u64 end)
5484{
5485	struct fs_path *p = NULL;
5486	u64 read_size = max_send_read_size(sctx);
5487	u64 offset = sctx->cur_inode_last_extent;
5488	int ret = 0;
5489
5490	/*
5491	 * A hole that starts at EOF or beyond it. Since we do not yet support
5492	 * fallocate (for extent preallocation and hole punching), sending a
5493	 * write of zeroes starting at EOF or beyond would later require issuing
5494	 * a truncate operation which would undo the write and achieve nothing.
5495	 */
5496	if (offset >= sctx->cur_inode_size)
5497		return 0;
5498
5499	/*
5500	 * Don't go beyond the inode's i_size due to prealloc extents that start
5501	 * after the i_size.
5502	 */
5503	end = min_t(u64, end, sctx->cur_inode_size);
5504
5505	if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5506		return send_update_extent(sctx, offset, end - offset);
5507
5508	p = fs_path_alloc();
5509	if (!p)
5510		return -ENOMEM;
5511	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5512	if (ret < 0)
5513		goto tlv_put_failure;
5514	while (offset < end) {
5515		u64 len = min(end - offset, read_size);
5516
5517		ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5518		if (ret < 0)
5519			break;
5520		TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5521		TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5522		ret = put_data_header(sctx, len);
5523		if (ret < 0)
5524			break;
5525		memset(sctx->send_buf + sctx->send_size, 0, len);
5526		sctx->send_size += len;
5527		ret = send_cmd(sctx);
5528		if (ret < 0)
5529			break;
5530		offset += len;
5531	}
5532	sctx->cur_inode_next_write_offset = offset;
5533tlv_put_failure:
5534	fs_path_free(p);
5535	return ret;
5536}
5537
5538static int send_encoded_inline_extent(struct send_ctx *sctx,
5539				      struct btrfs_path *path, u64 offset,
5540				      u64 len)
5541{
5542	struct btrfs_root *root = sctx->send_root;
5543	struct btrfs_fs_info *fs_info = root->fs_info;
5544	struct inode *inode;
5545	struct fs_path *fspath;
5546	struct extent_buffer *leaf = path->nodes[0];
5547	struct btrfs_key key;
5548	struct btrfs_file_extent_item *ei;
5549	u64 ram_bytes;
5550	size_t inline_size;
5551	int ret;
5552
5553	inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
5554	if (IS_ERR(inode))
5555		return PTR_ERR(inode);
5556
5557	fspath = fs_path_alloc();
5558	if (!fspath) {
5559		ret = -ENOMEM;
5560		goto out;
5561	}
5562
5563	ret = begin_cmd(sctx, BTRFS_SEND_C_ENCODED_WRITE);
5564	if (ret < 0)
5565		goto out;
5566
5567	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5568	if (ret < 0)
5569		goto out;
5570
5571	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
5572	ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
5573	ram_bytes = btrfs_file_extent_ram_bytes(leaf, ei);
5574	inline_size = btrfs_file_extent_inline_item_len(leaf, path->slots[0]);
5575
5576	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, fspath);
5577	TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5578	TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_FILE_LEN,
5579		    min(key.offset + ram_bytes - offset, len));
5580	TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_LEN, ram_bytes);
5581	TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_OFFSET, offset - key.offset);
5582	ret = btrfs_encoded_io_compression_from_extent(fs_info,
5583				btrfs_file_extent_compression(leaf, ei));
5584	if (ret < 0)
5585		goto out;
5586	TLV_PUT_U32(sctx, BTRFS_SEND_A_COMPRESSION, ret);
5587
5588	ret = put_data_header(sctx, inline_size);
5589	if (ret < 0)
5590		goto out;
5591	read_extent_buffer(leaf, sctx->send_buf + sctx->send_size,
5592			   btrfs_file_extent_inline_start(ei), inline_size);
5593	sctx->send_size += inline_size;
5594
5595	ret = send_cmd(sctx);
5596
5597tlv_put_failure:
5598out:
5599	fs_path_free(fspath);
5600	iput(inode);
5601	return ret;
5602}
5603
5604static int send_encoded_extent(struct send_ctx *sctx, struct btrfs_path *path,
5605			       u64 offset, u64 len)
5606{
5607	struct btrfs_root *root = sctx->send_root;
5608	struct btrfs_fs_info *fs_info = root->fs_info;
5609	struct inode *inode;
5610	struct fs_path *fspath;
5611	struct extent_buffer *leaf = path->nodes[0];
5612	struct btrfs_key key;
5613	struct btrfs_file_extent_item *ei;
5614	u64 disk_bytenr, disk_num_bytes;
5615	u32 data_offset;
5616	struct btrfs_cmd_header *hdr;
5617	u32 crc;
5618	int ret;
5619
5620	inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
5621	if (IS_ERR(inode))
5622		return PTR_ERR(inode);
5623
5624	fspath = fs_path_alloc();
5625	if (!fspath) {
5626		ret = -ENOMEM;
5627		goto out;
5628	}
5629
5630	ret = begin_cmd(sctx, BTRFS_SEND_C_ENCODED_WRITE);
5631	if (ret < 0)
5632		goto out;
5633
5634	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5635	if (ret < 0)
5636		goto out;
5637
5638	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
5639	ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
5640	disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
5641	disk_num_bytes = btrfs_file_extent_disk_num_bytes(leaf, ei);
5642
5643	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, fspath);
5644	TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5645	TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_FILE_LEN,
5646		    min(key.offset + btrfs_file_extent_num_bytes(leaf, ei) - offset,
5647			len));
5648	TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_LEN,
5649		    btrfs_file_extent_ram_bytes(leaf, ei));
5650	TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_OFFSET,
5651		    offset - key.offset + btrfs_file_extent_offset(leaf, ei));
5652	ret = btrfs_encoded_io_compression_from_extent(fs_info,
5653				btrfs_file_extent_compression(leaf, ei));
5654	if (ret < 0)
5655		goto out;
5656	TLV_PUT_U32(sctx, BTRFS_SEND_A_COMPRESSION, ret);
5657	TLV_PUT_U32(sctx, BTRFS_SEND_A_ENCRYPTION, 0);
5658
5659	ret = put_data_header(sctx, disk_num_bytes);
5660	if (ret < 0)
5661		goto out;
5662
5663	/*
5664	 * We want to do I/O directly into the send buffer, so get the next page
5665	 * boundary in the send buffer. This means that there may be a gap
5666	 * between the beginning of the command and the file data.
5667	 */
5668	data_offset = PAGE_ALIGN(sctx->send_size);
5669	if (data_offset > sctx->send_max_size ||
5670	    sctx->send_max_size - data_offset < disk_num_bytes) {
5671		ret = -EOVERFLOW;
5672		goto out;
5673	}
5674
5675	/*
5676	 * Note that send_buf is a mapping of send_buf_pages, so this is really
5677	 * reading into send_buf.
5678	 */
5679	ret = btrfs_encoded_read_regular_fill_pages(BTRFS_I(inode), offset,
5680						    disk_bytenr, disk_num_bytes,
5681						    sctx->send_buf_pages +
5682						    (data_offset >> PAGE_SHIFT));
5683	if (ret)
5684		goto out;
5685
5686	hdr = (struct btrfs_cmd_header *)sctx->send_buf;
5687	hdr->len = cpu_to_le32(sctx->send_size + disk_num_bytes - sizeof(*hdr));
5688	hdr->crc = 0;
5689	crc = crc32c(0, sctx->send_buf, sctx->send_size);
5690	crc = crc32c(crc, sctx->send_buf + data_offset, disk_num_bytes);
5691	hdr->crc = cpu_to_le32(crc);
5692
5693	ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
5694			&sctx->send_off);
5695	if (!ret) {
5696		ret = write_buf(sctx->send_filp, sctx->send_buf + data_offset,
5697				disk_num_bytes, &sctx->send_off);
5698	}
5699	sctx->send_size = 0;
5700	sctx->put_data = false;
5701
5702tlv_put_failure:
5703out:
5704	fs_path_free(fspath);
5705	iput(inode);
5706	return ret;
5707}
5708
5709static int send_extent_data(struct send_ctx *sctx, struct btrfs_path *path,
5710			    const u64 offset, const u64 len)
5711{
5712	const u64 end = offset + len;
5713	struct extent_buffer *leaf = path->nodes[0];
5714	struct btrfs_file_extent_item *ei;
5715	u64 read_size = max_send_read_size(sctx);
5716	u64 sent = 0;
5717
5718	if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5719		return send_update_extent(sctx, offset, len);
5720
5721	ei = btrfs_item_ptr(leaf, path->slots[0],
5722			    struct btrfs_file_extent_item);
5723	if ((sctx->flags & BTRFS_SEND_FLAG_COMPRESSED) &&
5724	    btrfs_file_extent_compression(leaf, ei) != BTRFS_COMPRESS_NONE) {
5725		bool is_inline = (btrfs_file_extent_type(leaf, ei) ==
5726				  BTRFS_FILE_EXTENT_INLINE);
5727
5728		/*
5729		 * Send the compressed extent unless the compressed data is
5730		 * larger than the decompressed data. This can happen if we're
5731		 * not sending the entire extent, either because it has been
5732		 * partially overwritten/truncated or because this is a part of
5733		 * the extent that we couldn't clone in clone_range().
5734		 */
5735		if (is_inline &&
5736		    btrfs_file_extent_inline_item_len(leaf,
5737						      path->slots[0]) <= len) {
5738			return send_encoded_inline_extent(sctx, path, offset,
5739							  len);
5740		} else if (!is_inline &&
5741			   btrfs_file_extent_disk_num_bytes(leaf, ei) <= len) {
5742			return send_encoded_extent(sctx, path, offset, len);
5743		}
5744	}
5745
5746	if (sctx->cur_inode == NULL) {
5747		struct btrfs_root *root = sctx->send_root;
5748
5749		sctx->cur_inode = btrfs_iget(root->fs_info->sb, sctx->cur_ino, root);
5750		if (IS_ERR(sctx->cur_inode)) {
5751			int err = PTR_ERR(sctx->cur_inode);
5752
5753			sctx->cur_inode = NULL;
5754			return err;
5755		}
5756		memset(&sctx->ra, 0, sizeof(struct file_ra_state));
5757		file_ra_state_init(&sctx->ra, sctx->cur_inode->i_mapping);
5758
5759		/*
5760		 * It's very likely there are no pages from this inode in the page
5761		 * cache, so after reading extents and sending their data, we clean
5762		 * the page cache to avoid trashing the page cache (adding pressure
5763		 * to the page cache and forcing eviction of other data more useful
5764		 * for applications).
5765		 *
5766		 * We decide if we should clean the page cache simply by checking
5767		 * if the inode's mapping nrpages is 0 when we first open it, and
5768		 * not by using something like filemap_range_has_page() before
5769		 * reading an extent because when we ask the readahead code to
5770		 * read a given file range, it may (and almost always does) read
5771		 * pages from beyond that range (see the documentation for
5772		 * page_cache_sync_readahead()), so it would not be reliable,
5773		 * because after reading the first extent future calls to
5774		 * filemap_range_has_page() would return true because the readahead
5775		 * on the previous extent resulted in reading pages of the current
5776		 * extent as well.
5777		 */
5778		sctx->clean_page_cache = (sctx->cur_inode->i_mapping->nrpages == 0);
5779		sctx->page_cache_clear_start = round_down(offset, PAGE_SIZE);
5780	}
5781
5782	while (sent < len) {
5783		u64 size = min(len - sent, read_size);
5784		int ret;
5785
5786		ret = send_write(sctx, offset + sent, size);
5787		if (ret < 0)
5788			return ret;
5789		sent += size;
5790	}
5791
5792	if (sctx->clean_page_cache && PAGE_ALIGNED(end)) {
5793		/*
5794		 * Always operate only on ranges that are a multiple of the page
5795		 * size. This is not only to prevent zeroing parts of a page in
5796		 * the case of subpage sector size, but also to guarantee we evict
5797		 * pages, as passing a range that is smaller than page size does
5798		 * not evict the respective page (only zeroes part of its content).
5799		 *
5800		 * Always start from the end offset of the last range cleared.
5801		 * This is because the readahead code may (and very often does)
5802		 * reads pages beyond the range we request for readahead. So if
5803		 * we have an extent layout like this:
5804		 *
5805		 *            [ extent A ] [ extent B ] [ extent C ]
5806		 *
5807		 * When we ask page_cache_sync_readahead() to read extent A, it
5808		 * may also trigger reads for pages of extent B. If we are doing
5809		 * an incremental send and extent B has not changed between the
5810		 * parent and send snapshots, some or all of its pages may end
5811		 * up being read and placed in the page cache. So when truncating
5812		 * the page cache we always start from the end offset of the
5813		 * previously processed extent up to the end of the current
5814		 * extent.
5815		 */
5816		truncate_inode_pages_range(&sctx->cur_inode->i_data,
5817					   sctx->page_cache_clear_start,
5818					   end - 1);
5819		sctx->page_cache_clear_start = end;
5820	}
5821
5822	return 0;
5823}
5824
5825/*
5826 * Search for a capability xattr related to sctx->cur_ino. If the capability is
5827 * found, call send_set_xattr function to emit it.
5828 *
5829 * Return 0 if there isn't a capability, or when the capability was emitted
5830 * successfully, or < 0 if an error occurred.
5831 */
5832static int send_capabilities(struct send_ctx *sctx)
5833{
5834	struct fs_path *fspath = NULL;
5835	struct btrfs_path *path;
5836	struct btrfs_dir_item *di;
5837	struct extent_buffer *leaf;
5838	unsigned long data_ptr;
5839	char *buf = NULL;
5840	int buf_len;
5841	int ret = 0;
5842
5843	path = alloc_path_for_send();
5844	if (!path)
5845		return -ENOMEM;
5846
5847	di = btrfs_lookup_xattr(NULL, sctx->send_root, path, sctx->cur_ino,
5848				XATTR_NAME_CAPS, strlen(XATTR_NAME_CAPS), 0);
5849	if (!di) {
5850		/* There is no xattr for this inode */
5851		goto out;
5852	} else if (IS_ERR(di)) {
5853		ret = PTR_ERR(di);
5854		goto out;
5855	}
5856
5857	leaf = path->nodes[0];
5858	buf_len = btrfs_dir_data_len(leaf, di);
5859
5860	fspath = fs_path_alloc();
5861	buf = kmalloc(buf_len, GFP_KERNEL);
5862	if (!fspath || !buf) {
5863		ret = -ENOMEM;
5864		goto out;
5865	}
5866
5867	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5868	if (ret < 0)
5869		goto out;
5870
5871	data_ptr = (unsigned long)(di + 1) + btrfs_dir_name_len(leaf, di);
5872	read_extent_buffer(leaf, buf, data_ptr, buf_len);
5873
5874	ret = send_set_xattr(sctx, fspath, XATTR_NAME_CAPS,
5875			strlen(XATTR_NAME_CAPS), buf, buf_len);
5876out:
5877	kfree(buf);
5878	fs_path_free(fspath);
5879	btrfs_free_path(path);
5880	return ret;
5881}
5882
5883static int clone_range(struct send_ctx *sctx, struct btrfs_path *dst_path,
5884		       struct clone_root *clone_root, const u64 disk_byte,
5885		       u64 data_offset, u64 offset, u64 len)
5886{
5887	struct btrfs_path *path;
5888	struct btrfs_key key;
5889	int ret;
5890	struct btrfs_inode_info info;
5891	u64 clone_src_i_size = 0;
5892
5893	/*
5894	 * Prevent cloning from a zero offset with a length matching the sector
5895	 * size because in some scenarios this will make the receiver fail.
5896	 *
5897	 * For example, if in the source filesystem the extent at offset 0
5898	 * has a length of sectorsize and it was written using direct IO, then
5899	 * it can never be an inline extent (even if compression is enabled).
5900	 * Then this extent can be cloned in the original filesystem to a non
5901	 * zero file offset, but it may not be possible to clone in the
5902	 * destination filesystem because it can be inlined due to compression
5903	 * on the destination filesystem (as the receiver's write operations are
5904	 * always done using buffered IO). The same happens when the original
5905	 * filesystem does not have compression enabled but the destination
5906	 * filesystem has.
5907	 */
5908	if (clone_root->offset == 0 &&
5909	    len == sctx->send_root->fs_info->sectorsize)
5910		return send_extent_data(sctx, dst_path, offset, len);
5911
5912	path = alloc_path_for_send();
5913	if (!path)
5914		return -ENOMEM;
5915
5916	/*
5917	 * There are inodes that have extents that lie behind its i_size. Don't
5918	 * accept clones from these extents.
5919	 */
5920	ret = get_inode_info(clone_root->root, clone_root->ino, &info);
5921	btrfs_release_path(path);
5922	if (ret < 0)
5923		goto out;
5924	clone_src_i_size = info.size;
5925
5926	/*
5927	 * We can't send a clone operation for the entire range if we find
5928	 * extent items in the respective range in the source file that
5929	 * refer to different extents or if we find holes.
5930	 * So check for that and do a mix of clone and regular write/copy
5931	 * operations if needed.
5932	 *
5933	 * Example:
5934	 *
5935	 * mkfs.btrfs -f /dev/sda
5936	 * mount /dev/sda /mnt
5937	 * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
5938	 * cp --reflink=always /mnt/foo /mnt/bar
5939	 * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
5940	 * btrfs subvolume snapshot -r /mnt /mnt/snap
5941	 *
5942	 * If when we send the snapshot and we are processing file bar (which
5943	 * has a higher inode number than foo) we blindly send a clone operation
5944	 * for the [0, 100K[ range from foo to bar, the receiver ends up getting
5945	 * a file bar that matches the content of file foo - iow, doesn't match
5946	 * the content from bar in the original filesystem.
5947	 */
5948	key.objectid = clone_root->ino;
5949	key.type = BTRFS_EXTENT_DATA_KEY;
5950	key.offset = clone_root->offset;
5951	ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
5952	if (ret < 0)
5953		goto out;
5954	if (ret > 0 && path->slots[0] > 0) {
5955		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
5956		if (key.objectid == clone_root->ino &&
5957		    key.type == BTRFS_EXTENT_DATA_KEY)
5958			path->slots[0]--;
5959	}
5960
5961	while (true) {
5962		struct extent_buffer *leaf = path->nodes[0];
5963		int slot = path->slots[0];
5964		struct btrfs_file_extent_item *ei;
5965		u8 type;
5966		u64 ext_len;
5967		u64 clone_len;
5968		u64 clone_data_offset;
5969		bool crossed_src_i_size = false;
5970
5971		if (slot >= btrfs_header_nritems(leaf)) {
5972			ret = btrfs_next_leaf(clone_root->root, path);
5973			if (ret < 0)
5974				goto out;
5975			else if (ret > 0)
5976				break;
5977			continue;
5978		}
5979
5980		btrfs_item_key_to_cpu(leaf, &key, slot);
5981
5982		/*
5983		 * We might have an implicit trailing hole (NO_HOLES feature
5984		 * enabled). We deal with it after leaving this loop.
5985		 */
5986		if (key.objectid != clone_root->ino ||
5987		    key.type != BTRFS_EXTENT_DATA_KEY)
5988			break;
5989
5990		ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5991		type = btrfs_file_extent_type(leaf, ei);
5992		if (type == BTRFS_FILE_EXTENT_INLINE) {
5993			ext_len = btrfs_file_extent_ram_bytes(leaf, ei);
5994			ext_len = PAGE_ALIGN(ext_len);
5995		} else {
5996			ext_len = btrfs_file_extent_num_bytes(leaf, ei);
5997		}
5998
5999		if (key.offset + ext_len <= clone_root->offset)
6000			goto next;
6001
6002		if (key.offset > clone_root->offset) {
6003			/* Implicit hole, NO_HOLES feature enabled. */
6004			u64 hole_len = key.offset - clone_root->offset;
6005
6006			if (hole_len > len)
6007				hole_len = len;
6008			ret = send_extent_data(sctx, dst_path, offset,
6009					       hole_len);
6010			if (ret < 0)
6011				goto out;
6012
6013			len -= hole_len;
6014			if (len == 0)
6015				break;
6016			offset += hole_len;
6017			clone_root->offset += hole_len;
6018			data_offset += hole_len;
6019		}
6020
6021		if (key.offset >= clone_root->offset + len)
6022			break;
6023
6024		if (key.offset >= clone_src_i_size)
6025			break;
6026
6027		if (key.offset + ext_len > clone_src_i_size) {
6028			ext_len = clone_src_i_size - key.offset;
6029			crossed_src_i_size = true;
6030		}
6031
6032		clone_data_offset = btrfs_file_extent_offset(leaf, ei);
6033		if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte) {
6034			clone_root->offset = key.offset;
6035			if (clone_data_offset < data_offset &&
6036				clone_data_offset + ext_len > data_offset) {
6037				u64 extent_offset;
6038
6039				extent_offset = data_offset - clone_data_offset;
6040				ext_len -= extent_offset;
6041				clone_data_offset += extent_offset;
6042				clone_root->offset += extent_offset;
6043			}
6044		}
6045
6046		clone_len = min_t(u64, ext_len, len);
6047
6048		if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
6049		    clone_data_offset == data_offset) {
6050			const u64 src_end = clone_root->offset + clone_len;
6051			const u64 sectorsize = SZ_64K;
6052
6053			/*
6054			 * We can't clone the last block, when its size is not
6055			 * sector size aligned, into the middle of a file. If we
6056			 * do so, the receiver will get a failure (-EINVAL) when
6057			 * trying to clone or will silently corrupt the data in
6058			 * the destination file if it's on a kernel without the
6059			 * fix introduced by commit ac765f83f1397646
6060			 * ("Btrfs: fix data corruption due to cloning of eof
6061			 * block).
6062			 *
6063			 * So issue a clone of the aligned down range plus a
6064			 * regular write for the eof block, if we hit that case.
6065			 *
6066			 * Also, we use the maximum possible sector size, 64K,
6067			 * because we don't know what's the sector size of the
6068			 * filesystem that receives the stream, so we have to
6069			 * assume the largest possible sector size.
6070			 */
6071			if (src_end == clone_src_i_size &&
6072			    !IS_ALIGNED(src_end, sectorsize) &&
6073			    offset + clone_len < sctx->cur_inode_size) {
6074				u64 slen;
6075
6076				slen = ALIGN_DOWN(src_end - clone_root->offset,
6077						  sectorsize);
6078				if (slen > 0) {
6079					ret = send_clone(sctx, offset, slen,
6080							 clone_root);
6081					if (ret < 0)
6082						goto out;
6083				}
6084				ret = send_extent_data(sctx, dst_path,
6085						       offset + slen,
6086						       clone_len - slen);
6087			} else {
6088				ret = send_clone(sctx, offset, clone_len,
6089						 clone_root);
6090			}
6091		} else if (crossed_src_i_size && clone_len < len) {
6092			/*
6093			 * If we are at i_size of the clone source inode and we
6094			 * can not clone from it, terminate the loop. This is
6095			 * to avoid sending two write operations, one with a
6096			 * length matching clone_len and the final one after
6097			 * this loop with a length of len - clone_len.
6098			 *
6099			 * When using encoded writes (BTRFS_SEND_FLAG_COMPRESSED
6100			 * was passed to the send ioctl), this helps avoid
6101			 * sending an encoded write for an offset that is not
6102			 * sector size aligned, in case the i_size of the source
6103			 * inode is not sector size aligned. That will make the
6104			 * receiver fallback to decompression of the data and
6105			 * writing it using regular buffered IO, therefore while
6106			 * not incorrect, it's not optimal due decompression and
6107			 * possible re-compression at the receiver.
6108			 */
6109			break;
6110		} else {
6111			ret = send_extent_data(sctx, dst_path, offset,
6112					       clone_len);
6113		}
6114
6115		if (ret < 0)
6116			goto out;
6117
6118		len -= clone_len;
6119		if (len == 0)
6120			break;
6121		offset += clone_len;
6122		clone_root->offset += clone_len;
6123
6124		/*
6125		 * If we are cloning from the file we are currently processing,
6126		 * and using the send root as the clone root, we must stop once
6127		 * the current clone offset reaches the current eof of the file
6128		 * at the receiver, otherwise we would issue an invalid clone
6129		 * operation (source range going beyond eof) and cause the
6130		 * receiver to fail. So if we reach the current eof, bail out
6131		 * and fallback to a regular write.
6132		 */
6133		if (clone_root->root == sctx->send_root &&
6134		    clone_root->ino == sctx->cur_ino &&
6135		    clone_root->offset >= sctx->cur_inode_next_write_offset)
6136			break;
6137
6138		data_offset += clone_len;
6139next:
6140		path->slots[0]++;
6141	}
6142
6143	if (len > 0)
6144		ret = send_extent_data(sctx, dst_path, offset, len);
6145	else
6146		ret = 0;
6147out:
6148	btrfs_free_path(path);
6149	return ret;
6150}
6151
6152static int send_write_or_clone(struct send_ctx *sctx,
6153			       struct btrfs_path *path,
6154			       struct btrfs_key *key,
6155			       struct clone_root *clone_root)
6156{
6157	int ret = 0;
6158	u64 offset = key->offset;
6159	u64 end;
6160	u64 bs = sctx->send_root->fs_info->sectorsize;
6161
6162	end = min_t(u64, btrfs_file_extent_end(path), sctx->cur_inode_size);
6163	if (offset >= end)
6164		return 0;
6165
6166	if (clone_root && IS_ALIGNED(end, bs)) {
6167		struct btrfs_file_extent_item *ei;
6168		u64 disk_byte;
6169		u64 data_offset;
6170
6171		ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
6172				    struct btrfs_file_extent_item);
6173		disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
6174		data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
6175		ret = clone_range(sctx, path, clone_root, disk_byte,
6176				  data_offset, offset, end - offset);
6177	} else {
6178		ret = send_extent_data(sctx, path, offset, end - offset);
6179	}
6180	sctx->cur_inode_next_write_offset = end;
6181	return ret;
6182}
6183
6184static int is_extent_unchanged(struct send_ctx *sctx,
6185			       struct btrfs_path *left_path,
6186			       struct btrfs_key *ekey)
6187{
6188	int ret = 0;
6189	struct btrfs_key key;
6190	struct btrfs_path *path = NULL;
6191	struct extent_buffer *eb;
6192	int slot;
6193	struct btrfs_key found_key;
6194	struct btrfs_file_extent_item *ei;
6195	u64 left_disknr;
6196	u64 right_disknr;
6197	u64 left_offset;
6198	u64 right_offset;
6199	u64 left_offset_fixed;
6200	u64 left_len;
6201	u64 right_len;
6202	u64 left_gen;
6203	u64 right_gen;
6204	u8 left_type;
6205	u8 right_type;
6206
6207	path = alloc_path_for_send();
6208	if (!path)
6209		return -ENOMEM;
6210
6211	eb = left_path->nodes[0];
6212	slot = left_path->slots[0];
6213	ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
6214	left_type = btrfs_file_extent_type(eb, ei);
6215
6216	if (left_type != BTRFS_FILE_EXTENT_REG) {
6217		ret = 0;
6218		goto out;
6219	}
6220	left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
6221	left_len = btrfs_file_extent_num_bytes(eb, ei);
6222	left_offset = btrfs_file_extent_offset(eb, ei);
6223	left_gen = btrfs_file_extent_generation(eb, ei);
6224
6225	/*
6226	 * Following comments will refer to these graphics. L is the left
6227	 * extents which we are checking at the moment. 1-8 are the right
6228	 * extents that we iterate.
6229	 *
6230	 *       |-----L-----|
6231	 * |-1-|-2a-|-3-|-4-|-5-|-6-|
6232	 *
6233	 *       |-----L-----|
6234	 * |--1--|-2b-|...(same as above)
6235	 *
6236	 * Alternative situation. Happens on files where extents got split.
6237	 *       |-----L-----|
6238	 * |-----------7-----------|-6-|
6239	 *
6240	 * Alternative situation. Happens on files which got larger.
6241	 *       |-----L-----|
6242	 * |-8-|
6243	 * Nothing follows after 8.
6244	 */
6245
6246	key.objectid = ekey->objectid;
6247	key.type = BTRFS_EXTENT_DATA_KEY;
6248	key.offset = ekey->offset;
6249	ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
6250	if (ret < 0)
6251		goto out;
6252	if (ret) {
6253		ret = 0;
6254		goto out;
6255	}
6256
6257	/*
6258	 * Handle special case where the right side has no extents at all.
6259	 */
6260	eb = path->nodes[0];
6261	slot = path->slots[0];
6262	btrfs_item_key_to_cpu(eb, &found_key, slot);
6263	if (found_key.objectid != key.objectid ||
6264	    found_key.type != key.type) {
6265		/* If we're a hole then just pretend nothing changed */
6266		ret = (left_disknr) ? 0 : 1;
6267		goto out;
6268	}
6269
6270	/*
6271	 * We're now on 2a, 2b or 7.
6272	 */
6273	key = found_key;
6274	while (key.offset < ekey->offset + left_len) {
6275		ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
6276		right_type = btrfs_file_extent_type(eb, ei);
6277		if (right_type != BTRFS_FILE_EXTENT_REG &&
6278		    right_type != BTRFS_FILE_EXTENT_INLINE) {
6279			ret = 0;
6280			goto out;
6281		}
6282
6283		if (right_type == BTRFS_FILE_EXTENT_INLINE) {
6284			right_len = btrfs_file_extent_ram_bytes(eb, ei);
6285			right_len = PAGE_ALIGN(right_len);
6286		} else {
6287			right_len = btrfs_file_extent_num_bytes(eb, ei);
6288		}
6289
6290		/*
6291		 * Are we at extent 8? If yes, we know the extent is changed.
6292		 * This may only happen on the first iteration.
6293		 */
6294		if (found_key.offset + right_len <= ekey->offset) {
6295			/* If we're a hole just pretend nothing changed */
6296			ret = (left_disknr) ? 0 : 1;
6297			goto out;
6298		}
6299
6300		/*
6301		 * We just wanted to see if when we have an inline extent, what
6302		 * follows it is a regular extent (wanted to check the above
6303		 * condition for inline extents too). This should normally not
6304		 * happen but it's possible for example when we have an inline
6305		 * compressed extent representing data with a size matching
6306		 * the page size (currently the same as sector size).
6307		 */
6308		if (right_type == BTRFS_FILE_EXTENT_INLINE) {
6309			ret = 0;
6310			goto out;
6311		}
6312
6313		right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
6314		right_offset = btrfs_file_extent_offset(eb, ei);
6315		right_gen = btrfs_file_extent_generation(eb, ei);
6316
6317		left_offset_fixed = left_offset;
6318		if (key.offset < ekey->offset) {
6319			/* Fix the right offset for 2a and 7. */
6320			right_offset += ekey->offset - key.offset;
6321		} else {
6322			/* Fix the left offset for all behind 2a and 2b */
6323			left_offset_fixed += key.offset - ekey->offset;
6324		}
6325
6326		/*
6327		 * Check if we have the same extent.
6328		 */
6329		if (left_disknr != right_disknr ||
6330		    left_offset_fixed != right_offset ||
6331		    left_gen != right_gen) {
6332			ret = 0;
6333			goto out;
6334		}
6335
6336		/*
6337		 * Go to the next extent.
6338		 */
6339		ret = btrfs_next_item(sctx->parent_root, path);
6340		if (ret < 0)
6341			goto out;
6342		if (!ret) {
6343			eb = path->nodes[0];
6344			slot = path->slots[0];
6345			btrfs_item_key_to_cpu(eb, &found_key, slot);
6346		}
6347		if (ret || found_key.objectid != key.objectid ||
6348		    found_key.type != key.type) {
6349			key.offset += right_len;
6350			break;
6351		}
6352		if (found_key.offset != key.offset + right_len) {
6353			ret = 0;
6354			goto out;
6355		}
6356		key = found_key;
6357	}
6358
6359	/*
6360	 * We're now behind the left extent (treat as unchanged) or at the end
6361	 * of the right side (treat as changed).
6362	 */
6363	if (key.offset >= ekey->offset + left_len)
6364		ret = 1;
6365	else
6366		ret = 0;
6367
6368
6369out:
6370	btrfs_free_path(path);
6371	return ret;
6372}
6373
6374static int get_last_extent(struct send_ctx *sctx, u64 offset)
6375{
6376	struct btrfs_path *path;
6377	struct btrfs_root *root = sctx->send_root;
6378	struct btrfs_key key;
6379	int ret;
6380
6381	path = alloc_path_for_send();
6382	if (!path)
6383		return -ENOMEM;
6384
6385	sctx->cur_inode_last_extent = 0;
6386
6387	key.objectid = sctx->cur_ino;
6388	key.type = BTRFS_EXTENT_DATA_KEY;
6389	key.offset = offset;
6390	ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
6391	if (ret < 0)
6392		goto out;
6393	ret = 0;
6394	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
6395	if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
6396		goto out;
6397
6398	sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
6399out:
6400	btrfs_free_path(path);
6401	return ret;
6402}
6403
6404static int range_is_hole_in_parent(struct send_ctx *sctx,
6405				   const u64 start,
6406				   const u64 end)
6407{
6408	struct btrfs_path *path;
6409	struct btrfs_key key;
6410	struct btrfs_root *root = sctx->parent_root;
6411	u64 search_start = start;
6412	int ret;
6413
6414	path = alloc_path_for_send();
6415	if (!path)
6416		return -ENOMEM;
6417
6418	key.objectid = sctx->cur_ino;
6419	key.type = BTRFS_EXTENT_DATA_KEY;
6420	key.offset = search_start;
6421	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6422	if (ret < 0)
6423		goto out;
6424	if (ret > 0 && path->slots[0] > 0)
6425		path->slots[0]--;
6426
6427	while (search_start < end) {
6428		struct extent_buffer *leaf = path->nodes[0];
6429		int slot = path->slots[0];
6430		struct btrfs_file_extent_item *fi;
6431		u64 extent_end;
6432
6433		if (slot >= btrfs_header_nritems(leaf)) {
6434			ret = btrfs_next_leaf(root, path);
6435			if (ret < 0)
6436				goto out;
6437			else if (ret > 0)
6438				break;
6439			continue;
6440		}
6441
6442		btrfs_item_key_to_cpu(leaf, &key, slot);
6443		if (key.objectid < sctx->cur_ino ||
6444		    key.type < BTRFS_EXTENT_DATA_KEY)
6445			goto next;
6446		if (key.objectid > sctx->cur_ino ||
6447		    key.type > BTRFS_EXTENT_DATA_KEY ||
6448		    key.offset >= end)
6449			break;
6450
6451		fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
6452		extent_end = btrfs_file_extent_end(path);
6453		if (extent_end <= start)
6454			goto next;
6455		if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
6456			search_start = extent_end;
6457			goto next;
6458		}
6459		ret = 0;
6460		goto out;
6461next:
6462		path->slots[0]++;
6463	}
6464	ret = 1;
6465out:
6466	btrfs_free_path(path);
6467	return ret;
6468}
6469
6470static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
6471			   struct btrfs_key *key)
6472{
6473	int ret = 0;
6474
6475	if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
6476		return 0;
6477
6478	/*
6479	 * Get last extent's end offset (exclusive) if we haven't determined it
6480	 * yet (we're processing the first file extent item that is new), or if
6481	 * we're at the first slot of a leaf and the last extent's end is less
6482	 * than the current extent's offset, because we might have skipped
6483	 * entire leaves that contained only file extent items for our current
6484	 * inode. These leaves have a generation number smaller (older) than the
6485	 * one in the current leaf and the leaf our last extent came from, and
6486	 * are located between these 2 leaves.
6487	 */
6488	if ((sctx->cur_inode_last_extent == (u64)-1) ||
6489	    (path->slots[0] == 0 && sctx->cur_inode_last_extent < key->offset)) {
 
 
 
6490		ret = get_last_extent(sctx, key->offset - 1);
6491		if (ret)
6492			return ret;
6493	}
6494
6495	if (sctx->cur_inode_last_extent < key->offset) {
6496		ret = range_is_hole_in_parent(sctx,
6497					      sctx->cur_inode_last_extent,
6498					      key->offset);
6499		if (ret < 0)
6500			return ret;
6501		else if (ret == 0)
6502			ret = send_hole(sctx, key->offset);
6503		else
6504			ret = 0;
6505	}
6506	sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
6507	return ret;
6508}
6509
6510static int process_extent(struct send_ctx *sctx,
6511			  struct btrfs_path *path,
6512			  struct btrfs_key *key)
6513{
6514	struct clone_root *found_clone = NULL;
6515	int ret = 0;
6516
6517	if (S_ISLNK(sctx->cur_inode_mode))
6518		return 0;
6519
6520	if (sctx->parent_root && !sctx->cur_inode_new) {
6521		ret = is_extent_unchanged(sctx, path, key);
6522		if (ret < 0)
6523			goto out;
6524		if (ret) {
6525			ret = 0;
6526			goto out_hole;
6527		}
6528	} else {
6529		struct btrfs_file_extent_item *ei;
6530		u8 type;
6531
6532		ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
6533				    struct btrfs_file_extent_item);
6534		type = btrfs_file_extent_type(path->nodes[0], ei);
6535		if (type == BTRFS_FILE_EXTENT_PREALLOC ||
6536		    type == BTRFS_FILE_EXTENT_REG) {
6537			/*
6538			 * The send spec does not have a prealloc command yet,
6539			 * so just leave a hole for prealloc'ed extents until
6540			 * we have enough commands queued up to justify rev'ing
6541			 * the send spec.
6542			 */
6543			if (type == BTRFS_FILE_EXTENT_PREALLOC) {
6544				ret = 0;
6545				goto out;
6546			}
6547
6548			/* Have a hole, just skip it. */
6549			if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
6550				ret = 0;
6551				goto out;
6552			}
6553		}
6554	}
6555
6556	ret = find_extent_clone(sctx, path, key->objectid, key->offset,
6557			sctx->cur_inode_size, &found_clone);
6558	if (ret != -ENOENT && ret < 0)
6559		goto out;
6560
6561	ret = send_write_or_clone(sctx, path, key, found_clone);
6562	if (ret)
6563		goto out;
6564out_hole:
6565	ret = maybe_send_hole(sctx, path, key);
6566out:
6567	return ret;
6568}
6569
6570static int process_all_extents(struct send_ctx *sctx)
6571{
6572	int ret = 0;
6573	int iter_ret = 0;
6574	struct btrfs_root *root;
6575	struct btrfs_path *path;
6576	struct btrfs_key key;
6577	struct btrfs_key found_key;
6578
6579	root = sctx->send_root;
6580	path = alloc_path_for_send();
6581	if (!path)
6582		return -ENOMEM;
6583
6584	key.objectid = sctx->cmp_key->objectid;
6585	key.type = BTRFS_EXTENT_DATA_KEY;
6586	key.offset = 0;
6587	btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
6588		if (found_key.objectid != key.objectid ||
6589		    found_key.type != key.type) {
6590			ret = 0;
6591			break;
6592		}
6593
6594		ret = process_extent(sctx, path, &found_key);
6595		if (ret < 0)
6596			break;
6597	}
6598	/* Catch error found during iteration */
6599	if (iter_ret < 0)
6600		ret = iter_ret;
6601
6602	btrfs_free_path(path);
6603	return ret;
6604}
6605
6606static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
6607					   int *pending_move,
6608					   int *refs_processed)
6609{
6610	int ret = 0;
6611
6612	if (sctx->cur_ino == 0)
6613		goto out;
6614	if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
6615	    sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
6616		goto out;
6617	if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
6618		goto out;
6619
6620	ret = process_recorded_refs(sctx, pending_move);
6621	if (ret < 0)
6622		goto out;
6623
6624	*refs_processed = 1;
6625out:
6626	return ret;
6627}
6628
6629static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
6630{
6631	int ret = 0;
6632	struct btrfs_inode_info info;
6633	u64 left_mode;
6634	u64 left_uid;
6635	u64 left_gid;
6636	u64 left_fileattr;
6637	u64 right_mode;
6638	u64 right_uid;
6639	u64 right_gid;
6640	u64 right_fileattr;
6641	int need_chmod = 0;
6642	int need_chown = 0;
6643	bool need_fileattr = false;
6644	int need_truncate = 1;
6645	int pending_move = 0;
6646	int refs_processed = 0;
6647
6648	if (sctx->ignore_cur_inode)
6649		return 0;
6650
6651	ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
6652					      &refs_processed);
6653	if (ret < 0)
6654		goto out;
6655
6656	/*
6657	 * We have processed the refs and thus need to advance send_progress.
6658	 * Now, calls to get_cur_xxx will take the updated refs of the current
6659	 * inode into account.
6660	 *
6661	 * On the other hand, if our current inode is a directory and couldn't
6662	 * be moved/renamed because its parent was renamed/moved too and it has
6663	 * a higher inode number, we can only move/rename our current inode
6664	 * after we moved/renamed its parent. Therefore in this case operate on
6665	 * the old path (pre move/rename) of our current inode, and the
6666	 * move/rename will be performed later.
6667	 */
6668	if (refs_processed && !pending_move)
6669		sctx->send_progress = sctx->cur_ino + 1;
6670
6671	if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
6672		goto out;
6673	if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
6674		goto out;
6675	ret = get_inode_info(sctx->send_root, sctx->cur_ino, &info);
6676	if (ret < 0)
6677		goto out;
6678	left_mode = info.mode;
6679	left_uid = info.uid;
6680	left_gid = info.gid;
6681	left_fileattr = info.fileattr;
6682
6683	if (!sctx->parent_root || sctx->cur_inode_new) {
6684		need_chown = 1;
6685		if (!S_ISLNK(sctx->cur_inode_mode))
6686			need_chmod = 1;
6687		if (sctx->cur_inode_next_write_offset == sctx->cur_inode_size)
6688			need_truncate = 0;
6689	} else {
6690		u64 old_size;
6691
6692		ret = get_inode_info(sctx->parent_root, sctx->cur_ino, &info);
6693		if (ret < 0)
6694			goto out;
6695		old_size = info.size;
6696		right_mode = info.mode;
6697		right_uid = info.uid;
6698		right_gid = info.gid;
6699		right_fileattr = info.fileattr;
6700
6701		if (left_uid != right_uid || left_gid != right_gid)
6702			need_chown = 1;
6703		if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
6704			need_chmod = 1;
6705		if (!S_ISLNK(sctx->cur_inode_mode) && left_fileattr != right_fileattr)
6706			need_fileattr = true;
6707		if ((old_size == sctx->cur_inode_size) ||
6708		    (sctx->cur_inode_size > old_size &&
6709		     sctx->cur_inode_next_write_offset == sctx->cur_inode_size))
6710			need_truncate = 0;
6711	}
6712
6713	if (S_ISREG(sctx->cur_inode_mode)) {
6714		if (need_send_hole(sctx)) {
6715			if (sctx->cur_inode_last_extent == (u64)-1 ||
6716			    sctx->cur_inode_last_extent <
6717			    sctx->cur_inode_size) {
6718				ret = get_last_extent(sctx, (u64)-1);
6719				if (ret)
6720					goto out;
6721			}
6722			if (sctx->cur_inode_last_extent < sctx->cur_inode_size) {
6723				ret = range_is_hole_in_parent(sctx,
6724						      sctx->cur_inode_last_extent,
6725						      sctx->cur_inode_size);
6726				if (ret < 0) {
6727					goto out;
6728				} else if (ret == 0) {
6729					ret = send_hole(sctx, sctx->cur_inode_size);
6730					if (ret < 0)
6731						goto out;
6732				} else {
6733					/* Range is already a hole, skip. */
6734					ret = 0;
6735				}
6736			}
6737		}
6738		if (need_truncate) {
6739			ret = send_truncate(sctx, sctx->cur_ino,
6740					    sctx->cur_inode_gen,
6741					    sctx->cur_inode_size);
6742			if (ret < 0)
6743				goto out;
6744		}
6745	}
6746
6747	if (need_chown) {
6748		ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6749				left_uid, left_gid);
6750		if (ret < 0)
6751			goto out;
6752	}
6753	if (need_chmod) {
6754		ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6755				left_mode);
6756		if (ret < 0)
6757			goto out;
6758	}
6759	if (need_fileattr) {
6760		ret = send_fileattr(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6761				    left_fileattr);
6762		if (ret < 0)
6763			goto out;
6764	}
6765
6766	if (proto_cmd_ok(sctx, BTRFS_SEND_C_ENABLE_VERITY)
6767	    && sctx->cur_inode_needs_verity) {
6768		ret = process_verity(sctx);
6769		if (ret < 0)
6770			goto out;
6771	}
6772
6773	ret = send_capabilities(sctx);
6774	if (ret < 0)
6775		goto out;
6776
6777	/*
6778	 * If other directory inodes depended on our current directory
6779	 * inode's move/rename, now do their move/rename operations.
6780	 */
6781	if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
6782		ret = apply_children_dir_moves(sctx);
6783		if (ret)
6784			goto out;
6785		/*
6786		 * Need to send that every time, no matter if it actually
6787		 * changed between the two trees as we have done changes to
6788		 * the inode before. If our inode is a directory and it's
6789		 * waiting to be moved/renamed, we will send its utimes when
6790		 * it's moved/renamed, therefore we don't need to do it here.
6791		 */
6792		sctx->send_progress = sctx->cur_ino + 1;
6793
6794		/*
6795		 * If the current inode is a non-empty directory, delay issuing
6796		 * the utimes command for it, as it's very likely we have inodes
6797		 * with an higher number inside it. We want to issue the utimes
6798		 * command only after adding all dentries to it.
6799		 */
6800		if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_size > 0)
6801			ret = cache_dir_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6802		else
6803			ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6804
6805		if (ret < 0)
6806			goto out;
6807	}
6808
6809out:
6810	if (!ret)
6811		ret = trim_dir_utimes_cache(sctx);
6812
6813	return ret;
6814}
6815
6816static void close_current_inode(struct send_ctx *sctx)
6817{
6818	u64 i_size;
6819
6820	if (sctx->cur_inode == NULL)
6821		return;
6822
6823	i_size = i_size_read(sctx->cur_inode);
6824
6825	/*
6826	 * If we are doing an incremental send, we may have extents between the
6827	 * last processed extent and the i_size that have not been processed
6828	 * because they haven't changed but we may have read some of their pages
6829	 * through readahead, see the comments at send_extent_data().
6830	 */
6831	if (sctx->clean_page_cache && sctx->page_cache_clear_start < i_size)
6832		truncate_inode_pages_range(&sctx->cur_inode->i_data,
6833					   sctx->page_cache_clear_start,
6834					   round_up(i_size, PAGE_SIZE) - 1);
6835
6836	iput(sctx->cur_inode);
6837	sctx->cur_inode = NULL;
6838}
6839
6840static int changed_inode(struct send_ctx *sctx,
6841			 enum btrfs_compare_tree_result result)
6842{
6843	int ret = 0;
6844	struct btrfs_key *key = sctx->cmp_key;
6845	struct btrfs_inode_item *left_ii = NULL;
6846	struct btrfs_inode_item *right_ii = NULL;
6847	u64 left_gen = 0;
6848	u64 right_gen = 0;
6849
6850	close_current_inode(sctx);
6851
6852	sctx->cur_ino = key->objectid;
6853	sctx->cur_inode_new_gen = false;
6854	sctx->cur_inode_last_extent = (u64)-1;
6855	sctx->cur_inode_next_write_offset = 0;
6856	sctx->ignore_cur_inode = false;
6857
6858	/*
6859	 * Set send_progress to current inode. This will tell all get_cur_xxx
6860	 * functions that the current inode's refs are not updated yet. Later,
6861	 * when process_recorded_refs is finished, it is set to cur_ino + 1.
6862	 */
6863	sctx->send_progress = sctx->cur_ino;
6864
6865	if (result == BTRFS_COMPARE_TREE_NEW ||
6866	    result == BTRFS_COMPARE_TREE_CHANGED) {
6867		left_ii = btrfs_item_ptr(sctx->left_path->nodes[0],
6868				sctx->left_path->slots[0],
6869				struct btrfs_inode_item);
6870		left_gen = btrfs_inode_generation(sctx->left_path->nodes[0],
6871				left_ii);
6872	} else {
6873		right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6874				sctx->right_path->slots[0],
6875				struct btrfs_inode_item);
6876		right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6877				right_ii);
6878	}
6879	if (result == BTRFS_COMPARE_TREE_CHANGED) {
6880		right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6881				sctx->right_path->slots[0],
6882				struct btrfs_inode_item);
6883
6884		right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6885				right_ii);
6886
6887		/*
6888		 * The cur_ino = root dir case is special here. We can't treat
6889		 * the inode as deleted+reused because it would generate a
6890		 * stream that tries to delete/mkdir the root dir.
6891		 */
6892		if (left_gen != right_gen &&
6893		    sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6894			sctx->cur_inode_new_gen = true;
6895	}
6896
6897	/*
6898	 * Normally we do not find inodes with a link count of zero (orphans)
6899	 * because the most common case is to create a snapshot and use it
6900	 * for a send operation. However other less common use cases involve
6901	 * using a subvolume and send it after turning it to RO mode just
6902	 * after deleting all hard links of a file while holding an open
6903	 * file descriptor against it or turning a RO snapshot into RW mode,
6904	 * keep an open file descriptor against a file, delete it and then
6905	 * turn the snapshot back to RO mode before using it for a send
6906	 * operation. The former is what the receiver operation does.
6907	 * Therefore, if we want to send these snapshots soon after they're
6908	 * received, we need to handle orphan inodes as well. Moreover, orphans
6909	 * can appear not only in the send snapshot but also in the parent
6910	 * snapshot. Here are several cases:
6911	 *
6912	 * Case 1: BTRFS_COMPARE_TREE_NEW
6913	 *       |  send snapshot  | action
6914	 * --------------------------------
6915	 * nlink |        0        | ignore
6916	 *
6917	 * Case 2: BTRFS_COMPARE_TREE_DELETED
6918	 *       | parent snapshot | action
6919	 * ----------------------------------
6920	 * nlink |        0        | as usual
6921	 * Note: No unlinks will be sent because there're no paths for it.
6922	 *
6923	 * Case 3: BTRFS_COMPARE_TREE_CHANGED
6924	 *           |       | parent snapshot | send snapshot | action
6925	 * -----------------------------------------------------------------------
6926	 * subcase 1 | nlink |        0        |       0       | ignore
6927	 * subcase 2 | nlink |       >0        |       0       | new_gen(deletion)
6928	 * subcase 3 | nlink |        0        |      >0       | new_gen(creation)
6929	 *
6930	 */
6931	if (result == BTRFS_COMPARE_TREE_NEW) {
6932		if (btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii) == 0) {
6933			sctx->ignore_cur_inode = true;
6934			goto out;
6935		}
6936		sctx->cur_inode_gen = left_gen;
6937		sctx->cur_inode_new = true;
6938		sctx->cur_inode_deleted = false;
6939		sctx->cur_inode_size = btrfs_inode_size(
6940				sctx->left_path->nodes[0], left_ii);
6941		sctx->cur_inode_mode = btrfs_inode_mode(
6942				sctx->left_path->nodes[0], left_ii);
6943		sctx->cur_inode_rdev = btrfs_inode_rdev(
6944				sctx->left_path->nodes[0], left_ii);
6945		if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6946			ret = send_create_inode_if_needed(sctx);
6947	} else if (result == BTRFS_COMPARE_TREE_DELETED) {
6948		sctx->cur_inode_gen = right_gen;
6949		sctx->cur_inode_new = false;
6950		sctx->cur_inode_deleted = true;
6951		sctx->cur_inode_size = btrfs_inode_size(
6952				sctx->right_path->nodes[0], right_ii);
6953		sctx->cur_inode_mode = btrfs_inode_mode(
6954				sctx->right_path->nodes[0], right_ii);
6955	} else if (result == BTRFS_COMPARE_TREE_CHANGED) {
6956		u32 new_nlinks, old_nlinks;
6957
6958		new_nlinks = btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii);
6959		old_nlinks = btrfs_inode_nlink(sctx->right_path->nodes[0], right_ii);
6960		if (new_nlinks == 0 && old_nlinks == 0) {
6961			sctx->ignore_cur_inode = true;
6962			goto out;
6963		} else if (new_nlinks == 0 || old_nlinks == 0) {
6964			sctx->cur_inode_new_gen = 1;
6965		}
6966		/*
6967		 * We need to do some special handling in case the inode was
6968		 * reported as changed with a changed generation number. This
6969		 * means that the original inode was deleted and new inode
6970		 * reused the same inum. So we have to treat the old inode as
6971		 * deleted and the new one as new.
6972		 */
6973		if (sctx->cur_inode_new_gen) {
6974			/*
6975			 * First, process the inode as if it was deleted.
6976			 */
6977			if (old_nlinks > 0) {
6978				sctx->cur_inode_gen = right_gen;
6979				sctx->cur_inode_new = false;
6980				sctx->cur_inode_deleted = true;
6981				sctx->cur_inode_size = btrfs_inode_size(
6982						sctx->right_path->nodes[0], right_ii);
6983				sctx->cur_inode_mode = btrfs_inode_mode(
6984						sctx->right_path->nodes[0], right_ii);
6985				ret = process_all_refs(sctx,
6986						BTRFS_COMPARE_TREE_DELETED);
6987				if (ret < 0)
6988					goto out;
6989			}
6990
6991			/*
6992			 * Now process the inode as if it was new.
6993			 */
6994			if (new_nlinks > 0) {
6995				sctx->cur_inode_gen = left_gen;
6996				sctx->cur_inode_new = true;
6997				sctx->cur_inode_deleted = false;
6998				sctx->cur_inode_size = btrfs_inode_size(
6999						sctx->left_path->nodes[0],
7000						left_ii);
7001				sctx->cur_inode_mode = btrfs_inode_mode(
7002						sctx->left_path->nodes[0],
7003						left_ii);
7004				sctx->cur_inode_rdev = btrfs_inode_rdev(
7005						sctx->left_path->nodes[0],
7006						left_ii);
7007				ret = send_create_inode_if_needed(sctx);
7008				if (ret < 0)
7009					goto out;
7010
7011				ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW);
7012				if (ret < 0)
7013					goto out;
7014				/*
7015				 * Advance send_progress now as we did not get
7016				 * into process_recorded_refs_if_needed in the
7017				 * new_gen case.
7018				 */
7019				sctx->send_progress = sctx->cur_ino + 1;
7020
7021				/*
7022				 * Now process all extents and xattrs of the
7023				 * inode as if they were all new.
7024				 */
7025				ret = process_all_extents(sctx);
7026				if (ret < 0)
7027					goto out;
7028				ret = process_all_new_xattrs(sctx);
7029				if (ret < 0)
7030					goto out;
7031			}
7032		} else {
7033			sctx->cur_inode_gen = left_gen;
7034			sctx->cur_inode_new = false;
7035			sctx->cur_inode_new_gen = false;
7036			sctx->cur_inode_deleted = false;
7037			sctx->cur_inode_size = btrfs_inode_size(
7038					sctx->left_path->nodes[0], left_ii);
7039			sctx->cur_inode_mode = btrfs_inode_mode(
7040					sctx->left_path->nodes[0], left_ii);
7041		}
7042	}
7043
7044out:
7045	return ret;
7046}
7047
7048/*
7049 * We have to process new refs before deleted refs, but compare_trees gives us
7050 * the new and deleted refs mixed. To fix this, we record the new/deleted refs
7051 * first and later process them in process_recorded_refs.
7052 * For the cur_inode_new_gen case, we skip recording completely because
7053 * changed_inode did already initiate processing of refs. The reason for this is
7054 * that in this case, compare_tree actually compares the refs of 2 different
7055 * inodes. To fix this, process_all_refs is used in changed_inode to handle all
7056 * refs of the right tree as deleted and all refs of the left tree as new.
7057 */
7058static int changed_ref(struct send_ctx *sctx,
7059		       enum btrfs_compare_tree_result result)
7060{
7061	int ret = 0;
7062
7063	if (sctx->cur_ino != sctx->cmp_key->objectid) {
7064		inconsistent_snapshot_error(sctx, result, "reference");
7065		return -EIO;
7066	}
7067
7068	if (!sctx->cur_inode_new_gen &&
7069	    sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) {
7070		if (result == BTRFS_COMPARE_TREE_NEW)
7071			ret = record_new_ref(sctx);
7072		else if (result == BTRFS_COMPARE_TREE_DELETED)
7073			ret = record_deleted_ref(sctx);
7074		else if (result == BTRFS_COMPARE_TREE_CHANGED)
7075			ret = record_changed_ref(sctx);
7076	}
7077
7078	return ret;
7079}
7080
7081/*
7082 * Process new/deleted/changed xattrs. We skip processing in the
7083 * cur_inode_new_gen case because changed_inode did already initiate processing
7084 * of xattrs. The reason is the same as in changed_ref
7085 */
7086static int changed_xattr(struct send_ctx *sctx,
7087			 enum btrfs_compare_tree_result result)
7088{
7089	int ret = 0;
7090
7091	if (sctx->cur_ino != sctx->cmp_key->objectid) {
7092		inconsistent_snapshot_error(sctx, result, "xattr");
7093		return -EIO;
7094	}
7095
7096	if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7097		if (result == BTRFS_COMPARE_TREE_NEW)
7098			ret = process_new_xattr(sctx);
7099		else if (result == BTRFS_COMPARE_TREE_DELETED)
7100			ret = process_deleted_xattr(sctx);
7101		else if (result == BTRFS_COMPARE_TREE_CHANGED)
7102			ret = process_changed_xattr(sctx);
7103	}
7104
7105	return ret;
7106}
7107
7108/*
7109 * Process new/deleted/changed extents. We skip processing in the
7110 * cur_inode_new_gen case because changed_inode did already initiate processing
7111 * of extents. The reason is the same as in changed_ref
7112 */
7113static int changed_extent(struct send_ctx *sctx,
7114			  enum btrfs_compare_tree_result result)
7115{
7116	int ret = 0;
7117
7118	/*
7119	 * We have found an extent item that changed without the inode item
7120	 * having changed. This can happen either after relocation (where the
7121	 * disk_bytenr of an extent item is replaced at
7122	 * relocation.c:replace_file_extents()) or after deduplication into a
7123	 * file in both the parent and send snapshots (where an extent item can
7124	 * get modified or replaced with a new one). Note that deduplication
7125	 * updates the inode item, but it only changes the iversion (sequence
7126	 * field in the inode item) of the inode, so if a file is deduplicated
7127	 * the same amount of times in both the parent and send snapshots, its
7128	 * iversion becomes the same in both snapshots, whence the inode item is
7129	 * the same on both snapshots.
7130	 */
7131	if (sctx->cur_ino != sctx->cmp_key->objectid)
7132		return 0;
7133
7134	if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7135		if (result != BTRFS_COMPARE_TREE_DELETED)
7136			ret = process_extent(sctx, sctx->left_path,
7137					sctx->cmp_key);
7138	}
7139
7140	return ret;
7141}
7142
7143static int changed_verity(struct send_ctx *sctx, enum btrfs_compare_tree_result result)
7144{
7145	int ret = 0;
7146
7147	if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7148		if (result == BTRFS_COMPARE_TREE_NEW)
7149			sctx->cur_inode_needs_verity = true;
7150	}
7151	return ret;
7152}
7153
7154static int dir_changed(struct send_ctx *sctx, u64 dir)
7155{
7156	u64 orig_gen, new_gen;
7157	int ret;
7158
7159	ret = get_inode_gen(sctx->send_root, dir, &new_gen);
7160	if (ret)
7161		return ret;
7162
7163	ret = get_inode_gen(sctx->parent_root, dir, &orig_gen);
7164	if (ret)
7165		return ret;
7166
7167	return (orig_gen != new_gen) ? 1 : 0;
7168}
7169
7170static int compare_refs(struct send_ctx *sctx, struct btrfs_path *path,
7171			struct btrfs_key *key)
7172{
7173	struct btrfs_inode_extref *extref;
7174	struct extent_buffer *leaf;
7175	u64 dirid = 0, last_dirid = 0;
7176	unsigned long ptr;
7177	u32 item_size;
7178	u32 cur_offset = 0;
7179	int ref_name_len;
7180	int ret = 0;
7181
7182	/* Easy case, just check this one dirid */
7183	if (key->type == BTRFS_INODE_REF_KEY) {
7184		dirid = key->offset;
7185
7186		ret = dir_changed(sctx, dirid);
7187		goto out;
7188	}
7189
7190	leaf = path->nodes[0];
7191	item_size = btrfs_item_size(leaf, path->slots[0]);
7192	ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
7193	while (cur_offset < item_size) {
7194		extref = (struct btrfs_inode_extref *)(ptr +
7195						       cur_offset);
7196		dirid = btrfs_inode_extref_parent(leaf, extref);
7197		ref_name_len = btrfs_inode_extref_name_len(leaf, extref);
7198		cur_offset += ref_name_len + sizeof(*extref);
7199		if (dirid == last_dirid)
7200			continue;
7201		ret = dir_changed(sctx, dirid);
7202		if (ret)
7203			break;
7204		last_dirid = dirid;
7205	}
7206out:
7207	return ret;
7208}
7209
7210/*
7211 * Updates compare related fields in sctx and simply forwards to the actual
7212 * changed_xxx functions.
7213 */
7214static int changed_cb(struct btrfs_path *left_path,
7215		      struct btrfs_path *right_path,
7216		      struct btrfs_key *key,
7217		      enum btrfs_compare_tree_result result,
7218		      struct send_ctx *sctx)
7219{
7220	int ret = 0;
7221
7222	/*
7223	 * We can not hold the commit root semaphore here. This is because in
7224	 * the case of sending and receiving to the same filesystem, using a
7225	 * pipe, could result in a deadlock:
7226	 *
7227	 * 1) The task running send blocks on the pipe because it's full;
7228	 *
7229	 * 2) The task running receive, which is the only consumer of the pipe,
7230	 *    is waiting for a transaction commit (for example due to a space
7231	 *    reservation when doing a write or triggering a transaction commit
7232	 *    when creating a subvolume);
7233	 *
7234	 * 3) The transaction is waiting to write lock the commit root semaphore,
7235	 *    but can not acquire it since it's being held at 1).
7236	 *
7237	 * Down this call chain we write to the pipe through kernel_write().
7238	 * The same type of problem can also happen when sending to a file that
7239	 * is stored in the same filesystem - when reserving space for a write
7240	 * into the file, we can trigger a transaction commit.
7241	 *
7242	 * Our caller has supplied us with clones of leaves from the send and
7243	 * parent roots, so we're safe here from a concurrent relocation and
7244	 * further reallocation of metadata extents while we are here. Below we
7245	 * also assert that the leaves are clones.
7246	 */
7247	lockdep_assert_not_held(&sctx->send_root->fs_info->commit_root_sem);
7248
7249	/*
7250	 * We always have a send root, so left_path is never NULL. We will not
7251	 * have a leaf when we have reached the end of the send root but have
7252	 * not yet reached the end of the parent root.
7253	 */
7254	if (left_path->nodes[0])
7255		ASSERT(test_bit(EXTENT_BUFFER_UNMAPPED,
7256				&left_path->nodes[0]->bflags));
7257	/*
7258	 * When doing a full send we don't have a parent root, so right_path is
7259	 * NULL. When doing an incremental send, we may have reached the end of
7260	 * the parent root already, so we don't have a leaf at right_path.
7261	 */
7262	if (right_path && right_path->nodes[0])
7263		ASSERT(test_bit(EXTENT_BUFFER_UNMAPPED,
7264				&right_path->nodes[0]->bflags));
7265
7266	if (result == BTRFS_COMPARE_TREE_SAME) {
7267		if (key->type == BTRFS_INODE_REF_KEY ||
7268		    key->type == BTRFS_INODE_EXTREF_KEY) {
7269			ret = compare_refs(sctx, left_path, key);
7270			if (!ret)
7271				return 0;
7272			if (ret < 0)
7273				return ret;
7274		} else if (key->type == BTRFS_EXTENT_DATA_KEY) {
7275			return maybe_send_hole(sctx, left_path, key);
7276		} else {
7277			return 0;
7278		}
7279		result = BTRFS_COMPARE_TREE_CHANGED;
7280		ret = 0;
7281	}
7282
7283	sctx->left_path = left_path;
7284	sctx->right_path = right_path;
7285	sctx->cmp_key = key;
7286
7287	ret = finish_inode_if_needed(sctx, 0);
7288	if (ret < 0)
7289		goto out;
7290
7291	/* Ignore non-FS objects */
7292	if (key->objectid == BTRFS_FREE_INO_OBJECTID ||
7293	    key->objectid == BTRFS_FREE_SPACE_OBJECTID)
7294		goto out;
7295
7296	if (key->type == BTRFS_INODE_ITEM_KEY) {
7297		ret = changed_inode(sctx, result);
7298	} else if (!sctx->ignore_cur_inode) {
7299		if (key->type == BTRFS_INODE_REF_KEY ||
7300		    key->type == BTRFS_INODE_EXTREF_KEY)
7301			ret = changed_ref(sctx, result);
7302		else if (key->type == BTRFS_XATTR_ITEM_KEY)
7303			ret = changed_xattr(sctx, result);
7304		else if (key->type == BTRFS_EXTENT_DATA_KEY)
7305			ret = changed_extent(sctx, result);
7306		else if (key->type == BTRFS_VERITY_DESC_ITEM_KEY &&
7307			 key->offset == 0)
7308			ret = changed_verity(sctx, result);
7309	}
7310
7311out:
7312	return ret;
7313}
7314
7315static int search_key_again(const struct send_ctx *sctx,
7316			    struct btrfs_root *root,
7317			    struct btrfs_path *path,
7318			    const struct btrfs_key *key)
7319{
7320	int ret;
7321
7322	if (!path->need_commit_sem)
7323		lockdep_assert_held_read(&root->fs_info->commit_root_sem);
7324
7325	/*
7326	 * Roots used for send operations are readonly and no one can add,
7327	 * update or remove keys from them, so we should be able to find our
7328	 * key again. The only exception is deduplication, which can operate on
7329	 * readonly roots and add, update or remove keys to/from them - but at
7330	 * the moment we don't allow it to run in parallel with send.
7331	 */
7332	ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7333	ASSERT(ret <= 0);
7334	if (ret > 0) {
7335		btrfs_print_tree(path->nodes[path->lowest_level], false);
7336		btrfs_err(root->fs_info,
7337"send: key (%llu %u %llu) not found in %s root %llu, lowest_level %d, slot %d",
7338			  key->objectid, key->type, key->offset,
7339			  (root == sctx->parent_root ? "parent" : "send"),
7340			  root->root_key.objectid, path->lowest_level,
7341			  path->slots[path->lowest_level]);
7342		return -EUCLEAN;
7343	}
7344
7345	return ret;
7346}
7347
7348static int full_send_tree(struct send_ctx *sctx)
7349{
7350	int ret;
7351	struct btrfs_root *send_root = sctx->send_root;
7352	struct btrfs_key key;
7353	struct btrfs_fs_info *fs_info = send_root->fs_info;
7354	struct btrfs_path *path;
7355
7356	path = alloc_path_for_send();
7357	if (!path)
7358		return -ENOMEM;
7359	path->reada = READA_FORWARD_ALWAYS;
7360
7361	key.objectid = BTRFS_FIRST_FREE_OBJECTID;
7362	key.type = BTRFS_INODE_ITEM_KEY;
7363	key.offset = 0;
7364
7365	down_read(&fs_info->commit_root_sem);
7366	sctx->last_reloc_trans = fs_info->last_reloc_trans;
7367	up_read(&fs_info->commit_root_sem);
7368
7369	ret = btrfs_search_slot_for_read(send_root, &key, path, 1, 0);
7370	if (ret < 0)
7371		goto out;
7372	if (ret)
7373		goto out_finish;
7374
7375	while (1) {
7376		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
7377
7378		ret = changed_cb(path, NULL, &key,
7379				 BTRFS_COMPARE_TREE_NEW, sctx);
7380		if (ret < 0)
7381			goto out;
7382
7383		down_read(&fs_info->commit_root_sem);
7384		if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
7385			sctx->last_reloc_trans = fs_info->last_reloc_trans;
7386			up_read(&fs_info->commit_root_sem);
7387			/*
7388			 * A transaction used for relocating a block group was
7389			 * committed or is about to finish its commit. Release
7390			 * our path (leaf) and restart the search, so that we
7391			 * avoid operating on any file extent items that are
7392			 * stale, with a disk_bytenr that reflects a pre
7393			 * relocation value. This way we avoid as much as
7394			 * possible to fallback to regular writes when checking
7395			 * if we can clone file ranges.
7396			 */
7397			btrfs_release_path(path);
7398			ret = search_key_again(sctx, send_root, path, &key);
7399			if (ret < 0)
7400				goto out;
7401		} else {
7402			up_read(&fs_info->commit_root_sem);
7403		}
7404
7405		ret = btrfs_next_item(send_root, path);
7406		if (ret < 0)
7407			goto out;
7408		if (ret) {
7409			ret  = 0;
7410			break;
7411		}
7412	}
7413
7414out_finish:
7415	ret = finish_inode_if_needed(sctx, 1);
7416
7417out:
7418	btrfs_free_path(path);
7419	return ret;
7420}
7421
7422static int replace_node_with_clone(struct btrfs_path *path, int level)
7423{
7424	struct extent_buffer *clone;
7425
7426	clone = btrfs_clone_extent_buffer(path->nodes[level]);
7427	if (!clone)
7428		return -ENOMEM;
7429
7430	free_extent_buffer(path->nodes[level]);
7431	path->nodes[level] = clone;
7432
7433	return 0;
7434}
7435
7436static int tree_move_down(struct btrfs_path *path, int *level, u64 reada_min_gen)
7437{
7438	struct extent_buffer *eb;
7439	struct extent_buffer *parent = path->nodes[*level];
7440	int slot = path->slots[*level];
7441	const int nritems = btrfs_header_nritems(parent);
7442	u64 reada_max;
7443	u64 reada_done = 0;
7444
7445	lockdep_assert_held_read(&parent->fs_info->commit_root_sem);
7446	ASSERT(*level != 0);
7447
 
7448	eb = btrfs_read_node_slot(parent, slot);
7449	if (IS_ERR(eb))
7450		return PTR_ERR(eb);
7451
7452	/*
7453	 * Trigger readahead for the next leaves we will process, so that it is
7454	 * very likely that when we need them they are already in memory and we
7455	 * will not block on disk IO. For nodes we only do readahead for one,
7456	 * since the time window between processing nodes is typically larger.
7457	 */
7458	reada_max = (*level == 1 ? SZ_128K : eb->fs_info->nodesize);
7459
7460	for (slot++; slot < nritems && reada_done < reada_max; slot++) {
7461		if (btrfs_node_ptr_generation(parent, slot) > reada_min_gen) {
7462			btrfs_readahead_node_child(parent, slot);
7463			reada_done += eb->fs_info->nodesize;
7464		}
7465	}
7466
7467	path->nodes[*level - 1] = eb;
7468	path->slots[*level - 1] = 0;
7469	(*level)--;
7470
7471	if (*level == 0)
7472		return replace_node_with_clone(path, 0);
7473
7474	return 0;
7475}
7476
7477static int tree_move_next_or_upnext(struct btrfs_path *path,
7478				    int *level, int root_level)
7479{
7480	int ret = 0;
7481	int nritems;
7482	nritems = btrfs_header_nritems(path->nodes[*level]);
7483
7484	path->slots[*level]++;
7485
7486	while (path->slots[*level] >= nritems) {
7487		if (*level == root_level) {
7488			path->slots[*level] = nritems - 1;
7489			return -1;
7490		}
7491
7492		/* move upnext */
7493		path->slots[*level] = 0;
7494		free_extent_buffer(path->nodes[*level]);
7495		path->nodes[*level] = NULL;
7496		(*level)++;
7497		path->slots[*level]++;
7498
7499		nritems = btrfs_header_nritems(path->nodes[*level]);
7500		ret = 1;
7501	}
7502	return ret;
7503}
7504
7505/*
7506 * Returns 1 if it had to move up and next. 0 is returned if it moved only next
7507 * or down.
7508 */
7509static int tree_advance(struct btrfs_path *path,
7510			int *level, int root_level,
7511			int allow_down,
7512			struct btrfs_key *key,
7513			u64 reada_min_gen)
7514{
7515	int ret;
7516
7517	if (*level == 0 || !allow_down) {
7518		ret = tree_move_next_or_upnext(path, level, root_level);
7519	} else {
7520		ret = tree_move_down(path, level, reada_min_gen);
7521	}
7522
7523	/*
7524	 * Even if we have reached the end of a tree, ret is -1, update the key
7525	 * anyway, so that in case we need to restart due to a block group
7526	 * relocation, we can assert that the last key of the root node still
7527	 * exists in the tree.
7528	 */
7529	if (*level == 0)
7530		btrfs_item_key_to_cpu(path->nodes[*level], key,
7531				      path->slots[*level]);
7532	else
7533		btrfs_node_key_to_cpu(path->nodes[*level], key,
7534				      path->slots[*level]);
7535
7536	return ret;
7537}
7538
7539static int tree_compare_item(struct btrfs_path *left_path,
7540			     struct btrfs_path *right_path,
7541			     char *tmp_buf)
7542{
7543	int cmp;
7544	int len1, len2;
7545	unsigned long off1, off2;
7546
7547	len1 = btrfs_item_size(left_path->nodes[0], left_path->slots[0]);
7548	len2 = btrfs_item_size(right_path->nodes[0], right_path->slots[0]);
7549	if (len1 != len2)
7550		return 1;
7551
7552	off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
7553	off2 = btrfs_item_ptr_offset(right_path->nodes[0],
7554				right_path->slots[0]);
7555
7556	read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
7557
7558	cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
7559	if (cmp)
7560		return 1;
7561	return 0;
7562}
7563
7564/*
7565 * A transaction used for relocating a block group was committed or is about to
7566 * finish its commit. Release our paths and restart the search, so that we are
7567 * not using stale extent buffers:
7568 *
7569 * 1) For levels > 0, we are only holding references of extent buffers, without
7570 *    any locks on them, which does not prevent them from having been relocated
7571 *    and reallocated after the last time we released the commit root semaphore.
7572 *    The exception are the root nodes, for which we always have a clone, see
7573 *    the comment at btrfs_compare_trees();
7574 *
7575 * 2) For leaves, level 0, we are holding copies (clones) of extent buffers, so
7576 *    we are safe from the concurrent relocation and reallocation. However they
7577 *    can have file extent items with a pre relocation disk_bytenr value, so we
7578 *    restart the start from the current commit roots and clone the new leaves so
7579 *    that we get the post relocation disk_bytenr values. Not doing so, could
7580 *    make us clone the wrong data in case there are new extents using the old
7581 *    disk_bytenr that happen to be shared.
7582 */
7583static int restart_after_relocation(struct btrfs_path *left_path,
7584				    struct btrfs_path *right_path,
7585				    const struct btrfs_key *left_key,
7586				    const struct btrfs_key *right_key,
7587				    int left_level,
7588				    int right_level,
7589				    const struct send_ctx *sctx)
7590{
7591	int root_level;
7592	int ret;
7593
7594	lockdep_assert_held_read(&sctx->send_root->fs_info->commit_root_sem);
7595
7596	btrfs_release_path(left_path);
7597	btrfs_release_path(right_path);
7598
7599	/*
7600	 * Since keys can not be added or removed to/from our roots because they
7601	 * are readonly and we do not allow deduplication to run in parallel
7602	 * (which can add, remove or change keys), the layout of the trees should
7603	 * not change.
7604	 */
7605	left_path->lowest_level = left_level;
7606	ret = search_key_again(sctx, sctx->send_root, left_path, left_key);
7607	if (ret < 0)
7608		return ret;
7609
7610	right_path->lowest_level = right_level;
7611	ret = search_key_again(sctx, sctx->parent_root, right_path, right_key);
7612	if (ret < 0)
7613		return ret;
7614
7615	/*
7616	 * If the lowest level nodes are leaves, clone them so that they can be
7617	 * safely used by changed_cb() while not under the protection of the
7618	 * commit root semaphore, even if relocation and reallocation happens in
7619	 * parallel.
7620	 */
7621	if (left_level == 0) {
7622		ret = replace_node_with_clone(left_path, 0);
7623		if (ret < 0)
7624			return ret;
7625	}
7626
7627	if (right_level == 0) {
7628		ret = replace_node_with_clone(right_path, 0);
7629		if (ret < 0)
7630			return ret;
7631	}
7632
7633	/*
7634	 * Now clone the root nodes (unless they happen to be the leaves we have
7635	 * already cloned). This is to protect against concurrent snapshotting of
7636	 * the send and parent roots (see the comment at btrfs_compare_trees()).
7637	 */
7638	root_level = btrfs_header_level(sctx->send_root->commit_root);
7639	if (root_level > 0) {
7640		ret = replace_node_with_clone(left_path, root_level);
7641		if (ret < 0)
7642			return ret;
7643	}
7644
7645	root_level = btrfs_header_level(sctx->parent_root->commit_root);
7646	if (root_level > 0) {
7647		ret = replace_node_with_clone(right_path, root_level);
7648		if (ret < 0)
7649			return ret;
7650	}
7651
7652	return 0;
7653}
7654
7655/*
7656 * This function compares two trees and calls the provided callback for
7657 * every changed/new/deleted item it finds.
7658 * If shared tree blocks are encountered, whole subtrees are skipped, making
7659 * the compare pretty fast on snapshotted subvolumes.
7660 *
7661 * This currently works on commit roots only. As commit roots are read only,
7662 * we don't do any locking. The commit roots are protected with transactions.
7663 * Transactions are ended and rejoined when a commit is tried in between.
7664 *
7665 * This function checks for modifications done to the trees while comparing.
7666 * If it detects a change, it aborts immediately.
7667 */
7668static int btrfs_compare_trees(struct btrfs_root *left_root,
7669			struct btrfs_root *right_root, struct send_ctx *sctx)
7670{
7671	struct btrfs_fs_info *fs_info = left_root->fs_info;
7672	int ret;
7673	int cmp;
7674	struct btrfs_path *left_path = NULL;
7675	struct btrfs_path *right_path = NULL;
7676	struct btrfs_key left_key;
7677	struct btrfs_key right_key;
7678	char *tmp_buf = NULL;
7679	int left_root_level;
7680	int right_root_level;
7681	int left_level;
7682	int right_level;
7683	int left_end_reached = 0;
7684	int right_end_reached = 0;
7685	int advance_left = 0;
7686	int advance_right = 0;
7687	u64 left_blockptr;
7688	u64 right_blockptr;
7689	u64 left_gen;
7690	u64 right_gen;
7691	u64 reada_min_gen;
7692
7693	left_path = btrfs_alloc_path();
7694	if (!left_path) {
7695		ret = -ENOMEM;
7696		goto out;
7697	}
7698	right_path = btrfs_alloc_path();
7699	if (!right_path) {
7700		ret = -ENOMEM;
7701		goto out;
7702	}
7703
7704	tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
7705	if (!tmp_buf) {
7706		ret = -ENOMEM;
7707		goto out;
7708	}
7709
7710	left_path->search_commit_root = 1;
7711	left_path->skip_locking = 1;
7712	right_path->search_commit_root = 1;
7713	right_path->skip_locking = 1;
7714
7715	/*
7716	 * Strategy: Go to the first items of both trees. Then do
7717	 *
7718	 * If both trees are at level 0
7719	 *   Compare keys of current items
7720	 *     If left < right treat left item as new, advance left tree
7721	 *       and repeat
7722	 *     If left > right treat right item as deleted, advance right tree
7723	 *       and repeat
7724	 *     If left == right do deep compare of items, treat as changed if
7725	 *       needed, advance both trees and repeat
7726	 * If both trees are at the same level but not at level 0
7727	 *   Compare keys of current nodes/leafs
7728	 *     If left < right advance left tree and repeat
7729	 *     If left > right advance right tree and repeat
7730	 *     If left == right compare blockptrs of the next nodes/leafs
7731	 *       If they match advance both trees but stay at the same level
7732	 *         and repeat
7733	 *       If they don't match advance both trees while allowing to go
7734	 *         deeper and repeat
7735	 * If tree levels are different
7736	 *   Advance the tree that needs it and repeat
7737	 *
7738	 * Advancing a tree means:
7739	 *   If we are at level 0, try to go to the next slot. If that's not
7740	 *   possible, go one level up and repeat. Stop when we found a level
7741	 *   where we could go to the next slot. We may at this point be on a
7742	 *   node or a leaf.
7743	 *
7744	 *   If we are not at level 0 and not on shared tree blocks, go one
7745	 *   level deeper.
7746	 *
7747	 *   If we are not at level 0 and on shared tree blocks, go one slot to
7748	 *   the right if possible or go up and right.
7749	 */
7750
7751	down_read(&fs_info->commit_root_sem);
7752	left_level = btrfs_header_level(left_root->commit_root);
7753	left_root_level = left_level;
7754	/*
7755	 * We clone the root node of the send and parent roots to prevent races
7756	 * with snapshot creation of these roots. Snapshot creation COWs the
7757	 * root node of a tree, so after the transaction is committed the old
7758	 * extent can be reallocated while this send operation is still ongoing.
7759	 * So we clone them, under the commit root semaphore, to be race free.
7760	 */
7761	left_path->nodes[left_level] =
7762			btrfs_clone_extent_buffer(left_root->commit_root);
7763	if (!left_path->nodes[left_level]) {
7764		ret = -ENOMEM;
7765		goto out_unlock;
7766	}
7767
7768	right_level = btrfs_header_level(right_root->commit_root);
7769	right_root_level = right_level;
7770	right_path->nodes[right_level] =
7771			btrfs_clone_extent_buffer(right_root->commit_root);
7772	if (!right_path->nodes[right_level]) {
7773		ret = -ENOMEM;
7774		goto out_unlock;
7775	}
7776	/*
7777	 * Our right root is the parent root, while the left root is the "send"
7778	 * root. We know that all new nodes/leaves in the left root must have
7779	 * a generation greater than the right root's generation, so we trigger
7780	 * readahead for those nodes and leaves of the left root, as we know we
7781	 * will need to read them at some point.
7782	 */
7783	reada_min_gen = btrfs_header_generation(right_root->commit_root);
7784
7785	if (left_level == 0)
7786		btrfs_item_key_to_cpu(left_path->nodes[left_level],
7787				&left_key, left_path->slots[left_level]);
7788	else
7789		btrfs_node_key_to_cpu(left_path->nodes[left_level],
7790				&left_key, left_path->slots[left_level]);
7791	if (right_level == 0)
7792		btrfs_item_key_to_cpu(right_path->nodes[right_level],
7793				&right_key, right_path->slots[right_level]);
7794	else
7795		btrfs_node_key_to_cpu(right_path->nodes[right_level],
7796				&right_key, right_path->slots[right_level]);
7797
7798	sctx->last_reloc_trans = fs_info->last_reloc_trans;
7799
7800	while (1) {
7801		if (need_resched() ||
7802		    rwsem_is_contended(&fs_info->commit_root_sem)) {
7803			up_read(&fs_info->commit_root_sem);
7804			cond_resched();
7805			down_read(&fs_info->commit_root_sem);
7806		}
7807
7808		if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
7809			ret = restart_after_relocation(left_path, right_path,
7810						       &left_key, &right_key,
7811						       left_level, right_level,
7812						       sctx);
7813			if (ret < 0)
7814				goto out_unlock;
7815			sctx->last_reloc_trans = fs_info->last_reloc_trans;
7816		}
7817
7818		if (advance_left && !left_end_reached) {
7819			ret = tree_advance(left_path, &left_level,
7820					left_root_level,
7821					advance_left != ADVANCE_ONLY_NEXT,
7822					&left_key, reada_min_gen);
7823			if (ret == -1)
7824				left_end_reached = ADVANCE;
7825			else if (ret < 0)
7826				goto out_unlock;
7827			advance_left = 0;
7828		}
7829		if (advance_right && !right_end_reached) {
7830			ret = tree_advance(right_path, &right_level,
7831					right_root_level,
7832					advance_right != ADVANCE_ONLY_NEXT,
7833					&right_key, reada_min_gen);
7834			if (ret == -1)
7835				right_end_reached = ADVANCE;
7836			else if (ret < 0)
7837				goto out_unlock;
7838			advance_right = 0;
7839		}
7840
7841		if (left_end_reached && right_end_reached) {
7842			ret = 0;
7843			goto out_unlock;
7844		} else if (left_end_reached) {
7845			if (right_level == 0) {
7846				up_read(&fs_info->commit_root_sem);
7847				ret = changed_cb(left_path, right_path,
7848						&right_key,
7849						BTRFS_COMPARE_TREE_DELETED,
7850						sctx);
7851				if (ret < 0)
7852					goto out;
7853				down_read(&fs_info->commit_root_sem);
7854			}
7855			advance_right = ADVANCE;
7856			continue;
7857		} else if (right_end_reached) {
7858			if (left_level == 0) {
7859				up_read(&fs_info->commit_root_sem);
7860				ret = changed_cb(left_path, right_path,
7861						&left_key,
7862						BTRFS_COMPARE_TREE_NEW,
7863						sctx);
7864				if (ret < 0)
7865					goto out;
7866				down_read(&fs_info->commit_root_sem);
7867			}
7868			advance_left = ADVANCE;
7869			continue;
7870		}
7871
7872		if (left_level == 0 && right_level == 0) {
7873			up_read(&fs_info->commit_root_sem);
7874			cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7875			if (cmp < 0) {
7876				ret = changed_cb(left_path, right_path,
7877						&left_key,
7878						BTRFS_COMPARE_TREE_NEW,
7879						sctx);
7880				advance_left = ADVANCE;
7881			} else if (cmp > 0) {
7882				ret = changed_cb(left_path, right_path,
7883						&right_key,
7884						BTRFS_COMPARE_TREE_DELETED,
7885						sctx);
7886				advance_right = ADVANCE;
7887			} else {
7888				enum btrfs_compare_tree_result result;
7889
7890				WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
7891				ret = tree_compare_item(left_path, right_path,
7892							tmp_buf);
7893				if (ret)
7894					result = BTRFS_COMPARE_TREE_CHANGED;
7895				else
7896					result = BTRFS_COMPARE_TREE_SAME;
7897				ret = changed_cb(left_path, right_path,
7898						 &left_key, result, sctx);
7899				advance_left = ADVANCE;
7900				advance_right = ADVANCE;
7901			}
7902
7903			if (ret < 0)
7904				goto out;
7905			down_read(&fs_info->commit_root_sem);
7906		} else if (left_level == right_level) {
7907			cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7908			if (cmp < 0) {
7909				advance_left = ADVANCE;
7910			} else if (cmp > 0) {
7911				advance_right = ADVANCE;
7912			} else {
7913				left_blockptr = btrfs_node_blockptr(
7914						left_path->nodes[left_level],
7915						left_path->slots[left_level]);
7916				right_blockptr = btrfs_node_blockptr(
7917						right_path->nodes[right_level],
7918						right_path->slots[right_level]);
7919				left_gen = btrfs_node_ptr_generation(
7920						left_path->nodes[left_level],
7921						left_path->slots[left_level]);
7922				right_gen = btrfs_node_ptr_generation(
7923						right_path->nodes[right_level],
7924						right_path->slots[right_level]);
7925				if (left_blockptr == right_blockptr &&
7926				    left_gen == right_gen) {
7927					/*
7928					 * As we're on a shared block, don't
7929					 * allow to go deeper.
7930					 */
7931					advance_left = ADVANCE_ONLY_NEXT;
7932					advance_right = ADVANCE_ONLY_NEXT;
7933				} else {
7934					advance_left = ADVANCE;
7935					advance_right = ADVANCE;
7936				}
7937			}
7938		} else if (left_level < right_level) {
7939			advance_right = ADVANCE;
7940		} else {
7941			advance_left = ADVANCE;
7942		}
7943	}
7944
7945out_unlock:
7946	up_read(&fs_info->commit_root_sem);
7947out:
7948	btrfs_free_path(left_path);
7949	btrfs_free_path(right_path);
7950	kvfree(tmp_buf);
7951	return ret;
7952}
7953
7954static int send_subvol(struct send_ctx *sctx)
7955{
7956	int ret;
7957
7958	if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_STREAM_HEADER)) {
7959		ret = send_header(sctx);
7960		if (ret < 0)
7961			goto out;
7962	}
7963
7964	ret = send_subvol_begin(sctx);
7965	if (ret < 0)
7966		goto out;
7967
7968	if (sctx->parent_root) {
7969		ret = btrfs_compare_trees(sctx->send_root, sctx->parent_root, sctx);
7970		if (ret < 0)
7971			goto out;
7972		ret = finish_inode_if_needed(sctx, 1);
7973		if (ret < 0)
7974			goto out;
7975	} else {
7976		ret = full_send_tree(sctx);
7977		if (ret < 0)
7978			goto out;
7979	}
7980
7981out:
7982	free_recorded_refs(sctx);
7983	return ret;
7984}
7985
7986/*
7987 * If orphan cleanup did remove any orphans from a root, it means the tree
7988 * was modified and therefore the commit root is not the same as the current
7989 * root anymore. This is a problem, because send uses the commit root and
7990 * therefore can see inode items that don't exist in the current root anymore,
7991 * and for example make calls to btrfs_iget, which will do tree lookups based
7992 * on the current root and not on the commit root. Those lookups will fail,
7993 * returning a -ESTALE error, and making send fail with that error. So make
7994 * sure a send does not see any orphans we have just removed, and that it will
7995 * see the same inodes regardless of whether a transaction commit happened
7996 * before it started (meaning that the commit root will be the same as the
7997 * current root) or not.
7998 */
7999static int ensure_commit_roots_uptodate(struct send_ctx *sctx)
8000{
8001	int i;
8002	struct btrfs_trans_handle *trans = NULL;
8003
8004again:
8005	if (sctx->parent_root &&
8006	    sctx->parent_root->node != sctx->parent_root->commit_root)
8007		goto commit_trans;
8008
8009	for (i = 0; i < sctx->clone_roots_cnt; i++)
8010		if (sctx->clone_roots[i].root->node !=
8011		    sctx->clone_roots[i].root->commit_root)
8012			goto commit_trans;
8013
8014	if (trans)
8015		return btrfs_end_transaction(trans);
8016
8017	return 0;
8018
8019commit_trans:
8020	/* Use any root, all fs roots will get their commit roots updated. */
8021	if (!trans) {
8022		trans = btrfs_join_transaction(sctx->send_root);
8023		if (IS_ERR(trans))
8024			return PTR_ERR(trans);
8025		goto again;
8026	}
8027
8028	return btrfs_commit_transaction(trans);
8029}
8030
8031/*
8032 * Make sure any existing dellaloc is flushed for any root used by a send
8033 * operation so that we do not miss any data and we do not race with writeback
8034 * finishing and changing a tree while send is using the tree. This could
8035 * happen if a subvolume is in RW mode, has delalloc, is turned to RO mode and
8036 * a send operation then uses the subvolume.
8037 * After flushing delalloc ensure_commit_roots_uptodate() must be called.
8038 */
8039static int flush_delalloc_roots(struct send_ctx *sctx)
8040{
8041	struct btrfs_root *root = sctx->parent_root;
8042	int ret;
8043	int i;
8044
8045	if (root) {
8046		ret = btrfs_start_delalloc_snapshot(root, false);
8047		if (ret)
8048			return ret;
8049		btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
8050	}
8051
8052	for (i = 0; i < sctx->clone_roots_cnt; i++) {
8053		root = sctx->clone_roots[i].root;
8054		ret = btrfs_start_delalloc_snapshot(root, false);
8055		if (ret)
8056			return ret;
8057		btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
8058	}
8059
8060	return 0;
8061}
8062
8063static void btrfs_root_dec_send_in_progress(struct btrfs_root* root)
8064{
8065	spin_lock(&root->root_item_lock);
8066	root->send_in_progress--;
8067	/*
8068	 * Not much left to do, we don't know why it's unbalanced and
8069	 * can't blindly reset it to 0.
8070	 */
8071	if (root->send_in_progress < 0)
8072		btrfs_err(root->fs_info,
8073			  "send_in_progress unbalanced %d root %llu",
8074			  root->send_in_progress, root->root_key.objectid);
8075	spin_unlock(&root->root_item_lock);
8076}
8077
8078static void dedupe_in_progress_warn(const struct btrfs_root *root)
8079{
8080	btrfs_warn_rl(root->fs_info,
8081"cannot use root %llu for send while deduplications on it are in progress (%d in progress)",
8082		      root->root_key.objectid, root->dedupe_in_progress);
8083}
8084
8085long btrfs_ioctl_send(struct inode *inode, struct btrfs_ioctl_send_args *arg)
8086{
8087	int ret = 0;
8088	struct btrfs_root *send_root = BTRFS_I(inode)->root;
8089	struct btrfs_fs_info *fs_info = send_root->fs_info;
8090	struct btrfs_root *clone_root;
8091	struct send_ctx *sctx = NULL;
8092	u32 i;
8093	u64 *clone_sources_tmp = NULL;
8094	int clone_sources_to_rollback = 0;
8095	size_t alloc_size;
8096	int sort_clone_roots = 0;
8097	struct btrfs_lru_cache_entry *entry;
8098	struct btrfs_lru_cache_entry *tmp;
8099
8100	if (!capable(CAP_SYS_ADMIN))
8101		return -EPERM;
8102
8103	/*
8104	 * The subvolume must remain read-only during send, protect against
8105	 * making it RW. This also protects against deletion.
8106	 */
8107	spin_lock(&send_root->root_item_lock);
8108	if (btrfs_root_readonly(send_root) && send_root->dedupe_in_progress) {
8109		dedupe_in_progress_warn(send_root);
8110		spin_unlock(&send_root->root_item_lock);
8111		return -EAGAIN;
8112	}
8113	send_root->send_in_progress++;
8114	spin_unlock(&send_root->root_item_lock);
8115
8116	/*
8117	 * Userspace tools do the checks and warn the user if it's
8118	 * not RO.
8119	 */
8120	if (!btrfs_root_readonly(send_root)) {
8121		ret = -EPERM;
8122		goto out;
8123	}
8124
8125	/*
8126	 * Check that we don't overflow at later allocations, we request
8127	 * clone_sources_count + 1 items, and compare to unsigned long inside
8128	 * access_ok. Also set an upper limit for allocation size so this can't
8129	 * easily exhaust memory. Max number of clone sources is about 200K.
8130	 */
8131	if (arg->clone_sources_count > SZ_8M / sizeof(struct clone_root)) {
8132		ret = -EINVAL;
8133		goto out;
8134	}
8135
8136	if (arg->flags & ~BTRFS_SEND_FLAG_MASK) {
8137		ret = -EOPNOTSUPP;
8138		goto out;
8139	}
8140
8141	sctx = kzalloc(sizeof(struct send_ctx), GFP_KERNEL);
8142	if (!sctx) {
8143		ret = -ENOMEM;
8144		goto out;
8145	}
8146
8147	INIT_LIST_HEAD(&sctx->new_refs);
8148	INIT_LIST_HEAD(&sctx->deleted_refs);
8149
8150	btrfs_lru_cache_init(&sctx->name_cache, SEND_MAX_NAME_CACHE_SIZE);
8151	btrfs_lru_cache_init(&sctx->backref_cache, SEND_MAX_BACKREF_CACHE_SIZE);
8152	btrfs_lru_cache_init(&sctx->dir_created_cache,
8153			     SEND_MAX_DIR_CREATED_CACHE_SIZE);
8154	/*
8155	 * This cache is periodically trimmed to a fixed size elsewhere, see
8156	 * cache_dir_utimes() and trim_dir_utimes_cache().
8157	 */
8158	btrfs_lru_cache_init(&sctx->dir_utimes_cache, 0);
8159
8160	sctx->pending_dir_moves = RB_ROOT;
8161	sctx->waiting_dir_moves = RB_ROOT;
8162	sctx->orphan_dirs = RB_ROOT;
8163	sctx->rbtree_new_refs = RB_ROOT;
8164	sctx->rbtree_deleted_refs = RB_ROOT;
8165
8166	sctx->flags = arg->flags;
8167
8168	if (arg->flags & BTRFS_SEND_FLAG_VERSION) {
8169		if (arg->version > BTRFS_SEND_STREAM_VERSION) {
8170			ret = -EPROTO;
8171			goto out;
8172		}
8173		/* Zero means "use the highest version" */
8174		sctx->proto = arg->version ?: BTRFS_SEND_STREAM_VERSION;
8175	} else {
8176		sctx->proto = 1;
8177	}
8178	if ((arg->flags & BTRFS_SEND_FLAG_COMPRESSED) && sctx->proto < 2) {
8179		ret = -EINVAL;
8180		goto out;
8181	}
8182
8183	sctx->send_filp = fget(arg->send_fd);
8184	if (!sctx->send_filp || !(sctx->send_filp->f_mode & FMODE_WRITE)) {
8185		ret = -EBADF;
8186		goto out;
8187	}
8188
8189	sctx->send_root = send_root;
8190	/*
8191	 * Unlikely but possible, if the subvolume is marked for deletion but
8192	 * is slow to remove the directory entry, send can still be started
8193	 */
8194	if (btrfs_root_dead(sctx->send_root)) {
8195		ret = -EPERM;
8196		goto out;
8197	}
8198
8199	sctx->clone_roots_cnt = arg->clone_sources_count;
8200
8201	if (sctx->proto >= 2) {
8202		u32 send_buf_num_pages;
8203
8204		sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V2;
8205		sctx->send_buf = vmalloc(sctx->send_max_size);
8206		if (!sctx->send_buf) {
8207			ret = -ENOMEM;
8208			goto out;
8209		}
8210		send_buf_num_pages = sctx->send_max_size >> PAGE_SHIFT;
8211		sctx->send_buf_pages = kcalloc(send_buf_num_pages,
8212					       sizeof(*sctx->send_buf_pages),
8213					       GFP_KERNEL);
8214		if (!sctx->send_buf_pages) {
8215			ret = -ENOMEM;
8216			goto out;
8217		}
8218		for (i = 0; i < send_buf_num_pages; i++) {
8219			sctx->send_buf_pages[i] =
8220				vmalloc_to_page(sctx->send_buf + (i << PAGE_SHIFT));
8221		}
8222	} else {
8223		sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V1;
8224		sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL);
8225	}
8226	if (!sctx->send_buf) {
8227		ret = -ENOMEM;
8228		goto out;
8229	}
8230
8231	sctx->clone_roots = kvcalloc(arg->clone_sources_count + 1,
8232				     sizeof(*sctx->clone_roots),
8233				     GFP_KERNEL);
8234	if (!sctx->clone_roots) {
8235		ret = -ENOMEM;
8236		goto out;
8237	}
8238
8239	alloc_size = array_size(sizeof(*arg->clone_sources),
8240				arg->clone_sources_count);
8241
8242	if (arg->clone_sources_count) {
8243		clone_sources_tmp = kvmalloc(alloc_size, GFP_KERNEL);
8244		if (!clone_sources_tmp) {
8245			ret = -ENOMEM;
8246			goto out;
8247		}
8248
8249		ret = copy_from_user(clone_sources_tmp, arg->clone_sources,
8250				alloc_size);
8251		if (ret) {
8252			ret = -EFAULT;
8253			goto out;
8254		}
8255
8256		for (i = 0; i < arg->clone_sources_count; i++) {
8257			clone_root = btrfs_get_fs_root(fs_info,
8258						clone_sources_tmp[i], true);
8259			if (IS_ERR(clone_root)) {
8260				ret = PTR_ERR(clone_root);
8261				goto out;
8262			}
8263			spin_lock(&clone_root->root_item_lock);
8264			if (!btrfs_root_readonly(clone_root) ||
8265			    btrfs_root_dead(clone_root)) {
8266				spin_unlock(&clone_root->root_item_lock);
8267				btrfs_put_root(clone_root);
8268				ret = -EPERM;
8269				goto out;
8270			}
8271			if (clone_root->dedupe_in_progress) {
8272				dedupe_in_progress_warn(clone_root);
8273				spin_unlock(&clone_root->root_item_lock);
8274				btrfs_put_root(clone_root);
8275				ret = -EAGAIN;
8276				goto out;
8277			}
8278			clone_root->send_in_progress++;
8279			spin_unlock(&clone_root->root_item_lock);
8280
8281			sctx->clone_roots[i].root = clone_root;
8282			clone_sources_to_rollback = i + 1;
8283		}
8284		kvfree(clone_sources_tmp);
8285		clone_sources_tmp = NULL;
8286	}
8287
8288	if (arg->parent_root) {
8289		sctx->parent_root = btrfs_get_fs_root(fs_info, arg->parent_root,
8290						      true);
8291		if (IS_ERR(sctx->parent_root)) {
8292			ret = PTR_ERR(sctx->parent_root);
8293			goto out;
8294		}
8295
8296		spin_lock(&sctx->parent_root->root_item_lock);
8297		sctx->parent_root->send_in_progress++;
8298		if (!btrfs_root_readonly(sctx->parent_root) ||
8299				btrfs_root_dead(sctx->parent_root)) {
8300			spin_unlock(&sctx->parent_root->root_item_lock);
8301			ret = -EPERM;
8302			goto out;
8303		}
8304		if (sctx->parent_root->dedupe_in_progress) {
8305			dedupe_in_progress_warn(sctx->parent_root);
8306			spin_unlock(&sctx->parent_root->root_item_lock);
8307			ret = -EAGAIN;
8308			goto out;
8309		}
8310		spin_unlock(&sctx->parent_root->root_item_lock);
8311	}
8312
8313	/*
8314	 * Clones from send_root are allowed, but only if the clone source
8315	 * is behind the current send position. This is checked while searching
8316	 * for possible clone sources.
8317	 */
8318	sctx->clone_roots[sctx->clone_roots_cnt++].root =
8319		btrfs_grab_root(sctx->send_root);
8320
8321	/* We do a bsearch later */
8322	sort(sctx->clone_roots, sctx->clone_roots_cnt,
8323			sizeof(*sctx->clone_roots), __clone_root_cmp_sort,
8324			NULL);
8325	sort_clone_roots = 1;
8326
8327	ret = flush_delalloc_roots(sctx);
8328	if (ret)
8329		goto out;
8330
8331	ret = ensure_commit_roots_uptodate(sctx);
8332	if (ret)
8333		goto out;
8334
8335	ret = send_subvol(sctx);
8336	if (ret < 0)
8337		goto out;
8338
8339	btrfs_lru_cache_for_each_entry_safe(&sctx->dir_utimes_cache, entry, tmp) {
8340		ret = send_utimes(sctx, entry->key, entry->gen);
8341		if (ret < 0)
8342			goto out;
8343		btrfs_lru_cache_remove(&sctx->dir_utimes_cache, entry);
8344	}
8345
8346	if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_END_CMD)) {
8347		ret = begin_cmd(sctx, BTRFS_SEND_C_END);
8348		if (ret < 0)
8349			goto out;
8350		ret = send_cmd(sctx);
8351		if (ret < 0)
8352			goto out;
8353	}
8354
8355out:
8356	WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->pending_dir_moves));
8357	while (sctx && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)) {
8358		struct rb_node *n;
8359		struct pending_dir_move *pm;
8360
8361		n = rb_first(&sctx->pending_dir_moves);
8362		pm = rb_entry(n, struct pending_dir_move, node);
8363		while (!list_empty(&pm->list)) {
8364			struct pending_dir_move *pm2;
8365
8366			pm2 = list_first_entry(&pm->list,
8367					       struct pending_dir_move, list);
8368			free_pending_move(sctx, pm2);
8369		}
8370		free_pending_move(sctx, pm);
8371	}
8372
8373	WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves));
8374	while (sctx && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) {
8375		struct rb_node *n;
8376		struct waiting_dir_move *dm;
8377
8378		n = rb_first(&sctx->waiting_dir_moves);
8379		dm = rb_entry(n, struct waiting_dir_move, node);
8380		rb_erase(&dm->node, &sctx->waiting_dir_moves);
8381		kfree(dm);
8382	}
8383
8384	WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->orphan_dirs));
8385	while (sctx && !RB_EMPTY_ROOT(&sctx->orphan_dirs)) {
8386		struct rb_node *n;
8387		struct orphan_dir_info *odi;
8388
8389		n = rb_first(&sctx->orphan_dirs);
8390		odi = rb_entry(n, struct orphan_dir_info, node);
8391		free_orphan_dir_info(sctx, odi);
8392	}
8393
8394	if (sort_clone_roots) {
8395		for (i = 0; i < sctx->clone_roots_cnt; i++) {
8396			btrfs_root_dec_send_in_progress(
8397					sctx->clone_roots[i].root);
8398			btrfs_put_root(sctx->clone_roots[i].root);
8399		}
8400	} else {
8401		for (i = 0; sctx && i < clone_sources_to_rollback; i++) {
8402			btrfs_root_dec_send_in_progress(
8403					sctx->clone_roots[i].root);
8404			btrfs_put_root(sctx->clone_roots[i].root);
8405		}
8406
8407		btrfs_root_dec_send_in_progress(send_root);
8408	}
8409	if (sctx && !IS_ERR_OR_NULL(sctx->parent_root)) {
8410		btrfs_root_dec_send_in_progress(sctx->parent_root);
8411		btrfs_put_root(sctx->parent_root);
8412	}
8413
8414	kvfree(clone_sources_tmp);
8415
8416	if (sctx) {
8417		if (sctx->send_filp)
8418			fput(sctx->send_filp);
8419
8420		kvfree(sctx->clone_roots);
8421		kfree(sctx->send_buf_pages);
8422		kvfree(sctx->send_buf);
8423		kvfree(sctx->verity_descriptor);
8424
8425		close_current_inode(sctx);
8426
8427		btrfs_lru_cache_clear(&sctx->name_cache);
8428		btrfs_lru_cache_clear(&sctx->backref_cache);
8429		btrfs_lru_cache_clear(&sctx->dir_created_cache);
8430		btrfs_lru_cache_clear(&sctx->dir_utimes_cache);
8431
8432		kfree(sctx);
8433	}
8434
8435	return ret;
8436}