1/*
   2 * Copyright (C) 2012 Alexander Block.  All rights reserved.
   3 *
   4 * This program is free software; you can redistribute it and/or
   5 * modify it under the terms of the GNU General Public
   6 * License v2 as published by the Free Software Foundation.
   7 *
   8 * This program is distributed in the hope that it will be useful,
   9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11 * General Public License for more details.
  12 *
  13 * You should have received a copy of the GNU General Public
  14 * License along with this program; if not, write to the
  15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16 * Boston, MA 021110-1307, USA.
  17 */
  18
  19#include <linux/bsearch.h>
  20#include <linux/fs.h>
  21#include <linux/file.h>
  22#include <linux/sort.h>
  23#include <linux/mount.h>
  24#include <linux/xattr.h>
  25#include <linux/posix_acl_xattr.h>
  26#include <linux/radix-tree.h>
  27#include <linux/vmalloc.h>
  28#include <linux/string.h>
  29
  30#include "send.h"
  31#include "backref.h"
  32#include "hash.h"
  33#include "locking.h"
  34#include "disk-io.h"
  35#include "btrfs_inode.h"
  36#include "transaction.h"
  37#include "compression.h"
  38
  39/*
  40 * A fs_path is a helper to dynamically build path names with unknown size.
  41 * It reallocates the internal buffer on demand.
  42 * It allows fast adding of path elements on the right side (normal path) and
  43 * fast adding to the left side (reversed path). A reversed path can also be
  44 * unreversed if needed.
  45 */
  46struct fs_path {
  47	union {
  48		struct {
  49			char *start;
  50			char *end;
  51
  52			char *buf;
  53			unsigned short buf_len:15;
  54			unsigned short reversed:1;
  55			char inline_buf[];
  56		};
  57		/*
  58		 * Average path length does not exceed 200 bytes, we'll have
  59		 * better packing in the slab and higher chance to satisfy
  60		 * a allocation later during send.
  61		 */
  62		char pad[256];
  63	};
  64};
  65#define FS_PATH_INLINE_SIZE \
  66	(sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
  67
  68
  69/* reused for each extent */
  70struct clone_root {
  71	struct btrfs_root *root;
  72	u64 ino;
  73	u64 offset;
  74
  75	u64 found_refs;
  76};
  77
  78#define SEND_CTX_MAX_NAME_CACHE_SIZE 128
  79#define SEND_CTX_NAME_CACHE_CLEAN_SIZE (SEND_CTX_MAX_NAME_CACHE_SIZE * 2)
  80
  81struct send_ctx {
  82	struct file *send_filp;
  83	loff_t send_off;
  84	char *send_buf;
  85	u32 send_size;
  86	u32 send_max_size;
  87	u64 total_send_size;
  88	u64 cmd_send_size[BTRFS_SEND_C_MAX + 1];
  89	u64 flags;	/* 'flags' member of btrfs_ioctl_send_args is u64 */
  90
  91	struct btrfs_root *send_root;
  92	struct btrfs_root *parent_root;
  93	struct clone_root *clone_roots;
  94	int clone_roots_cnt;
  95
  96	/* current state of the compare_tree call */
  97	struct btrfs_path *left_path;
  98	struct btrfs_path *right_path;
  99	struct btrfs_key *cmp_key;
 100
 101	/*
 102	 * infos of the currently processed inode. In case of deleted inodes,
 103	 * these are the values from the deleted inode.
 104	 */
 105	u64 cur_ino;
 106	u64 cur_inode_gen;
 107	int cur_inode_new;
 108	int cur_inode_new_gen;
 109	int cur_inode_deleted;
 110	u64 cur_inode_size;
 111	u64 cur_inode_mode;
 112	u64 cur_inode_rdev;
 113	u64 cur_inode_last_extent;
 114
 115	u64 send_progress;
 116
 117	struct list_head new_refs;
 118	struct list_head deleted_refs;
 119
 120	struct radix_tree_root name_cache;
 121	struct list_head name_cache_list;
 122	int name_cache_size;
 123
 124	struct file_ra_state ra;
 125
 126	char *read_buf;
 127
 128	/*
 129	 * We process inodes by their increasing order, so if before an
 130	 * incremental send we reverse the parent/child relationship of
 131	 * directories such that a directory with a lower inode number was
 132	 * the parent of a directory with a higher inode number, and the one
 133	 * becoming the new parent got renamed too, we can't rename/move the
 134	 * directory with lower inode number when we finish processing it - we
 135	 * must process the directory with higher inode number first, then
 136	 * rename/move it and then rename/move the directory with lower inode
 137	 * number. Example follows.
 138	 *
 139	 * Tree state when the first send was performed:
 140	 *
 141	 * .
 142	 * |-- a                   (ino 257)
 143	 *     |-- b               (ino 258)
 144	 *         |
 145	 *         |
 146	 *         |-- c           (ino 259)
 147	 *         |   |-- d       (ino 260)
 148	 *         |
 149	 *         |-- c2          (ino 261)
 150	 *
 151	 * Tree state when the second (incremental) send is performed:
 152	 *
 153	 * .
 154	 * |-- a                   (ino 257)
 155	 *     |-- b               (ino 258)
 156	 *         |-- c2          (ino 261)
 157	 *             |-- d2      (ino 260)
 158	 *                 |-- cc  (ino 259)
 159	 *
 160	 * The sequence of steps that lead to the second state was:
 161	 *
 162	 * mv /a/b/c/d /a/b/c2/d2
 163	 * mv /a/b/c /a/b/c2/d2/cc
 164	 *
 165	 * "c" has lower inode number, but we can't move it (2nd mv operation)
 166	 * before we move "d", which has higher inode number.
 167	 *
 168	 * So we just memorize which move/rename operations must be performed
 169	 * later when their respective parent is processed and moved/renamed.
 170	 */
 171
 172	/* Indexed by parent directory inode number. */
 173	struct rb_root pending_dir_moves;
 174
 175	/*
 176	 * Reverse index, indexed by the inode number of a directory that
 177	 * is waiting for the move/rename of its immediate parent before its
 178	 * own move/rename can be performed.
 179	 */
 180	struct rb_root waiting_dir_moves;
 181
 182	/*
 183	 * A directory that is going to be rm'ed might have a child directory
 184	 * which is in the pending directory moves index above. In this case,
 185	 * the directory can only be removed after the move/rename of its child
 186	 * is performed. Example:
 187	 *
 188	 * Parent snapshot:
 189	 *
 190	 * .                        (ino 256)
 191	 * |-- a/                   (ino 257)
 192	 *     |-- b/               (ino 258)
 193	 *         |-- c/           (ino 259)
 194	 *         |   |-- x/       (ino 260)
 195	 *         |
 196	 *         |-- y/           (ino 261)
 197	 *
 198	 * Send snapshot:
 199	 *
 200	 * .                        (ino 256)
 201	 * |-- a/                   (ino 257)
 202	 *     |-- b/               (ino 258)
 203	 *         |-- YY/          (ino 261)
 204	 *              |-- x/      (ino 260)
 205	 *
 206	 * Sequence of steps that lead to the send snapshot:
 207	 * rm -f /a/b/c/foo.txt
 208	 * mv /a/b/y /a/b/YY
 209	 * mv /a/b/c/x /a/b/YY
 210	 * rmdir /a/b/c
 211	 *
 212	 * When the child is processed, its move/rename is delayed until its
 213	 * parent is processed (as explained above), but all other operations
 214	 * like update utimes, chown, chgrp, etc, are performed and the paths
 215	 * that it uses for those operations must use the orphanized name of
 216	 * its parent (the directory we're going to rm later), so we need to
 217	 * memorize that name.
 218	 *
 219	 * Indexed by the inode number of the directory to be deleted.
 220	 */
 221	struct rb_root orphan_dirs;
 222};
 223
 224struct pending_dir_move {
 225	struct rb_node node;
 226	struct list_head list;
 227	u64 parent_ino;
 228	u64 ino;
 229	u64 gen;
 230	struct list_head update_refs;
 231};
 232
 233struct waiting_dir_move {
 234	struct rb_node node;
 235	u64 ino;
 236	/*
 237	 * There might be some directory that could not be removed because it
 238	 * was waiting for this directory inode to be moved first. Therefore
 239	 * after this directory is moved, we can try to rmdir the ino rmdir_ino.
 240	 */
 241	u64 rmdir_ino;
 242	bool orphanized;
 243};
 244
 245struct orphan_dir_info {
 246	struct rb_node node;
 247	u64 ino;
 248	u64 gen;
 249};
 250
 251struct name_cache_entry {
 252	struct list_head list;
 253	/*
 254	 * radix_tree has only 32bit entries but we need to handle 64bit inums.
 255	 * We use the lower 32bit of the 64bit inum to store it in the tree. If
 256	 * more then one inum would fall into the same entry, we use radix_list
 257	 * to store the additional entries. radix_list is also used to store
 258	 * entries where two entries have the same inum but different
 259	 * generations.
 260	 */
 261	struct list_head radix_list;
 262	u64 ino;
 263	u64 gen;
 264	u64 parent_ino;
 265	u64 parent_gen;
 266	int ret;
 267	int need_later_update;
 268	int name_len;
 269	char name[];
 270};
 271
 272static void inconsistent_snapshot_error(struct send_ctx *sctx,
 273					enum btrfs_compare_tree_result result,
 274					const char *what)
 275{
 276	const char *result_string;
 277
 278	switch (result) {
 279	case BTRFS_COMPARE_TREE_NEW:
 280		result_string = "new";
 281		break;
 282	case BTRFS_COMPARE_TREE_DELETED:
 283		result_string = "deleted";
 284		break;
 285	case BTRFS_COMPARE_TREE_CHANGED:
 286		result_string = "updated";
 287		break;
 288	case BTRFS_COMPARE_TREE_SAME:
 289		ASSERT(0);
 290		result_string = "unchanged";
 291		break;
 292	default:
 293		ASSERT(0);
 294		result_string = "unexpected";
 295	}
 296
 297	btrfs_err(sctx->send_root->fs_info,
 298		  "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
 299		  result_string, what, sctx->cmp_key->objectid,
 300		  sctx->send_root->root_key.objectid,
 301		  (sctx->parent_root ?
 302		   sctx->parent_root->root_key.objectid : 0));
 303}
 304
 305static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
 306
 307static struct waiting_dir_move *
 308get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
 309
 310static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino);
 311
 312static int need_send_hole(struct send_ctx *sctx)
 313{
 314	return (sctx->parent_root && !sctx->cur_inode_new &&
 315		!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
 316		S_ISREG(sctx->cur_inode_mode));
 317}
 318
 319static void fs_path_reset(struct fs_path *p)
 320{
 321	if (p->reversed) {
 322		p->start = p->buf + p->buf_len - 1;
 323		p->end = p->start;
 324		*p->start = 0;
 325	} else {
 326		p->start = p->buf;
 327		p->end = p->start;
 328		*p->start = 0;
 329	}
 330}
 331
 332static struct fs_path *fs_path_alloc(void)
 333{
 334	struct fs_path *p;
 335
 336	p = kmalloc(sizeof(*p), GFP_KERNEL);
 337	if (!p)
 338		return NULL;
 339	p->reversed = 0;
 340	p->buf = p->inline_buf;
 341	p->buf_len = FS_PATH_INLINE_SIZE;
 342	fs_path_reset(p);
 343	return p;
 344}
 345
 346static struct fs_path *fs_path_alloc_reversed(void)
 347{
 348	struct fs_path *p;
 349
 350	p = fs_path_alloc();
 351	if (!p)
 352		return NULL;
 353	p->reversed = 1;
 354	fs_path_reset(p);
 355	return p;
 356}
 357
 358static void fs_path_free(struct fs_path *p)
 359{
 360	if (!p)
 361		return;
 362	if (p->buf != p->inline_buf)
 363		kfree(p->buf);
 364	kfree(p);
 365}
 366
 367static int fs_path_len(struct fs_path *p)
 368{
 369	return p->end - p->start;
 370}
 371
 372static int fs_path_ensure_buf(struct fs_path *p, int len)
 373{
 374	char *tmp_buf;
 375	int path_len;
 376	int old_buf_len;
 377
 378	len++;
 379
 380	if (p->buf_len >= len)
 381		return 0;
 382
 383	if (len > PATH_MAX) {
 384		WARN_ON(1);
 385		return -ENOMEM;
 386	}
 387
 388	path_len = p->end - p->start;
 389	old_buf_len = p->buf_len;
 390
 391	/*
 392	 * First time the inline_buf does not suffice
 393	 */
 394	if (p->buf == p->inline_buf) {
 395		tmp_buf = kmalloc(len, GFP_KERNEL);
 396		if (tmp_buf)
 397			memcpy(tmp_buf, p->buf, old_buf_len);
 398	} else {
 399		tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
 400	}
 401	if (!tmp_buf)
 402		return -ENOMEM;
 403	p->buf = tmp_buf;
 404	/*
 405	 * The real size of the buffer is bigger, this will let the fast path
 406	 * happen most of the time
 407	 */
 408	p->buf_len = ksize(p->buf);
 409
 410	if (p->reversed) {
 411		tmp_buf = p->buf + old_buf_len - path_len - 1;
 412		p->end = p->buf + p->buf_len - 1;
 413		p->start = p->end - path_len;
 414		memmove(p->start, tmp_buf, path_len + 1);
 415	} else {
 416		p->start = p->buf;
 417		p->end = p->start + path_len;
 418	}
 419	return 0;
 420}
 421
 422static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
 423				   char **prepared)
 424{
 425	int ret;
 426	int new_len;
 427
 428	new_len = p->end - p->start + name_len;
 429	if (p->start != p->end)
 430		new_len++;
 431	ret = fs_path_ensure_buf(p, new_len);
 432	if (ret < 0)
 433		goto out;
 434
 435	if (p->reversed) {
 436		if (p->start != p->end)
 437			*--p->start = '/';
 438		p->start -= name_len;
 439		*prepared = p->start;
 440	} else {
 441		if (p->start != p->end)
 442			*p->end++ = '/';
 443		*prepared = p->end;
 444		p->end += name_len;
 445		*p->end = 0;
 446	}
 447
 448out:
 449	return ret;
 450}
 451
 452static int fs_path_add(struct fs_path *p, const char *name, int name_len)
 453{
 454	int ret;
 455	char *prepared;
 456
 457	ret = fs_path_prepare_for_add(p, name_len, &prepared);
 458	if (ret < 0)
 459		goto out;
 460	memcpy(prepared, name, name_len);
 461
 462out:
 463	return ret;
 464}
 465
 466static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
 467{
 468	int ret;
 469	char *prepared;
 470
 471	ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
 472	if (ret < 0)
 473		goto out;
 474	memcpy(prepared, p2->start, p2->end - p2->start);
 475
 476out:
 477	return ret;
 478}
 479
 480static int fs_path_add_from_extent_buffer(struct fs_path *p,
 481					  struct extent_buffer *eb,
 482					  unsigned long off, int len)
 483{
 484	int ret;
 485	char *prepared;
 486
 487	ret = fs_path_prepare_for_add(p, len, &prepared);
 488	if (ret < 0)
 489		goto out;
 490
 491	read_extent_buffer(eb, prepared, off, len);
 492
 493out:
 494	return ret;
 495}
 496
 497static int fs_path_copy(struct fs_path *p, struct fs_path *from)
 498{
 499	int ret;
 500
 501	p->reversed = from->reversed;
 502	fs_path_reset(p);
 503
 504	ret = fs_path_add_path(p, from);
 505
 506	return ret;
 507}
 508
 509
 510static void fs_path_unreverse(struct fs_path *p)
 511{
 512	char *tmp;
 513	int len;
 514
 515	if (!p->reversed)
 516		return;
 517
 518	tmp = p->start;
 519	len = p->end - p->start;
 520	p->start = p->buf;
 521	p->end = p->start + len;
 522	memmove(p->start, tmp, len + 1);
 523	p->reversed = 0;
 524}
 525
 526static struct btrfs_path *alloc_path_for_send(void)
 527{
 528	struct btrfs_path *path;
 529
 530	path = btrfs_alloc_path();
 531	if (!path)
 532		return NULL;
 533	path->search_commit_root = 1;
 534	path->skip_locking = 1;
 535	path->need_commit_sem = 1;
 536	return path;
 537}
 538
 539static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
 540{
 541	int ret;
 542	mm_segment_t old_fs;
 543	u32 pos = 0;
 544
 545	old_fs = get_fs();
 546	set_fs(KERNEL_DS);
 547
 548	while (pos < len) {
 549		ret = vfs_write(filp, (__force const char __user *)buf + pos,
 550				len - pos, off);
 551		/* TODO handle that correctly */
 552		/*if (ret == -ERESTARTSYS) {
 553			continue;
 554		}*/
 555		if (ret < 0)
 556			goto out;
 557		if (ret == 0) {
 558			ret = -EIO;
 559			goto out;
 560		}
 561		pos += ret;
 562	}
 563
 564	ret = 0;
 565
 566out:
 567	set_fs(old_fs);
 568	return ret;
 569}
 570
 571static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
 572{
 573	struct btrfs_tlv_header *hdr;
 574	int total_len = sizeof(*hdr) + len;
 575	int left = sctx->send_max_size - sctx->send_size;
 576
 577	if (unlikely(left < total_len))
 578		return -EOVERFLOW;
 579
 580	hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
 581	hdr->tlv_type = cpu_to_le16(attr);
 582	hdr->tlv_len = cpu_to_le16(len);
 583	memcpy(hdr + 1, data, len);
 584	sctx->send_size += total_len;
 585
 586	return 0;
 587}
 588
 589#define TLV_PUT_DEFINE_INT(bits) \
 590	static int tlv_put_u##bits(struct send_ctx *sctx,	 	\
 591			u##bits attr, u##bits value)			\
 592	{								\
 593		__le##bits __tmp = cpu_to_le##bits(value);		\
 594		return tlv_put(sctx, attr, &__tmp, sizeof(__tmp));	\
 595	}
 596
 597TLV_PUT_DEFINE_INT(64)
 598
 599static int tlv_put_string(struct send_ctx *sctx, u16 attr,
 600			  const char *str, int len)
 601{
 602	if (len == -1)
 603		len = strlen(str);
 604	return tlv_put(sctx, attr, str, len);
 605}
 606
 607static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
 608			const u8 *uuid)
 609{
 610	return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
 611}
 612
 613static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
 614				  struct extent_buffer *eb,
 615				  struct btrfs_timespec *ts)
 616{
 617	struct btrfs_timespec bts;
 618	read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
 619	return tlv_put(sctx, attr, &bts, sizeof(bts));
 620}
 621
 622
 623#define TLV_PUT(sctx, attrtype, attrlen, data) \
 624	do { \
 625		ret = tlv_put(sctx, attrtype, attrlen, data); \
 626		if (ret < 0) \
 627			goto tlv_put_failure; \
 628	} while (0)
 629
 630#define TLV_PUT_INT(sctx, attrtype, bits, value) \
 631	do { \
 632		ret = tlv_put_u##bits(sctx, attrtype, value); \
 633		if (ret < 0) \
 634			goto tlv_put_failure; \
 635	} while (0)
 636
 637#define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
 638#define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
 639#define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
 640#define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
 641#define TLV_PUT_STRING(sctx, attrtype, str, len) \
 642	do { \
 643		ret = tlv_put_string(sctx, attrtype, str, len); \
 644		if (ret < 0) \
 645			goto tlv_put_failure; \
 646	} while (0)
 647#define TLV_PUT_PATH(sctx, attrtype, p) \
 648	do { \
 649		ret = tlv_put_string(sctx, attrtype, p->start, \
 650			p->end - p->start); \
 651		if (ret < 0) \
 652			goto tlv_put_failure; \
 653	} while(0)
 654#define TLV_PUT_UUID(sctx, attrtype, uuid) \
 655	do { \
 656		ret = tlv_put_uuid(sctx, attrtype, uuid); \
 657		if (ret < 0) \
 658			goto tlv_put_failure; \
 659	} while (0)
 660#define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
 661	do { \
 662		ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
 663		if (ret < 0) \
 664			goto tlv_put_failure; \
 665	} while (0)
 666
 667static int send_header(struct send_ctx *sctx)
 668{
 669	struct btrfs_stream_header hdr;
 670
 671	strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
 672	hdr.version = cpu_to_le32(BTRFS_SEND_STREAM_VERSION);
 673
 674	return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
 675					&sctx->send_off);
 676}
 677
 678/*
 679 * For each command/item we want to send to userspace, we call this function.
 680 */
 681static int begin_cmd(struct send_ctx *sctx, int cmd)
 682{
 683	struct btrfs_cmd_header *hdr;
 684
 685	if (WARN_ON(!sctx->send_buf))
 686		return -EINVAL;
 687
 688	BUG_ON(sctx->send_size);
 689
 690	sctx->send_size += sizeof(*hdr);
 691	hdr = (struct btrfs_cmd_header *)sctx->send_buf;
 692	hdr->cmd = cpu_to_le16(cmd);
 693
 694	return 0;
 695}
 696
 697static int send_cmd(struct send_ctx *sctx)
 698{
 699	int ret;
 700	struct btrfs_cmd_header *hdr;
 701	u32 crc;
 702
 703	hdr = (struct btrfs_cmd_header *)sctx->send_buf;
 704	hdr->len = cpu_to_le32(sctx->send_size - sizeof(*hdr));
 705	hdr->crc = 0;
 706
 707	crc = btrfs_crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
 708	hdr->crc = cpu_to_le32(crc);
 709
 710	ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
 711					&sctx->send_off);
 712
 713	sctx->total_send_size += sctx->send_size;
 714	sctx->cmd_send_size[le16_to_cpu(hdr->cmd)] += sctx->send_size;
 715	sctx->send_size = 0;
 716
 717	return ret;
 718}
 719
 720/*
 721 * Sends a move instruction to user space
 722 */
 723static int send_rename(struct send_ctx *sctx,
 724		     struct fs_path *from, struct fs_path *to)
 725{
 726	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
 727	int ret;
 728
 729	btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
 730
 731	ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
 732	if (ret < 0)
 733		goto out;
 734
 735	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
 736	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
 737
 738	ret = send_cmd(sctx);
 739
 740tlv_put_failure:
 741out:
 742	return ret;
 743}
 744
 745/*
 746 * Sends a link instruction to user space
 747 */
 748static int send_link(struct send_ctx *sctx,
 749		     struct fs_path *path, struct fs_path *lnk)
 750{
 751	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
 752	int ret;
 753
 754	btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
 755
 756	ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
 757	if (ret < 0)
 758		goto out;
 759
 760	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
 761	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
 762
 763	ret = send_cmd(sctx);
 764
 765tlv_put_failure:
 766out:
 767	return ret;
 768}
 769
 770/*
 771 * Sends an unlink instruction to user space
 772 */
 773static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
 774{
 775	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
 776	int ret;
 777
 778	btrfs_debug(fs_info, "send_unlink %s", path->start);
 779
 780	ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
 781	if (ret < 0)
 782		goto out;
 783
 784	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
 785
 786	ret = send_cmd(sctx);
 787
 788tlv_put_failure:
 789out:
 790	return ret;
 791}
 792
 793/*
 794 * Sends a rmdir instruction to user space
 795 */
 796static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
 797{
 798	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
 799	int ret;
 800
 801	btrfs_debug(fs_info, "send_rmdir %s", path->start);
 802
 803	ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
 804	if (ret < 0)
 805		goto out;
 806
 807	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
 808
 809	ret = send_cmd(sctx);
 810
 811tlv_put_failure:
 812out:
 813	return ret;
 814}
 815
 816/*
 817 * Helper function to retrieve some fields from an inode item.
 818 */
 819static int __get_inode_info(struct btrfs_root *root, struct btrfs_path *path,
 820			  u64 ino, u64 *size, u64 *gen, u64 *mode, u64 *uid,
 821			  u64 *gid, u64 *rdev)
 822{
 823	int ret;
 824	struct btrfs_inode_item *ii;
 825	struct btrfs_key key;
 826
 827	key.objectid = ino;
 828	key.type = BTRFS_INODE_ITEM_KEY;
 829	key.offset = 0;
 830	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 831	if (ret) {
 832		if (ret > 0)
 833			ret = -ENOENT;
 834		return ret;
 835	}
 836
 837	ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
 838			struct btrfs_inode_item);
 839	if (size)
 840		*size = btrfs_inode_size(path->nodes[0], ii);
 841	if (gen)
 842		*gen = btrfs_inode_generation(path->nodes[0], ii);
 843	if (mode)
 844		*mode = btrfs_inode_mode(path->nodes[0], ii);
 845	if (uid)
 846		*uid = btrfs_inode_uid(path->nodes[0], ii);
 847	if (gid)
 848		*gid = btrfs_inode_gid(path->nodes[0], ii);
 849	if (rdev)
 850		*rdev = btrfs_inode_rdev(path->nodes[0], ii);
 851
 852	return ret;
 853}
 854
 855static int get_inode_info(struct btrfs_root *root,
 856			  u64 ino, u64 *size, u64 *gen,
 857			  u64 *mode, u64 *uid, u64 *gid,
 858			  u64 *rdev)
 859{
 860	struct btrfs_path *path;
 861	int ret;
 862
 863	path = alloc_path_for_send();
 864	if (!path)
 865		return -ENOMEM;
 866	ret = __get_inode_info(root, path, ino, size, gen, mode, uid, gid,
 867			       rdev);
 868	btrfs_free_path(path);
 869	return ret;
 870}
 871
 872typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
 873				   struct fs_path *p,
 874				   void *ctx);
 875
 876/*
 877 * Helper function to iterate the entries in ONE btrfs_inode_ref or
 878 * btrfs_inode_extref.
 879 * The iterate callback may return a non zero value to stop iteration. This can
 880 * be a negative value for error codes or 1 to simply stop it.
 881 *
 882 * path must point to the INODE_REF or INODE_EXTREF when called.
 883 */
 884static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
 885			     struct btrfs_key *found_key, int resolve,
 886			     iterate_inode_ref_t iterate, void *ctx)
 887{
 888	struct extent_buffer *eb = path->nodes[0];
 889	struct btrfs_item *item;
 890	struct btrfs_inode_ref *iref;
 891	struct btrfs_inode_extref *extref;
 892	struct btrfs_path *tmp_path;
 893	struct fs_path *p;
 894	u32 cur = 0;
 895	u32 total;
 896	int slot = path->slots[0];
 897	u32 name_len;
 898	char *start;
 899	int ret = 0;
 900	int num = 0;
 901	int index;
 902	u64 dir;
 903	unsigned long name_off;
 904	unsigned long elem_size;
 905	unsigned long ptr;
 906
 907	p = fs_path_alloc_reversed();
 908	if (!p)
 909		return -ENOMEM;
 910
 911	tmp_path = alloc_path_for_send();
 912	if (!tmp_path) {
 913		fs_path_free(p);
 914		return -ENOMEM;
 915	}
 916
 917
 918	if (found_key->type == BTRFS_INODE_REF_KEY) {
 919		ptr = (unsigned long)btrfs_item_ptr(eb, slot,
 920						    struct btrfs_inode_ref);
 921		item = btrfs_item_nr(slot);
 922		total = btrfs_item_size(eb, item);
 923		elem_size = sizeof(*iref);
 924	} else {
 925		ptr = btrfs_item_ptr_offset(eb, slot);
 926		total = btrfs_item_size_nr(eb, slot);
 927		elem_size = sizeof(*extref);
 928	}
 929
 930	while (cur < total) {
 931		fs_path_reset(p);
 932
 933		if (found_key->type == BTRFS_INODE_REF_KEY) {
 934			iref = (struct btrfs_inode_ref *)(ptr + cur);
 935			name_len = btrfs_inode_ref_name_len(eb, iref);
 936			name_off = (unsigned long)(iref + 1);
 937			index = btrfs_inode_ref_index(eb, iref);
 938			dir = found_key->offset;
 939		} else {
 940			extref = (struct btrfs_inode_extref *)(ptr + cur);
 941			name_len = btrfs_inode_extref_name_len(eb, extref);
 942			name_off = (unsigned long)&extref->name;
 943			index = btrfs_inode_extref_index(eb, extref);
 944			dir = btrfs_inode_extref_parent(eb, extref);
 945		}
 946
 947		if (resolve) {
 948			start = btrfs_ref_to_path(root, tmp_path, name_len,
 949						  name_off, eb, dir,
 950						  p->buf, p->buf_len);
 951			if (IS_ERR(start)) {
 952				ret = PTR_ERR(start);
 953				goto out;
 954			}
 955			if (start < p->buf) {
 956				/* overflow , try again with larger buffer */
 957				ret = fs_path_ensure_buf(p,
 958						p->buf_len + p->buf - start);
 959				if (ret < 0)
 960					goto out;
 961				start = btrfs_ref_to_path(root, tmp_path,
 962							  name_len, name_off,
 963							  eb, dir,
 964							  p->buf, p->buf_len);
 965				if (IS_ERR(start)) {
 966					ret = PTR_ERR(start);
 967					goto out;
 968				}
 969				BUG_ON(start < p->buf);
 970			}
 971			p->start = start;
 972		} else {
 973			ret = fs_path_add_from_extent_buffer(p, eb, name_off,
 974							     name_len);
 975			if (ret < 0)
 976				goto out;
 977		}
 978
 979		cur += elem_size + name_len;
 980		ret = iterate(num, dir, index, p, ctx);
 981		if (ret)
 982			goto out;
 983		num++;
 984	}
 985
 986out:
 987	btrfs_free_path(tmp_path);
 988	fs_path_free(p);
 989	return ret;
 990}
 991
 992typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
 993				  const char *name, int name_len,
 994				  const char *data, int data_len,
 995				  u8 type, void *ctx);
 996
 997/*
 998 * Helper function to iterate the entries in ONE btrfs_dir_item.
 999 * The iterate callback may return a non zero value to stop iteration. This can
1000 * be a negative value for error codes or 1 to simply stop it.
1001 *
1002 * path must point to the dir item when called.
1003 */
1004static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
1005			    struct btrfs_key *found_key,
1006			    iterate_dir_item_t iterate, void *ctx)
1007{
1008	int ret = 0;
1009	struct extent_buffer *eb;
1010	struct btrfs_item *item;
1011	struct btrfs_dir_item *di;
1012	struct btrfs_key di_key;
1013	char *buf = NULL;
1014	int buf_len;
1015	u32 name_len;
1016	u32 data_len;
1017	u32 cur;
1018	u32 len;
1019	u32 total;
1020	int slot;
1021	int num;
1022	u8 type;
1023
1024	/*
1025	 * Start with a small buffer (1 page). If later we end up needing more
1026	 * space, which can happen for xattrs on a fs with a leaf size greater
1027	 * then the page size, attempt to increase the buffer. Typically xattr
1028	 * values are small.
1029	 */
1030	buf_len = PATH_MAX;
1031	buf = kmalloc(buf_len, GFP_KERNEL);
1032	if (!buf) {
1033		ret = -ENOMEM;
1034		goto out;
1035	}
1036
1037	eb = path->nodes[0];
1038	slot = path->slots[0];
1039	item = btrfs_item_nr(slot);
1040	di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1041	cur = 0;
1042	len = 0;
1043	total = btrfs_item_size(eb, item);
1044
1045	num = 0;
1046	while (cur < total) {
1047		name_len = btrfs_dir_name_len(eb, di);
1048		data_len = btrfs_dir_data_len(eb, di);
1049		type = btrfs_dir_type(eb, di);
1050		btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1051
1052		if (type == BTRFS_FT_XATTR) {
1053			if (name_len > XATTR_NAME_MAX) {
1054				ret = -ENAMETOOLONG;
1055				goto out;
1056			}
1057			if (name_len + data_len >
1058					BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1059				ret = -E2BIG;
1060				goto out;
1061			}
1062		} else {
1063			/*
1064			 * Path too long
1065			 */
1066			if (name_len + data_len > PATH_MAX) {
1067				ret = -ENAMETOOLONG;
1068				goto out;
1069			}
1070		}
1071
1072		if (name_len + data_len > buf_len) {
1073			buf_len = name_len + data_len;
1074			if (is_vmalloc_addr(buf)) {
1075				vfree(buf);
1076				buf = NULL;
1077			} else {
1078				char *tmp = krealloc(buf, buf_len,
1079						GFP_KERNEL | __GFP_NOWARN);
1080
1081				if (!tmp)
1082					kfree(buf);
1083				buf = tmp;
1084			}
1085			if (!buf) {
1086				buf = vmalloc(buf_len);
1087				if (!buf) {
1088					ret = -ENOMEM;
1089					goto out;
1090				}
1091			}
1092		}
1093
1094		read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1095				name_len + data_len);
1096
1097		len = sizeof(*di) + name_len + data_len;
1098		di = (struct btrfs_dir_item *)((char *)di + len);
1099		cur += len;
1100
1101		ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1102				data_len, type, ctx);
1103		if (ret < 0)
1104			goto out;
1105		if (ret) {
1106			ret = 0;
1107			goto out;
1108		}
1109
1110		num++;
1111	}
1112
1113out:
1114	kvfree(buf);
1115	return ret;
1116}
1117
1118static int __copy_first_ref(int num, u64 dir, int index,
1119			    struct fs_path *p, void *ctx)
1120{
1121	int ret;
1122	struct fs_path *pt = ctx;
1123
1124	ret = fs_path_copy(pt, p);
1125	if (ret < 0)
1126		return ret;
1127
1128	/* we want the first only */
1129	return 1;
1130}
1131
1132/*
1133 * Retrieve the first path of an inode. If an inode has more then one
1134 * ref/hardlink, this is ignored.
1135 */
1136static int get_inode_path(struct btrfs_root *root,
1137			  u64 ino, struct fs_path *path)
1138{
1139	int ret;
1140	struct btrfs_key key, found_key;
1141	struct btrfs_path *p;
1142
1143	p = alloc_path_for_send();
1144	if (!p)
1145		return -ENOMEM;
1146
1147	fs_path_reset(path);
1148
1149	key.objectid = ino;
1150	key.type = BTRFS_INODE_REF_KEY;
1151	key.offset = 0;
1152
1153	ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1154	if (ret < 0)
1155		goto out;
1156	if (ret) {
1157		ret = 1;
1158		goto out;
1159	}
1160	btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1161	if (found_key.objectid != ino ||
1162	    (found_key.type != BTRFS_INODE_REF_KEY &&
1163	     found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1164		ret = -ENOENT;
1165		goto out;
1166	}
1167
1168	ret = iterate_inode_ref(root, p, &found_key, 1,
1169				__copy_first_ref, path);
1170	if (ret < 0)
1171		goto out;
1172	ret = 0;
1173
1174out:
1175	btrfs_free_path(p);
1176	return ret;
1177}
1178
1179struct backref_ctx {
1180	struct send_ctx *sctx;
1181
1182	struct btrfs_path *path;
1183	/* number of total found references */
1184	u64 found;
1185
1186	/*
1187	 * used for clones found in send_root. clones found behind cur_objectid
1188	 * and cur_offset are not considered as allowed clones.
1189	 */
1190	u64 cur_objectid;
1191	u64 cur_offset;
1192
1193	/* may be truncated in case it's the last extent in a file */
1194	u64 extent_len;
1195
1196	/* data offset in the file extent item */
1197	u64 data_offset;
1198
1199	/* Just to check for bugs in backref resolving */
1200	int found_itself;
1201};
1202
1203static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1204{
1205	u64 root = (u64)(uintptr_t)key;
1206	struct clone_root *cr = (struct clone_root *)elt;
1207
1208	if (root < cr->root->objectid)
1209		return -1;
1210	if (root > cr->root->objectid)
1211		return 1;
1212	return 0;
1213}
1214
1215static int __clone_root_cmp_sort(const void *e1, const void *e2)
1216{
1217	struct clone_root *cr1 = (struct clone_root *)e1;
1218	struct clone_root *cr2 = (struct clone_root *)e2;
1219
1220	if (cr1->root->objectid < cr2->root->objectid)
1221		return -1;
1222	if (cr1->root->objectid > cr2->root->objectid)
1223		return 1;
1224	return 0;
1225}
1226
1227/*
1228 * Called for every backref that is found for the current extent.
1229 * Results are collected in sctx->clone_roots->ino/offset/found_refs
1230 */
1231static int __iterate_backrefs(u64 ino, u64 offset, u64 root, void *ctx_)
1232{
1233	struct backref_ctx *bctx = ctx_;
1234	struct clone_root *found;
1235	int ret;
1236	u64 i_size;
1237
1238	/* First check if the root is in the list of accepted clone sources */
1239	found = bsearch((void *)(uintptr_t)root, bctx->sctx->clone_roots,
1240			bctx->sctx->clone_roots_cnt,
1241			sizeof(struct clone_root),
1242			__clone_root_cmp_bsearch);
1243	if (!found)
1244		return 0;
1245
1246	if (found->root == bctx->sctx->send_root &&
1247	    ino == bctx->cur_objectid &&
1248	    offset == bctx->cur_offset) {
1249		bctx->found_itself = 1;
1250	}
1251
1252	/*
1253	 * There are inodes that have extents that lie behind its i_size. Don't
1254	 * accept clones from these extents.
1255	 */
1256	ret = __get_inode_info(found->root, bctx->path, ino, &i_size, NULL, NULL,
1257			       NULL, NULL, NULL);
1258	btrfs_release_path(bctx->path);
1259	if (ret < 0)
1260		return ret;
1261
1262	if (offset + bctx->data_offset + bctx->extent_len > i_size)
1263		return 0;
1264
1265	/*
1266	 * Make sure we don't consider clones from send_root that are
1267	 * behind the current inode/offset.
1268	 */
1269	if (found->root == bctx->sctx->send_root) {
1270		/*
1271		 * TODO for the moment we don't accept clones from the inode
1272		 * that is currently send. We may change this when
1273		 * BTRFS_IOC_CLONE_RANGE supports cloning from and to the same
1274		 * file.
1275		 */
1276		if (ino >= bctx->cur_objectid)
1277			return 0;
1278#if 0
1279		if (ino > bctx->cur_objectid)
1280			return 0;
1281		if (offset + bctx->extent_len > bctx->cur_offset)
1282			return 0;
1283#endif
1284	}
1285
1286	bctx->found++;
1287	found->found_refs++;
1288	if (ino < found->ino) {
1289		found->ino = ino;
1290		found->offset = offset;
1291	} else if (found->ino == ino) {
1292		/*
1293		 * same extent found more then once in the same file.
1294		 */
1295		if (found->offset > offset + bctx->extent_len)
1296			found->offset = offset;
1297	}
1298
1299	return 0;
1300}
1301
1302/*
1303 * Given an inode, offset and extent item, it finds a good clone for a clone
1304 * instruction. Returns -ENOENT when none could be found. The function makes
1305 * sure that the returned clone is usable at the point where sending is at the
1306 * moment. This means, that no clones are accepted which lie behind the current
1307 * inode+offset.
1308 *
1309 * path must point to the extent item when called.
1310 */
1311static int find_extent_clone(struct send_ctx *sctx,
1312			     struct btrfs_path *path,
1313			     u64 ino, u64 data_offset,
1314			     u64 ino_size,
1315			     struct clone_root **found)
1316{
1317	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1318	int ret;
1319	int extent_type;
1320	u64 logical;
1321	u64 disk_byte;
1322	u64 num_bytes;
1323	u64 extent_item_pos;
1324	u64 flags = 0;
1325	struct btrfs_file_extent_item *fi;
1326	struct extent_buffer *eb = path->nodes[0];
1327	struct backref_ctx *backref_ctx = NULL;
1328	struct clone_root *cur_clone_root;
1329	struct btrfs_key found_key;
1330	struct btrfs_path *tmp_path;
1331	int compressed;
1332	u32 i;
1333
1334	tmp_path = alloc_path_for_send();
1335	if (!tmp_path)
1336		return -ENOMEM;
1337
1338	/* We only use this path under the commit sem */
1339	tmp_path->need_commit_sem = 0;
1340
1341	backref_ctx = kmalloc(sizeof(*backref_ctx), GFP_KERNEL);
1342	if (!backref_ctx) {
1343		ret = -ENOMEM;
1344		goto out;
1345	}
1346
1347	backref_ctx->path = tmp_path;
1348
1349	if (data_offset >= ino_size) {
1350		/*
1351		 * There may be extents that lie behind the file's size.
1352		 * I at least had this in combination with snapshotting while
1353		 * writing large files.
1354		 */
1355		ret = 0;
1356		goto out;
1357	}
1358
1359	fi = btrfs_item_ptr(eb, path->slots[0],
1360			struct btrfs_file_extent_item);
1361	extent_type = btrfs_file_extent_type(eb, fi);
1362	if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1363		ret = -ENOENT;
1364		goto out;
1365	}
1366	compressed = btrfs_file_extent_compression(eb, fi);
1367
1368	num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1369	disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1370	if (disk_byte == 0) {
1371		ret = -ENOENT;
1372		goto out;
1373	}
1374	logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1375
1376	down_read(&fs_info->commit_root_sem);
1377	ret = extent_from_logical(fs_info, disk_byte, tmp_path,
1378				  &found_key, &flags);
1379	up_read(&fs_info->commit_root_sem);
1380	btrfs_release_path(tmp_path);
1381
1382	if (ret < 0)
1383		goto out;
1384	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1385		ret = -EIO;
1386		goto out;
1387	}
1388
1389	/*
1390	 * Setup the clone roots.
1391	 */
1392	for (i = 0; i < sctx->clone_roots_cnt; i++) {
1393		cur_clone_root = sctx->clone_roots + i;
1394		cur_clone_root->ino = (u64)-1;
1395		cur_clone_root->offset = 0;
1396		cur_clone_root->found_refs = 0;
1397	}
1398
1399	backref_ctx->sctx = sctx;
1400	backref_ctx->found = 0;
1401	backref_ctx->cur_objectid = ino;
1402	backref_ctx->cur_offset = data_offset;
1403	backref_ctx->found_itself = 0;
1404	backref_ctx->extent_len = num_bytes;
1405	/*
1406	 * For non-compressed extents iterate_extent_inodes() gives us extent
1407	 * offsets that already take into account the data offset, but not for
1408	 * compressed extents, since the offset is logical and not relative to
1409	 * the physical extent locations. We must take this into account to
1410	 * avoid sending clone offsets that go beyond the source file's size,
1411	 * which would result in the clone ioctl failing with -EINVAL on the
1412	 * receiving end.
1413	 */
1414	if (compressed == BTRFS_COMPRESS_NONE)
1415		backref_ctx->data_offset = 0;
1416	else
1417		backref_ctx->data_offset = btrfs_file_extent_offset(eb, fi);
1418
1419	/*
1420	 * The last extent of a file may be too large due to page alignment.
1421	 * We need to adjust extent_len in this case so that the checks in
1422	 * __iterate_backrefs work.
1423	 */
1424	if (data_offset + num_bytes >= ino_size)
1425		backref_ctx->extent_len = ino_size - data_offset;
1426
1427	/*
1428	 * Now collect all backrefs.
1429	 */
1430	if (compressed == BTRFS_COMPRESS_NONE)
1431		extent_item_pos = logical - found_key.objectid;
1432	else
1433		extent_item_pos = 0;
1434	ret = iterate_extent_inodes(fs_info, found_key.objectid,
1435				    extent_item_pos, 1, __iterate_backrefs,
1436				    backref_ctx);
1437
1438	if (ret < 0)
1439		goto out;
1440
1441	if (!backref_ctx->found_itself) {
1442		/* found a bug in backref code? */
1443		ret = -EIO;
1444		btrfs_err(fs_info,
1445			  "did not find backref in send_root. inode=%llu, offset=%llu, disk_byte=%llu found extent=%llu",
1446			  ino, data_offset, disk_byte, found_key.objectid);
1447		goto out;
1448	}
1449
1450	btrfs_debug(fs_info,
1451		    "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1452		    data_offset, ino, num_bytes, logical);
1453
1454	if (!backref_ctx->found)
1455		btrfs_debug(fs_info, "no clones found");
1456
1457	cur_clone_root = NULL;
1458	for (i = 0; i < sctx->clone_roots_cnt; i++) {
1459		if (sctx->clone_roots[i].found_refs) {
1460			if (!cur_clone_root)
1461				cur_clone_root = sctx->clone_roots + i;
1462			else if (sctx->clone_roots[i].root == sctx->send_root)
1463				/* prefer clones from send_root over others */
1464				cur_clone_root = sctx->clone_roots + i;
1465		}
1466
1467	}
1468
1469	if (cur_clone_root) {
1470		*found = cur_clone_root;
1471		ret = 0;
1472	} else {
1473		ret = -ENOENT;
1474	}
1475
1476out:
1477	btrfs_free_path(tmp_path);
1478	kfree(backref_ctx);
1479	return ret;
1480}
1481
1482static int read_symlink(struct btrfs_root *root,
1483			u64 ino,
1484			struct fs_path *dest)
1485{
1486	int ret;
1487	struct btrfs_path *path;
1488	struct btrfs_key key;
1489	struct btrfs_file_extent_item *ei;
1490	u8 type;
1491	u8 compression;
1492	unsigned long off;
1493	int len;
1494
1495	path = alloc_path_for_send();
1496	if (!path)
1497		return -ENOMEM;
1498
1499	key.objectid = ino;
1500	key.type = BTRFS_EXTENT_DATA_KEY;
1501	key.offset = 0;
1502	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1503	if (ret < 0)
1504		goto out;
1505	if (ret) {
1506		/*
1507		 * An empty symlink inode. Can happen in rare error paths when
1508		 * creating a symlink (transaction committed before the inode
1509		 * eviction handler removed the symlink inode items and a crash
1510		 * happened in between or the subvol was snapshoted in between).
1511		 * Print an informative message to dmesg/syslog so that the user
1512		 * can delete the symlink.
1513		 */
1514		btrfs_err(root->fs_info,
1515			  "Found empty symlink inode %llu at root %llu",
1516			  ino, root->root_key.objectid);
1517		ret = -EIO;
1518		goto out;
1519	}
1520
1521	ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1522			struct btrfs_file_extent_item);
1523	type = btrfs_file_extent_type(path->nodes[0], ei);
1524	compression = btrfs_file_extent_compression(path->nodes[0], ei);
1525	BUG_ON(type != BTRFS_FILE_EXTENT_INLINE);
1526	BUG_ON(compression);
1527
1528	off = btrfs_file_extent_inline_start(ei);
1529	len = btrfs_file_extent_inline_len(path->nodes[0], path->slots[0], ei);
1530
1531	ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1532
1533out:
1534	btrfs_free_path(path);
1535	return ret;
1536}
1537
1538/*
1539 * Helper function to generate a file name that is unique in the root of
1540 * send_root and parent_root. This is used to generate names for orphan inodes.
1541 */
1542static int gen_unique_name(struct send_ctx *sctx,
1543			   u64 ino, u64 gen,
1544			   struct fs_path *dest)
1545{
1546	int ret = 0;
1547	struct btrfs_path *path;
1548	struct btrfs_dir_item *di;
1549	char tmp[64];
1550	int len;
1551	u64 idx = 0;
1552
1553	path = alloc_path_for_send();
1554	if (!path)
1555		return -ENOMEM;
1556
1557	while (1) {
1558		len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1559				ino, gen, idx);
1560		ASSERT(len < sizeof(tmp));
1561
1562		di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1563				path, BTRFS_FIRST_FREE_OBJECTID,
1564				tmp, strlen(tmp), 0);
1565		btrfs_release_path(path);
1566		if (IS_ERR(di)) {
1567			ret = PTR_ERR(di);
1568			goto out;
1569		}
1570		if (di) {
1571			/* not unique, try again */
1572			idx++;
1573			continue;
1574		}
1575
1576		if (!sctx->parent_root) {
1577			/* unique */
1578			ret = 0;
1579			break;
1580		}
1581
1582		di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1583				path, BTRFS_FIRST_FREE_OBJECTID,
1584				tmp, strlen(tmp), 0);
1585		btrfs_release_path(path);
1586		if (IS_ERR(di)) {
1587			ret = PTR_ERR(di);
1588			goto out;
1589		}
1590		if (di) {
1591			/* not unique, try again */
1592			idx++;
1593			continue;
1594		}
1595		/* unique */
1596		break;
1597	}
1598
1599	ret = fs_path_add(dest, tmp, strlen(tmp));
1600
1601out:
1602	btrfs_free_path(path);
1603	return ret;
1604}
1605
1606enum inode_state {
1607	inode_state_no_change,
1608	inode_state_will_create,
1609	inode_state_did_create,
1610	inode_state_will_delete,
1611	inode_state_did_delete,
1612};
1613
1614static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen)
1615{
1616	int ret;
1617	int left_ret;
1618	int right_ret;
1619	u64 left_gen;
1620	u64 right_gen;
1621
1622	ret = get_inode_info(sctx->send_root, ino, NULL, &left_gen, NULL, NULL,
1623			NULL, NULL);
1624	if (ret < 0 && ret != -ENOENT)
1625		goto out;
1626	left_ret = ret;
1627
1628	if (!sctx->parent_root) {
1629		right_ret = -ENOENT;
1630	} else {
1631		ret = get_inode_info(sctx->parent_root, ino, NULL, &right_gen,
1632				NULL, NULL, NULL, NULL);
1633		if (ret < 0 && ret != -ENOENT)
1634			goto out;
1635		right_ret = ret;
1636	}
1637
1638	if (!left_ret && !right_ret) {
1639		if (left_gen == gen && right_gen == gen) {
1640			ret = inode_state_no_change;
1641		} else if (left_gen == gen) {
1642			if (ino < sctx->send_progress)
1643				ret = inode_state_did_create;
1644			else
1645				ret = inode_state_will_create;
1646		} else if (right_gen == gen) {
1647			if (ino < sctx->send_progress)
1648				ret = inode_state_did_delete;
1649			else
1650				ret = inode_state_will_delete;
1651		} else  {
1652			ret = -ENOENT;
1653		}
1654	} else if (!left_ret) {
1655		if (left_gen == gen) {
1656			if (ino < sctx->send_progress)
1657				ret = inode_state_did_create;
1658			else
1659				ret = inode_state_will_create;
1660		} else {
1661			ret = -ENOENT;
1662		}
1663	} else if (!right_ret) {
1664		if (right_gen == gen) {
1665			if (ino < sctx->send_progress)
1666				ret = inode_state_did_delete;
1667			else
1668				ret = inode_state_will_delete;
1669		} else {
1670			ret = -ENOENT;
1671		}
1672	} else {
1673		ret = -ENOENT;
1674	}
1675
1676out:
1677	return ret;
1678}
1679
1680static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen)
1681{
1682	int ret;
1683
1684	ret = get_cur_inode_state(sctx, ino, gen);
1685	if (ret < 0)
1686		goto out;
1687
1688	if (ret == inode_state_no_change ||
1689	    ret == inode_state_did_create ||
1690	    ret == inode_state_will_delete)
1691		ret = 1;
1692	else
1693		ret = 0;
1694
1695out:
1696	return ret;
1697}
1698
1699/*
1700 * Helper function to lookup a dir item in a dir.
1701 */
1702static int lookup_dir_item_inode(struct btrfs_root *root,
1703				 u64 dir, const char *name, int name_len,
1704				 u64 *found_inode,
1705				 u8 *found_type)
1706{
1707	int ret = 0;
1708	struct btrfs_dir_item *di;
1709	struct btrfs_key key;
1710	struct btrfs_path *path;
1711
1712	path = alloc_path_for_send();
1713	if (!path)
1714		return -ENOMEM;
1715
1716	di = btrfs_lookup_dir_item(NULL, root, path,
1717			dir, name, name_len, 0);
1718	if (!di) {
1719		ret = -ENOENT;
1720		goto out;
1721	}
1722	if (IS_ERR(di)) {
1723		ret = PTR_ERR(di);
1724		goto out;
1725	}
1726	btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
1727	if (key.type == BTRFS_ROOT_ITEM_KEY) {
1728		ret = -ENOENT;
1729		goto out;
1730	}
1731	*found_inode = key.objectid;
1732	*found_type = btrfs_dir_type(path->nodes[0], di);
1733
1734out:
1735	btrfs_free_path(path);
1736	return ret;
1737}
1738
1739/*
1740 * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
1741 * generation of the parent dir and the name of the dir entry.
1742 */
1743static int get_first_ref(struct btrfs_root *root, u64 ino,
1744			 u64 *dir, u64 *dir_gen, struct fs_path *name)
1745{
1746	int ret;
1747	struct btrfs_key key;
1748	struct btrfs_key found_key;
1749	struct btrfs_path *path;
1750	int len;
1751	u64 parent_dir;
1752
1753	path = alloc_path_for_send();
1754	if (!path)
1755		return -ENOMEM;
1756
1757	key.objectid = ino;
1758	key.type = BTRFS_INODE_REF_KEY;
1759	key.offset = 0;
1760
1761	ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
1762	if (ret < 0)
1763		goto out;
1764	if (!ret)
1765		btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1766				path->slots[0]);
1767	if (ret || found_key.objectid != ino ||
1768	    (found_key.type != BTRFS_INODE_REF_KEY &&
1769	     found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1770		ret = -ENOENT;
1771		goto out;
1772	}
1773
1774	if (found_key.type == BTRFS_INODE_REF_KEY) {
1775		struct btrfs_inode_ref *iref;
1776		iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1777				      struct btrfs_inode_ref);
1778		len = btrfs_inode_ref_name_len(path->nodes[0], iref);
1779		ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1780						     (unsigned long)(iref + 1),
1781						     len);
1782		parent_dir = found_key.offset;
1783	} else {
1784		struct btrfs_inode_extref *extref;
1785		extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1786					struct btrfs_inode_extref);
1787		len = btrfs_inode_extref_name_len(path->nodes[0], extref);
1788		ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1789					(unsigned long)&extref->name, len);
1790		parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
1791	}
1792	if (ret < 0)
1793		goto out;
1794	btrfs_release_path(path);
1795
1796	if (dir_gen) {
1797		ret = get_inode_info(root, parent_dir, NULL, dir_gen, NULL,
1798				     NULL, NULL, NULL);
1799		if (ret < 0)
1800			goto out;
1801	}
1802
1803	*dir = parent_dir;
1804
1805out:
1806	btrfs_free_path(path);
1807	return ret;
1808}
1809
1810static int is_first_ref(struct btrfs_root *root,
1811			u64 ino, u64 dir,
1812			const char *name, int name_len)
1813{
1814	int ret;
1815	struct fs_path *tmp_name;
1816	u64 tmp_dir;
1817
1818	tmp_name = fs_path_alloc();
1819	if (!tmp_name)
1820		return -ENOMEM;
1821
1822	ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
1823	if (ret < 0)
1824		goto out;
1825
1826	if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
1827		ret = 0;
1828		goto out;
1829	}
1830
1831	ret = !memcmp(tmp_name->start, name, name_len);
1832
1833out:
1834	fs_path_free(tmp_name);
1835	return ret;
1836}
1837
1838/*
1839 * Used by process_recorded_refs to determine if a new ref would overwrite an
1840 * already existing ref. In case it detects an overwrite, it returns the
1841 * inode/gen in who_ino/who_gen.
1842 * When an overwrite is detected, process_recorded_refs does proper orphanizing
1843 * to make sure later references to the overwritten inode are possible.
1844 * Orphanizing is however only required for the first ref of an inode.
1845 * process_recorded_refs does an additional is_first_ref check to see if
1846 * orphanizing is really required.
1847 */
1848static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
1849			      const char *name, int name_len,
1850			      u64 *who_ino, u64 *who_gen)
1851{
1852	int ret = 0;
1853	u64 gen;
1854	u64 other_inode = 0;
1855	u8 other_type = 0;
1856
1857	if (!sctx->parent_root)
1858		goto out;
1859
1860	ret = is_inode_existent(sctx, dir, dir_gen);
1861	if (ret <= 0)
1862		goto out;
1863
1864	/*
1865	 * If we have a parent root we need to verify that the parent dir was
1866	 * not deleted and then re-created, if it was then we have no overwrite
1867	 * and we can just unlink this entry.
1868	 */
1869	if (sctx->parent_root) {
1870		ret = get_inode_info(sctx->parent_root, dir, NULL, &gen, NULL,
1871				     NULL, NULL, NULL);
1872		if (ret < 0 && ret != -ENOENT)
1873			goto out;
1874		if (ret) {
1875			ret = 0;
1876			goto out;
1877		}
1878		if (gen != dir_gen)
1879			goto out;
1880	}
1881
1882	ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
1883			&other_inode, &other_type);
1884	if (ret < 0 && ret != -ENOENT)
1885		goto out;
1886	if (ret) {
1887		ret = 0;
1888		goto out;
1889	}
1890
1891	/*
1892	 * Check if the overwritten ref was already processed. If yes, the ref
1893	 * was already unlinked/moved, so we can safely assume that we will not
1894	 * overwrite anything at this point in time.
1895	 */
1896	if (other_inode > sctx->send_progress ||
1897	    is_waiting_for_move(sctx, other_inode)) {
1898		ret = get_inode_info(sctx->parent_root, other_inode, NULL,
1899				who_gen, NULL, NULL, NULL, NULL);
1900		if (ret < 0)
1901			goto out;
1902
1903		ret = 1;
1904		*who_ino = other_inode;
1905	} else {
1906		ret = 0;
1907	}
1908
1909out:
1910	return ret;
1911}
1912
1913/*
1914 * Checks if the ref was overwritten by an already processed inode. This is
1915 * used by __get_cur_name_and_parent to find out if the ref was orphanized and
1916 * thus the orphan name needs be used.
1917 * process_recorded_refs also uses it to avoid unlinking of refs that were
1918 * overwritten.
1919 */
1920static int did_overwrite_ref(struct send_ctx *sctx,
1921			    u64 dir, u64 dir_gen,
1922			    u64 ino, u64 ino_gen,
1923			    const char *name, int name_len)
1924{
1925	int ret = 0;
1926	u64 gen;
1927	u64 ow_inode;
1928	u8 other_type;
1929
1930	if (!sctx->parent_root)
1931		goto out;
1932
1933	ret = is_inode_existent(sctx, dir, dir_gen);
1934	if (ret <= 0)
1935		goto out;
1936
1937	/* check if the ref was overwritten by another ref */
1938	ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
1939			&ow_inode, &other_type);
1940	if (ret < 0 && ret != -ENOENT)
1941		goto out;
1942	if (ret) {
1943		/* was never and will never be overwritten */
1944		ret = 0;
1945		goto out;
1946	}
1947
1948	ret = get_inode_info(sctx->send_root, ow_inode, NULL, &gen, NULL, NULL,
1949			NULL, NULL);
1950	if (ret < 0)
1951		goto out;
1952
1953	if (ow_inode == ino && gen == ino_gen) {
1954		ret = 0;
1955		goto out;
1956	}
1957
1958	/*
1959	 * We know that it is or will be overwritten. Check this now.
1960	 * The current inode being processed might have been the one that caused
1961	 * inode 'ino' to be orphanized, therefore check if ow_inode matches
1962	 * the current inode being processed.
1963	 */
1964	if ((ow_inode < sctx->send_progress) ||
1965	    (ino != sctx->cur_ino && ow_inode == sctx->cur_ino &&
1966	     gen == sctx->cur_inode_gen))
1967		ret = 1;
1968	else
1969		ret = 0;
1970
1971out:
1972	return ret;
1973}
1974
1975/*
1976 * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
1977 * that got overwritten. This is used by process_recorded_refs to determine
1978 * if it has to use the path as returned by get_cur_path or the orphan name.
1979 */
1980static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
1981{
1982	int ret = 0;
1983	struct fs_path *name = NULL;
1984	u64 dir;
1985	u64 dir_gen;
1986
1987	if (!sctx->parent_root)
1988		goto out;
1989
1990	name = fs_path_alloc();
1991	if (!name)
1992		return -ENOMEM;
1993
1994	ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
1995	if (ret < 0)
1996		goto out;
1997
1998	ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
1999			name->start, fs_path_len(name));
2000
2001out:
2002	fs_path_free(name);
2003	return ret;
2004}
2005
2006/*
2007 * Insert a name cache entry. On 32bit kernels the radix tree index is 32bit,
2008 * so we need to do some special handling in case we have clashes. This function
2009 * takes care of this with the help of name_cache_entry::radix_list.
2010 * In case of error, nce is kfreed.
2011 */
2012static int name_cache_insert(struct send_ctx *sctx,
2013			     struct name_cache_entry *nce)
2014{
2015	int ret = 0;
2016	struct list_head *nce_head;
2017
2018	nce_head = radix_tree_lookup(&sctx->name_cache,
2019			(unsigned long)nce->ino);
2020	if (!nce_head) {
2021		nce_head = kmalloc(sizeof(*nce_head), GFP_KERNEL);
2022		if (!nce_head) {
2023			kfree(nce);
2024			return -ENOMEM;
2025		}
2026		INIT_LIST_HEAD(nce_head);
2027
2028		ret = radix_tree_insert(&sctx->name_cache, nce->ino, nce_head);
2029		if (ret < 0) {
2030			kfree(nce_head);
2031			kfree(nce);
2032			return ret;
2033		}
2034	}
2035	list_add_tail(&nce->radix_list, nce_head);
2036	list_add_tail(&nce->list, &sctx->name_cache_list);
2037	sctx->name_cache_size++;
2038
2039	return ret;
2040}
2041
2042static void name_cache_delete(struct send_ctx *sctx,
2043			      struct name_cache_entry *nce)
2044{
2045	struct list_head *nce_head;
2046
2047	nce_head = radix_tree_lookup(&sctx->name_cache,
2048			(unsigned long)nce->ino);
2049	if (!nce_head) {
2050		btrfs_err(sctx->send_root->fs_info,
2051	      "name_cache_delete lookup failed ino %llu cache size %d, leaking memory",
2052			nce->ino, sctx->name_cache_size);
2053	}
2054
2055	list_del(&nce->radix_list);
2056	list_del(&nce->list);
2057	sctx->name_cache_size--;
2058
2059	/*
2060	 * We may not get to the final release of nce_head if the lookup fails
2061	 */
2062	if (nce_head && list_empty(nce_head)) {
2063		radix_tree_delete(&sctx->name_cache, (unsigned long)nce->ino);
2064		kfree(nce_head);
2065	}
2066}
2067
2068static struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2069						    u64 ino, u64 gen)
2070{
2071	struct list_head *nce_head;
2072	struct name_cache_entry *cur;
2073
2074	nce_head = radix_tree_lookup(&sctx->name_cache, (unsigned long)ino);
2075	if (!nce_head)
2076		return NULL;
2077
2078	list_for_each_entry(cur, nce_head, radix_list) {
2079		if (cur->ino == ino && cur->gen == gen)
2080			return cur;
2081	}
2082	return NULL;
2083}
2084
2085/*
2086 * Removes the entry from the list and adds it back to the end. This marks the
2087 * entry as recently used so that name_cache_clean_unused does not remove it.
2088 */
2089static void name_cache_used(struct send_ctx *sctx, struct name_cache_entry *nce)
2090{
2091	list_del(&nce->list);
2092	list_add_tail(&nce->list, &sctx->name_cache_list);
2093}
2094
2095/*
2096 * Remove some entries from the beginning of name_cache_list.
2097 */
2098static void name_cache_clean_unused(struct send_ctx *sctx)
2099{
2100	struct name_cache_entry *nce;
2101
2102	if (sctx->name_cache_size < SEND_CTX_NAME_CACHE_CLEAN_SIZE)
2103		return;
2104
2105	while (sctx->name_cache_size > SEND_CTX_MAX_NAME_CACHE_SIZE) {
2106		nce = list_entry(sctx->name_cache_list.next,
2107				struct name_cache_entry, list);
2108		name_cache_delete(sctx, nce);
2109		kfree(nce);
2110	}
2111}
2112
2113static void name_cache_free(struct send_ctx *sctx)
2114{
2115	struct name_cache_entry *nce;
2116
2117	while (!list_empty(&sctx->name_cache_list)) {
2118		nce = list_entry(sctx->name_cache_list.next,
2119				struct name_cache_entry, list);
2120		name_cache_delete(sctx, nce);
2121		kfree(nce);
2122	}
2123}
2124
2125/*
2126 * Used by get_cur_path for each ref up to the root.
2127 * Returns 0 if it succeeded.
2128 * Returns 1 if the inode is not existent or got overwritten. In that case, the
2129 * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2130 * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2131 * Returns <0 in case of error.
2132 */
2133static int __get_cur_name_and_parent(struct send_ctx *sctx,
2134				     u64 ino, u64 gen,
2135				     u64 *parent_ino,
2136				     u64 *parent_gen,
2137				     struct fs_path *dest)
2138{
2139	int ret;
2140	int nce_ret;
2141	struct name_cache_entry *nce = NULL;
2142
2143	/*
2144	 * First check if we already did a call to this function with the same
2145	 * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2146	 * return the cached result.
2147	 */
2148	nce = name_cache_search(sctx, ino, gen);
2149	if (nce) {
2150		if (ino < sctx->send_progress && nce->need_later_update) {
2151			name_cache_delete(sctx, nce);
2152			kfree(nce);
2153			nce = NULL;
2154		} else {
2155			name_cache_used(sctx, nce);
2156			*parent_ino = nce->parent_ino;
2157			*parent_gen = nce->parent_gen;
2158			ret = fs_path_add(dest, nce->name, nce->name_len);
2159			if (ret < 0)
2160				goto out;
2161			ret = nce->ret;
2162			goto out;
2163		}
2164	}
2165
2166	/*
2167	 * If the inode is not existent yet, add the orphan name and return 1.
2168	 * This should only happen for the parent dir that we determine in
2169	 * __record_new_ref
2170	 */
2171	ret = is_inode_existent(sctx, ino, gen);
2172	if (ret < 0)
2173		goto out;
2174
2175	if (!ret) {
2176		ret = gen_unique_name(sctx, ino, gen, dest);
2177		if (ret < 0)
2178			goto out;
2179		ret = 1;
2180		goto out_cache;
2181	}
2182
2183	/*
2184	 * Depending on whether the inode was already processed or not, use
2185	 * send_root or parent_root for ref lookup.
2186	 */
2187	if (ino < sctx->send_progress)
2188		ret = get_first_ref(sctx->send_root, ino,
2189				    parent_ino, parent_gen, dest);
2190	else
2191		ret = get_first_ref(sctx->parent_root, ino,
2192				    parent_ino, parent_gen, dest);
2193	if (ret < 0)
2194		goto out;
2195
2196	/*
2197	 * Check if the ref was overwritten by an inode's ref that was processed
2198	 * earlier. If yes, treat as orphan and return 1.
2199	 */
2200	ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2201			dest->start, dest->end - dest->start);
2202	if (ret < 0)
2203		goto out;
2204	if (ret) {
2205		fs_path_reset(dest);
2206		ret = gen_unique_name(sctx, ino, gen, dest);
2207		if (ret < 0)
2208			goto out;
2209		ret = 1;
2210	}
2211
2212out_cache:
2213	/*
2214	 * Store the result of the lookup in the name cache.
2215	 */
2216	nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2217	if (!nce) {
2218		ret = -ENOMEM;
2219		goto out;
2220	}
2221
2222	nce->ino = ino;
2223	nce->gen = gen;
2224	nce->parent_ino = *parent_ino;
2225	nce->parent_gen = *parent_gen;
2226	nce->name_len = fs_path_len(dest);
2227	nce->ret = ret;
2228	strcpy(nce->name, dest->start);
2229
2230	if (ino < sctx->send_progress)
2231		nce->need_later_update = 0;
2232	else
2233		nce->need_later_update = 1;
2234
2235	nce_ret = name_cache_insert(sctx, nce);
2236	if (nce_ret < 0)
2237		ret = nce_ret;
2238	name_cache_clean_unused(sctx);
2239
2240out:
2241	return ret;
2242}
2243
2244/*
2245 * Magic happens here. This function returns the first ref to an inode as it
2246 * would look like while receiving the stream at this point in time.
2247 * We walk the path up to the root. For every inode in between, we check if it
2248 * was already processed/sent. If yes, we continue with the parent as found
2249 * in send_root. If not, we continue with the parent as found in parent_root.
2250 * If we encounter an inode that was deleted at this point in time, we use the
2251 * inodes "orphan" name instead of the real name and stop. Same with new inodes
2252 * that were not created yet and overwritten inodes/refs.
2253 *
2254 * When do we have have orphan inodes:
2255 * 1. When an inode is freshly created and thus no valid refs are available yet
2256 * 2. When a directory lost all it's refs (deleted) but still has dir items
2257 *    inside which were not processed yet (pending for move/delete). If anyone
2258 *    tried to get the path to the dir items, it would get a path inside that
2259 *    orphan directory.
2260 * 3. When an inode is moved around or gets new links, it may overwrite the ref
2261 *    of an unprocessed inode. If in that case the first ref would be
2262 *    overwritten, the overwritten inode gets "orphanized". Later when we
2263 *    process this overwritten inode, it is restored at a new place by moving
2264 *    the orphan inode.
2265 *
2266 * sctx->send_progress tells this function at which point in time receiving
2267 * would be.
2268 */
2269static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2270			struct fs_path *dest)
2271{
2272	int ret = 0;
2273	struct fs_path *name = NULL;
2274	u64 parent_inode = 0;
2275	u64 parent_gen = 0;
2276	int stop = 0;
2277
2278	name = fs_path_alloc();
2279	if (!name) {
2280		ret = -ENOMEM;
2281		goto out;
2282	}
2283
2284	dest->reversed = 1;
2285	fs_path_reset(dest);
2286
2287	while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2288		struct waiting_dir_move *wdm;
2289
2290		fs_path_reset(name);
2291
2292		if (is_waiting_for_rm(sctx, ino)) {
2293			ret = gen_unique_name(sctx, ino, gen, name);
2294			if (ret < 0)
2295				goto out;
2296			ret = fs_path_add_path(dest, name);
2297			break;
2298		}
2299
2300		wdm = get_waiting_dir_move(sctx, ino);
2301		if (wdm && wdm->orphanized) {
2302			ret = gen_unique_name(sctx, ino, gen, name);
2303			stop = 1;
2304		} else if (wdm) {
2305			ret = get_first_ref(sctx->parent_root, ino,
2306					    &parent_inode, &parent_gen, name);
2307		} else {
2308			ret = __get_cur_name_and_parent(sctx, ino, gen,
2309							&parent_inode,
2310							&parent_gen, name);
2311			if (ret)
2312				stop = 1;
2313		}
2314
2315		if (ret < 0)
2316			goto out;
2317
2318		ret = fs_path_add_path(dest, name);
2319		if (ret < 0)
2320			goto out;
2321
2322		ino = parent_inode;
2323		gen = parent_gen;
2324	}
2325
2326out:
2327	fs_path_free(name);
2328	if (!ret)
2329		fs_path_unreverse(dest);
2330	return ret;
2331}
2332
2333/*
2334 * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2335 */
2336static int send_subvol_begin(struct send_ctx *sctx)
2337{
2338	int ret;
2339	struct btrfs_root *send_root = sctx->send_root;
2340	struct btrfs_root *parent_root = sctx->parent_root;
2341	struct btrfs_path *path;
2342	struct btrfs_key key;
2343	struct btrfs_root_ref *ref;
2344	struct extent_buffer *leaf;
2345	char *name = NULL;
2346	int namelen;
2347
2348	path = btrfs_alloc_path();
2349	if (!path)
2350		return -ENOMEM;
2351
2352	name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2353	if (!name) {
2354		btrfs_free_path(path);
2355		return -ENOMEM;
2356	}
2357
2358	key.objectid = send_root->objectid;
2359	key.type = BTRFS_ROOT_BACKREF_KEY;
2360	key.offset = 0;
2361
2362	ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2363				&key, path, 1, 0);
2364	if (ret < 0)
2365		goto out;
2366	if (ret) {
2367		ret = -ENOENT;
2368		goto out;
2369	}
2370
2371	leaf = path->nodes[0];
2372	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2373	if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2374	    key.objectid != send_root->objectid) {
2375		ret = -ENOENT;
2376		goto out;
2377	}
2378	ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2379	namelen = btrfs_root_ref_name_len(leaf, ref);
2380	read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2381	btrfs_release_path(path);
2382
2383	if (parent_root) {
2384		ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2385		if (ret < 0)
2386			goto out;
2387	} else {
2388		ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2389		if (ret < 0)
2390			goto out;
2391	}
2392
2393	TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2394
2395	if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2396		TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2397			    sctx->send_root->root_item.received_uuid);
2398	else
2399		TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2400			    sctx->send_root->root_item.uuid);
2401
2402	TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2403		    le64_to_cpu(sctx->send_root->root_item.ctransid));
2404	if (parent_root) {
2405		if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2406			TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2407				     parent_root->root_item.received_uuid);
2408		else
2409			TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2410				     parent_root->root_item.uuid);
2411		TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2412			    le64_to_cpu(sctx->parent_root->root_item.ctransid));
2413	}
2414
2415	ret = send_cmd(sctx);
2416
2417tlv_put_failure:
2418out:
2419	btrfs_free_path(path);
2420	kfree(name);
2421	return ret;
2422}
2423
2424static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2425{
2426	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2427	int ret = 0;
2428	struct fs_path *p;
2429
2430	btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2431
2432	p = fs_path_alloc();
2433	if (!p)
2434		return -ENOMEM;
2435
2436	ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2437	if (ret < 0)
2438		goto out;
2439
2440	ret = get_cur_path(sctx, ino, gen, p);
2441	if (ret < 0)
2442		goto out;
2443	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2444	TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2445
2446	ret = send_cmd(sctx);
2447
2448tlv_put_failure:
2449out:
2450	fs_path_free(p);
2451	return ret;
2452}
2453
2454static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2455{
2456	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2457	int ret = 0;
2458	struct fs_path *p;
2459
2460	btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2461
2462	p = fs_path_alloc();
2463	if (!p)
2464		return -ENOMEM;
2465
2466	ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2467	if (ret < 0)
2468		goto out;
2469
2470	ret = get_cur_path(sctx, ino, gen, p);
2471	if (ret < 0)
2472		goto out;
2473	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2474	TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2475
2476	ret = send_cmd(sctx);
2477
2478tlv_put_failure:
2479out:
2480	fs_path_free(p);
2481	return ret;
2482}
2483
2484static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2485{
2486	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2487	int ret = 0;
2488	struct fs_path *p;
2489
2490	btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2491		    ino, uid, gid);
2492
2493	p = fs_path_alloc();
2494	if (!p)
2495		return -ENOMEM;
2496
2497	ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2498	if (ret < 0)
2499		goto out;
2500
2501	ret = get_cur_path(sctx, ino, gen, p);
2502	if (ret < 0)
2503		goto out;
2504	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2505	TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2506	TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2507
2508	ret = send_cmd(sctx);
2509
2510tlv_put_failure:
2511out:
2512	fs_path_free(p);
2513	return ret;
2514}
2515
2516static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2517{
2518	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2519	int ret = 0;
2520	struct fs_path *p = NULL;
2521	struct btrfs_inode_item *ii;
2522	struct btrfs_path *path = NULL;
2523	struct extent_buffer *eb;
2524	struct btrfs_key key;
2525	int slot;
2526
2527	btrfs_debug(fs_info, "send_utimes %llu", ino);
2528
2529	p = fs_path_alloc();
2530	if (!p)
2531		return -ENOMEM;
2532
2533	path = alloc_path_for_send();
2534	if (!path) {
2535		ret = -ENOMEM;
2536		goto out;
2537	}
2538
2539	key.objectid = ino;
2540	key.type = BTRFS_INODE_ITEM_KEY;
2541	key.offset = 0;
2542	ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2543	if (ret > 0)
2544		ret = -ENOENT;
2545	if (ret < 0)
2546		goto out;
2547
2548	eb = path->nodes[0];
2549	slot = path->slots[0];
2550	ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2551
2552	ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2553	if (ret < 0)
2554		goto out;
2555
2556	ret = get_cur_path(sctx, ino, gen, p);
2557	if (ret < 0)
2558		goto out;
2559	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2560	TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2561	TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2562	TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2563	/* TODO Add otime support when the otime patches get into upstream */
2564
2565	ret = send_cmd(sctx);
2566
2567tlv_put_failure:
2568out:
2569	fs_path_free(p);
2570	btrfs_free_path(path);
2571	return ret;
2572}
2573
2574/*
2575 * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2576 * a valid path yet because we did not process the refs yet. So, the inode
2577 * is created as orphan.
2578 */
2579static int send_create_inode(struct send_ctx *sctx, u64 ino)
2580{
2581	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2582	int ret = 0;
2583	struct fs_path *p;
2584	int cmd;
2585	u64 gen;
2586	u64 mode;
2587	u64 rdev;
2588
2589	btrfs_debug(fs_info, "send_create_inode %llu", ino);
2590
2591	p = fs_path_alloc();
2592	if (!p)
2593		return -ENOMEM;
2594
2595	if (ino != sctx->cur_ino) {
2596		ret = get_inode_info(sctx->send_root, ino, NULL, &gen, &mode,
2597				     NULL, NULL, &rdev);
2598		if (ret < 0)
2599			goto out;
2600	} else {
2601		gen = sctx->cur_inode_gen;
2602		mode = sctx->cur_inode_mode;
2603		rdev = sctx->cur_inode_rdev;
2604	}
2605
2606	if (S_ISREG(mode)) {
2607		cmd = BTRFS_SEND_C_MKFILE;
2608	} else if (S_ISDIR(mode)) {
2609		cmd = BTRFS_SEND_C_MKDIR;
2610	} else if (S_ISLNK(mode)) {
2611		cmd = BTRFS_SEND_C_SYMLINK;
2612	} else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2613		cmd = BTRFS_SEND_C_MKNOD;
2614	} else if (S_ISFIFO(mode)) {
2615		cmd = BTRFS_SEND_C_MKFIFO;
2616	} else if (S_ISSOCK(mode)) {
2617		cmd = BTRFS_SEND_C_MKSOCK;
2618	} else {
2619		btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2620				(int)(mode & S_IFMT));
2621		ret = -ENOTSUPP;
2622		goto out;
2623	}
2624
2625	ret = begin_cmd(sctx, cmd);
2626	if (ret < 0)
2627		goto out;
2628
2629	ret = gen_unique_name(sctx, ino, gen, p);
2630	if (ret < 0)
2631		goto out;
2632
2633	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2634	TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2635
2636	if (S_ISLNK(mode)) {
2637		fs_path_reset(p);
2638		ret = read_symlink(sctx->send_root, ino, p);
2639		if (ret < 0)
2640			goto out;
2641		TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2642	} else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2643		   S_ISFIFO(mode) || S_ISSOCK(mode)) {
2644		TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2645		TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2646	}
2647
2648	ret = send_cmd(sctx);
2649	if (ret < 0)
2650		goto out;
2651
2652
2653tlv_put_failure:
2654out:
2655	fs_path_free(p);
2656	return ret;
2657}
2658
2659/*
2660 * We need some special handling for inodes that get processed before the parent
2661 * directory got created. See process_recorded_refs for details.
2662 * This function does the check if we already created the dir out of order.
2663 */
2664static int did_create_dir(struct send_ctx *sctx, u64 dir)
2665{
2666	int ret = 0;
2667	struct btrfs_path *path = NULL;
2668	struct btrfs_key key;
2669	struct btrfs_key found_key;
2670	struct btrfs_key di_key;
2671	struct extent_buffer *eb;
2672	struct btrfs_dir_item *di;
2673	int slot;
2674
2675	path = alloc_path_for_send();
2676	if (!path) {
2677		ret = -ENOMEM;
2678		goto out;
2679	}
2680
2681	key.objectid = dir;
2682	key.type = BTRFS_DIR_INDEX_KEY;
2683	key.offset = 0;
2684	ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2685	if (ret < 0)
2686		goto out;
2687
2688	while (1) {
2689		eb = path->nodes[0];
2690		slot = path->slots[0];
2691		if (slot >= btrfs_header_nritems(eb)) {
2692			ret = btrfs_next_leaf(sctx->send_root, path);
2693			if (ret < 0) {
2694				goto out;
2695			} else if (ret > 0) {
2696				ret = 0;
2697				break;
2698			}
2699			continue;
2700		}
2701
2702		btrfs_item_key_to_cpu(eb, &found_key, slot);
2703		if (found_key.objectid != key.objectid ||
2704		    found_key.type != key.type) {
2705			ret = 0;
2706			goto out;
2707		}
2708
2709		di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
2710		btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2711
2712		if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2713		    di_key.objectid < sctx->send_progress) {
2714			ret = 1;
2715			goto out;
2716		}
2717
2718		path->slots[0]++;
2719	}
2720
2721out:
2722	btrfs_free_path(path);
2723	return ret;
2724}
2725
2726/*
2727 * Only creates the inode if it is:
2728 * 1. Not a directory
2729 * 2. Or a directory which was not created already due to out of order
2730 *    directories. See did_create_dir and process_recorded_refs for details.
2731 */
2732static int send_create_inode_if_needed(struct send_ctx *sctx)
2733{
2734	int ret;
2735
2736	if (S_ISDIR(sctx->cur_inode_mode)) {
2737		ret = did_create_dir(sctx, sctx->cur_ino);
2738		if (ret < 0)
2739			goto out;
2740		if (ret) {
2741			ret = 0;
2742			goto out;
2743		}
2744	}
2745
2746	ret = send_create_inode(sctx, sctx->cur_ino);
2747	if (ret < 0)
2748		goto out;
2749
2750out:
2751	return ret;
2752}
2753
2754struct recorded_ref {
2755	struct list_head list;
2756	char *dir_path;
2757	char *name;
2758	struct fs_path *full_path;
2759	u64 dir;
2760	u64 dir_gen;
2761	int dir_path_len;
2762	int name_len;
2763};
2764
2765/*
2766 * We need to process new refs before deleted refs, but compare_tree gives us
2767 * everything mixed. So we first record all refs and later process them.
2768 * This function is a helper to record one ref.
2769 */
2770static int __record_ref(struct list_head *head, u64 dir,
2771		      u64 dir_gen, struct fs_path *path)
2772{
2773	struct recorded_ref *ref;
2774
2775	ref = kmalloc(sizeof(*ref), GFP_KERNEL);
2776	if (!ref)
2777		return -ENOMEM;
2778
2779	ref->dir = dir;
2780	ref->dir_gen = dir_gen;
2781	ref->full_path = path;
2782
2783	ref->name = (char *)kbasename(ref->full_path->start);
2784	ref->name_len = ref->full_path->end - ref->name;
2785	ref->dir_path = ref->full_path->start;
2786	if (ref->name == ref->full_path->start)
2787		ref->dir_path_len = 0;
2788	else
2789		ref->dir_path_len = ref->full_path->end -
2790				ref->full_path->start - 1 - ref->name_len;
2791
2792	list_add_tail(&ref->list, head);
2793	return 0;
2794}
2795
2796static int dup_ref(struct recorded_ref *ref, struct list_head *list)
2797{
2798	struct recorded_ref *new;
2799
2800	new = kmalloc(sizeof(*ref), GFP_KERNEL);
2801	if (!new)
2802		return -ENOMEM;
2803
2804	new->dir = ref->dir;
2805	new->dir_gen = ref->dir_gen;
2806	new->full_path = NULL;
2807	INIT_LIST_HEAD(&new->list);
2808	list_add_tail(&new->list, list);
2809	return 0;
2810}
2811
2812static void __free_recorded_refs(struct list_head *head)
2813{
2814	struct recorded_ref *cur;
2815
2816	while (!list_empty(head)) {
2817		cur = list_entry(head->next, struct recorded_ref, list);
2818		fs_path_free(cur->full_path);
2819		list_del(&cur->list);
2820		kfree(cur);
2821	}
2822}
2823
2824static void free_recorded_refs(struct send_ctx *sctx)
2825{
2826	__free_recorded_refs(&sctx->new_refs);
2827	__free_recorded_refs(&sctx->deleted_refs);
2828}
2829
2830/*
2831 * Renames/moves a file/dir to its orphan name. Used when the first
2832 * ref of an unprocessed inode gets overwritten and for all non empty
2833 * directories.
2834 */
2835static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
2836			  struct fs_path *path)
2837{
2838	int ret;
2839	struct fs_path *orphan;
2840
2841	orphan = fs_path_alloc();
2842	if (!orphan)
2843		return -ENOMEM;
2844
2845	ret = gen_unique_name(sctx, ino, gen, orphan);
2846	if (ret < 0)
2847		goto out;
2848
2849	ret = send_rename(sctx, path, orphan);
2850
2851out:
2852	fs_path_free(orphan);
2853	return ret;
2854}
2855
2856static struct orphan_dir_info *
2857add_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2858{
2859	struct rb_node **p = &sctx->orphan_dirs.rb_node;
2860	struct rb_node *parent = NULL;
2861	struct orphan_dir_info *entry, *odi;
2862
2863	odi = kmalloc(sizeof(*odi), GFP_KERNEL);
2864	if (!odi)
2865		return ERR_PTR(-ENOMEM);
2866	odi->ino = dir_ino;
2867	odi->gen = 0;
2868
2869	while (*p) {
2870		parent = *p;
2871		entry = rb_entry(parent, struct orphan_dir_info, node);
2872		if (dir_ino < entry->ino) {
2873			p = &(*p)->rb_left;
2874		} else if (dir_ino > entry->ino) {
2875			p = &(*p)->rb_right;
2876		} else {
2877			kfree(odi);
2878			return entry;
2879		}
2880	}
2881
2882	rb_link_node(&odi->node, parent, p);
2883	rb_insert_color(&odi->node, &sctx->orphan_dirs);
2884	return odi;
2885}
2886
2887static struct orphan_dir_info *
2888get_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2889{
2890	struct rb_node *n = sctx->orphan_dirs.rb_node;
2891	struct orphan_dir_info *entry;
2892
2893	while (n) {
2894		entry = rb_entry(n, struct orphan_dir_info, node);
2895		if (dir_ino < entry->ino)
2896			n = n->rb_left;
2897		else if (dir_ino > entry->ino)
2898			n = n->rb_right;
2899		else
2900			return entry;
2901	}
2902	return NULL;
2903}
2904
2905static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino)
2906{
2907	struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino);
2908
2909	return odi != NULL;
2910}
2911
2912static void free_orphan_dir_info(struct send_ctx *sctx,
2913				 struct orphan_dir_info *odi)
2914{
2915	if (!odi)
2916		return;
2917	rb_erase(&odi->node, &sctx->orphan_dirs);
2918	kfree(odi);
2919}
2920
2921/*
2922 * Returns 1 if a directory can be removed at this point in time.
2923 * We check this by iterating all dir items and checking if the inode behind
2924 * the dir item was already processed.
2925 */
2926static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2927		     u64 send_progress)
2928{
2929	int ret = 0;
2930	struct btrfs_root *root = sctx->parent_root;
2931	struct btrfs_path *path;
2932	struct btrfs_key key;
2933	struct btrfs_key found_key;
2934	struct btrfs_key loc;
2935	struct btrfs_dir_item *di;
2936
2937	/*
2938	 * Don't try to rmdir the top/root subvolume dir.
2939	 */
2940	if (dir == BTRFS_FIRST_FREE_OBJECTID)
2941		return 0;
2942
2943	path = alloc_path_for_send();
2944	if (!path)
2945		return -ENOMEM;
2946
2947	key.objectid = dir;
2948	key.type = BTRFS_DIR_INDEX_KEY;
2949	key.offset = 0;
2950	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2951	if (ret < 0)
2952		goto out;
2953
2954	while (1) {
2955		struct waiting_dir_move *dm;
2956
2957		if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2958			ret = btrfs_next_leaf(root, path);
2959			if (ret < 0)
2960				goto out;
2961			else if (ret > 0)
2962				break;
2963			continue;
2964		}
2965		btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2966				      path->slots[0]);
2967		if (found_key.objectid != key.objectid ||
2968		    found_key.type != key.type)
2969			break;
2970
2971		di = btrfs_item_ptr(path->nodes[0], path->slots[0],
2972				struct btrfs_dir_item);
2973		btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
2974
2975		dm = get_waiting_dir_move(sctx, loc.objectid);
2976		if (dm) {
2977			struct orphan_dir_info *odi;
2978
2979			odi = add_orphan_dir_info(sctx, dir);
2980			if (IS_ERR(odi)) {
2981				ret = PTR_ERR(odi);
2982				goto out;
2983			}
2984			odi->gen = dir_gen;
2985			dm->rmdir_ino = dir;
2986			ret = 0;
2987			goto out;
2988		}
2989
2990		if (loc.objectid > send_progress) {
2991			struct orphan_dir_info *odi;
2992
2993			odi = get_orphan_dir_info(sctx, dir);
2994			free_orphan_dir_info(sctx, odi);
2995			ret = 0;
2996			goto out;
2997		}
2998
2999		path->slots[0]++;
3000	}
3001
3002	ret = 1;
3003
3004out:
3005	btrfs_free_path(path);
3006	return ret;
3007}
3008
3009static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3010{
3011	struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3012
3013	return entry != NULL;
3014}
3015
3016static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3017{
3018	struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3019	struct rb_node *parent = NULL;
3020	struct waiting_dir_move *entry, *dm;
3021
3022	dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3023	if (!dm)
3024		return -ENOMEM;
3025	dm->ino = ino;
3026	dm->rmdir_ino = 0;
3027	dm->orphanized = orphanized;
3028
3029	while (*p) {
3030		parent = *p;
3031		entry = rb_entry(parent, struct waiting_dir_move, node);
3032		if (ino < entry->ino) {
3033			p = &(*p)->rb_left;
3034		} else if (ino > entry->ino) {
3035			p = &(*p)->rb_right;
3036		} else {
3037			kfree(dm);
3038			return -EEXIST;
3039		}
3040	}
3041
3042	rb_link_node(&dm->node, parent, p);
3043	rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3044	return 0;
3045}
3046
3047static struct waiting_dir_move *
3048get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3049{
3050	struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3051	struct waiting_dir_move *entry;
3052
3053	while (n) {
3054		entry = rb_entry(n, struct waiting_dir_move, node);
3055		if (ino < entry->ino)
3056			n = n->rb_left;
3057		else if (ino > entry->ino)
3058			n = n->rb_right;
3059		else
3060			return entry;
3061	}
3062	return NULL;
3063}
3064
3065static void free_waiting_dir_move(struct send_ctx *sctx,
3066				  struct waiting_dir_move *dm)
3067{
3068	if (!dm)
3069		return;
3070	rb_erase(&dm->node, &sctx->waiting_dir_moves);
3071	kfree(dm);
3072}
3073
3074static int add_pending_dir_move(struct send_ctx *sctx,
3075				u64 ino,
3076				u64 ino_gen,
3077				u64 parent_ino,
3078				struct list_head *new_refs,
3079				struct list_head *deleted_refs,
3080				const bool is_orphan)
3081{
3082	struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3083	struct rb_node *parent = NULL;
3084	struct pending_dir_move *entry = NULL, *pm;
3085	struct recorded_ref *cur;
3086	int exists = 0;
3087	int ret;
3088
3089	pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3090	if (!pm)
3091		return -ENOMEM;
3092	pm->parent_ino = parent_ino;
3093	pm->ino = ino;
3094	pm->gen = ino_gen;
3095	INIT_LIST_HEAD(&pm->list);
3096	INIT_LIST_HEAD(&pm->update_refs);
3097	RB_CLEAR_NODE(&pm->node);
3098
3099	while (*p) {
3100		parent = *p;
3101		entry = rb_entry(parent, struct pending_dir_move, node);
3102		if (parent_ino < entry->parent_ino) {
3103			p = &(*p)->rb_left;
3104		} else if (parent_ino > entry->parent_ino) {
3105			p = &(*p)->rb_right;
3106		} else {
3107			exists = 1;
3108			break;
3109		}
3110	}
3111
3112	list_for_each_entry(cur, deleted_refs, list) {
3113		ret = dup_ref(cur, &pm->update_refs);
3114		if (ret < 0)
3115			goto out;
3116	}
3117	list_for_each_entry(cur, new_refs, list) {
3118		ret = dup_ref(cur, &pm->update_refs);
3119		if (ret < 0)
3120			goto out;
3121	}
3122
3123	ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3124	if (ret)
3125		goto out;
3126
3127	if (exists) {
3128		list_add_tail(&pm->list, &entry->list);
3129	} else {
3130		rb_link_node(&pm->node, parent, p);
3131		rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3132	}
3133	ret = 0;
3134out:
3135	if (ret) {
3136		__free_recorded_refs(&pm->update_refs);
3137		kfree(pm);
3138	}
3139	return ret;
3140}
3141
3142static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3143						      u64 parent_ino)
3144{
3145	struct rb_node *n = sctx->pending_dir_moves.rb_node;
3146	struct pending_dir_move *entry;
3147
3148	while (n) {
3149		entry = rb_entry(n, struct pending_dir_move, node);
3150		if (parent_ino < entry->parent_ino)
3151			n = n->rb_left;
3152		else if (parent_ino > entry->parent_ino)
3153			n = n->rb_right;
3154		else
3155			return entry;
3156	}
3157	return NULL;
3158}
3159
3160static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3161		     u64 ino, u64 gen, u64 *ancestor_ino)
3162{
3163	int ret = 0;
3164	u64 parent_inode = 0;
3165	u64 parent_gen = 0;
3166	u64 start_ino = ino;
3167
3168	*ancestor_ino = 0;
3169	while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3170		fs_path_reset(name);
3171
3172		if (is_waiting_for_rm(sctx, ino))
3173			break;
3174		if (is_waiting_for_move(sctx, ino)) {
3175			if (*ancestor_ino == 0)
3176				*ancestor_ino = ino;
3177			ret = get_first_ref(sctx->parent_root, ino,
3178					    &parent_inode, &parent_gen, name);
3179		} else {
3180			ret = __get_cur_name_and_parent(sctx, ino, gen,
3181							&parent_inode,
3182							&parent_gen, name);
3183			if (ret > 0) {
3184				ret = 0;
3185				break;
3186			}
3187		}
3188		if (ret < 0)
3189			break;
3190		if (parent_inode == start_ino) {
3191			ret = 1;
3192			if (*ancestor_ino == 0)
3193				*ancestor_ino = ino;
3194			break;
3195		}
3196		ino = parent_inode;
3197		gen = parent_gen;
3198	}
3199	return ret;
3200}
3201
3202static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3203{
3204	struct fs_path *from_path = NULL;
3205	struct fs_path *to_path = NULL;
3206	struct fs_path *name = NULL;
3207	u64 orig_progress = sctx->send_progress;
3208	struct recorded_ref *cur;
3209	u64 parent_ino, parent_gen;
3210	struct waiting_dir_move *dm = NULL;
3211	u64 rmdir_ino = 0;
3212	u64 ancestor;
3213	bool is_orphan;
3214	int ret;
3215
3216	name = fs_path_alloc();
3217	from_path = fs_path_alloc();
3218	if (!name || !from_path) {
3219		ret = -ENOMEM;
3220		goto out;
3221	}
3222
3223	dm = get_waiting_dir_move(sctx, pm->ino);
3224	ASSERT(dm);
3225	rmdir_ino = dm->rmdir_ino;
3226	is_orphan = dm->orphanized;
3227	free_waiting_dir_move(sctx, dm);
3228
3229	if (is_orphan) {
3230		ret = gen_unique_name(sctx, pm->ino,
3231				      pm->gen, from_path);
3232	} else {
3233		ret = get_first_ref(sctx->parent_root, pm->ino,
3234				    &parent_ino, &parent_gen, name);
3235		if (ret < 0)
3236			goto out;
3237		ret = get_cur_path(sctx, parent_ino, parent_gen,
3238				   from_path);
3239		if (ret < 0)
3240			goto out;
3241		ret = fs_path_add_path(from_path, name);
3242	}
3243	if (ret < 0)
3244		goto out;
3245
3246	sctx->send_progress = sctx->cur_ino + 1;
3247	ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3248	if (ret < 0)
3249		goto out;
3250	if (ret) {
3251		LIST_HEAD(deleted_refs);
3252		ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3253		ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3254					   &pm->update_refs, &deleted_refs,
3255					   is_orphan);
3256		if (ret < 0)
3257			goto out;
3258		if (rmdir_ino) {
3259			dm = get_waiting_dir_move(sctx, pm->ino);
3260			ASSERT(dm);
3261			dm->rmdir_ino = rmdir_ino;
3262		}
3263		goto out;
3264	}
3265	fs_path_reset(name);
3266	to_path = name;
3267	name = NULL;
3268	ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3269	if (ret < 0)
3270		goto out;
3271
3272	ret = send_rename(sctx, from_path, to_path);
3273	if (ret < 0)
3274		goto out;
3275
3276	if (rmdir_ino) {
3277		struct orphan_dir_info *odi;
3278
3279		odi = get_orphan_dir_info(sctx, rmdir_ino);
3280		if (!odi) {
3281			/* already deleted */
3282			goto finish;
3283		}
3284		ret = can_rmdir(sctx, rmdir_ino, odi->gen, sctx->cur_ino);
3285		if (ret < 0)
3286			goto out;
3287		if (!ret)
3288			goto finish;
3289
3290		name = fs_path_alloc();
3291		if (!name) {
3292			ret = -ENOMEM;
3293			goto out;
3294		}
3295		ret = get_cur_path(sctx, rmdir_ino, odi->gen, name);
3296		if (ret < 0)
3297			goto out;
3298		ret = send_rmdir(sctx, name);
3299		if (ret < 0)
3300			goto out;
3301		free_orphan_dir_info(sctx, odi);
3302	}
3303
3304finish:
3305	ret = send_utimes(sctx, pm->ino, pm->gen);
3306	if (ret < 0)
3307		goto out;
3308
3309	/*
3310	 * After rename/move, need to update the utimes of both new parent(s)
3311	 * and old parent(s).
3312	 */
3313	list_for_each_entry(cur, &pm->update_refs, list) {
3314		/*
3315		 * The parent inode might have been deleted in the send snapshot
3316		 */
3317		ret = get_inode_info(sctx->send_root, cur->dir, NULL,
3318				     NULL, NULL, NULL, NULL, NULL);
3319		if (ret == -ENOENT) {
3320			ret = 0;
3321			continue;
3322		}
3323		if (ret < 0)
3324			goto out;
3325
3326		ret = send_utimes(sctx, cur->dir, cur->dir_gen);
3327		if (ret < 0)
3328			goto out;
3329	}
3330
3331out:
3332	fs_path_free(name);
3333	fs_path_free(from_path);
3334	fs_path_free(to_path);
3335	sctx->send_progress = orig_progress;
3336
3337	return ret;
3338}
3339
3340static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3341{
3342	if (!list_empty(&m->list))
3343		list_del(&m->list);
3344	if (!RB_EMPTY_NODE(&m->node))
3345		rb_erase(&m->node, &sctx->pending_dir_moves);
3346	__free_recorded_refs(&m->update_refs);
3347	kfree(m);
3348}
3349
3350static void tail_append_pending_moves(struct pending_dir_move *moves,
3351				      struct list_head *stack)
3352{
3353	if (list_empty(&moves->list)) {
3354		list_add_tail(&moves->list, stack);
3355	} else {
3356		LIST_HEAD(list);
3357		list_splice_init(&moves->list, &list);
3358		list_add_tail(&moves->list, stack);
3359		list_splice_tail(&list, stack);
3360	}
3361}
3362
3363static int apply_children_dir_moves(struct send_ctx *sctx)
3364{
3365	struct pending_dir_move *pm;
3366	struct list_head stack;
3367	u64 parent_ino = sctx->cur_ino;
3368	int ret = 0;
3369
3370	pm = get_pending_dir_moves(sctx, parent_ino);
3371	if (!pm)
3372		return 0;
3373
3374	INIT_LIST_HEAD(&stack);
3375	tail_append_pending_moves(pm, &stack);
3376
3377	while (!list_empty(&stack)) {
3378		pm = list_first_entry(&stack, struct pending_dir_move, list);
3379		parent_ino = pm->ino;
3380		ret = apply_dir_move(sctx, pm);
3381		free_pending_move(sctx, pm);
3382		if (ret)
3383			goto out;
3384		pm = get_pending_dir_moves(sctx, parent_ino);
3385		if (pm)
3386			tail_append_pending_moves(pm, &stack);
3387	}
3388	return 0;
3389
3390out:
3391	while (!list_empty(&stack)) {
3392		pm = list_first_entry(&stack, struct pending_dir_move, list);
3393		free_pending_move(sctx, pm);
3394	}
3395	return ret;
3396}
3397
3398/*
3399 * We might need to delay a directory rename even when no ancestor directory
3400 * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3401 * renamed. This happens when we rename a directory to the old name (the name
3402 * in the parent root) of some other unrelated directory that got its rename
3403 * delayed due to some ancestor with higher number that got renamed.
3404 *
3405 * Example:
3406 *
3407 * Parent snapshot:
3408 * .                                       (ino 256)
3409 * |---- a/                                (ino 257)
3410 * |     |---- file                        (ino 260)
3411 * |
3412 * |---- b/                                (ino 258)
3413 * |---- c/                                (ino 259)
3414 *
3415 * Send snapshot:
3416 * .                                       (ino 256)
3417 * |---- a/                                (ino 258)
3418 * |---- x/                                (ino 259)
3419 *       |---- y/                          (ino 257)
3420 *             |----- file                 (ino 260)
3421 *
3422 * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3423 * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3424 * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3425 * must issue is:
3426 *
3427 * 1 - rename 259 from 'c' to 'x'
3428 * 2 - rename 257 from 'a' to 'x/y'
3429 * 3 - rename 258 from 'b' to 'a'
3430 *
3431 * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3432 * be done right away and < 0 on error.
3433 */
3434static int wait_for_dest_dir_move(struct send_ctx *sctx,
3435				  struct recorded_ref *parent_ref,
3436				  const bool is_orphan)
3437{
3438	struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3439	struct btrfs_path *path;
3440	struct btrfs_key key;
3441	struct btrfs_key di_key;
3442	struct btrfs_dir_item *di;
3443	u64 left_gen;
3444	u64 right_gen;
3445	int ret = 0;
3446	struct waiting_dir_move *wdm;
3447
3448	if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3449		return 0;
3450
3451	path = alloc_path_for_send();
3452	if (!path)
3453		return -ENOMEM;
3454
3455	key.objectid = parent_ref->dir;
3456	key.type = BTRFS_DIR_ITEM_KEY;
3457	key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3458
3459	ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3460	if (ret < 0) {
3461		goto out;
3462	} else if (ret > 0) {
3463		ret = 0;
3464		goto out;
3465	}
3466
3467	di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3468				       parent_ref->name_len);
3469	if (!di) {
3470		ret = 0;
3471		goto out;
3472	}
3473	/*
3474	 * di_key.objectid has the number of the inode that has a dentry in the
3475	 * parent directory with the same name that sctx->cur_ino is being
3476	 * renamed to. We need to check if that inode is in the send root as
3477	 * well and if it is currently marked as an inode with a pending rename,
3478	 * if it is, we need to delay the rename of sctx->cur_ino as well, so
3479	 * that it happens after that other inode is renamed.
3480	 */
3481	btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3482	if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3483		ret = 0;
3484		goto out;
3485	}
3486
3487	ret = get_inode_info(sctx->parent_root, di_key.objectid, NULL,
3488			     &left_gen, NULL, NULL, NULL, NULL);
3489	if (ret < 0)
3490		goto out;
3491	ret = get_inode_info(sctx->send_root, di_key.objectid, NULL,
3492			     &right_gen, NULL, NULL, NULL, NULL);
3493	if (ret < 0) {
3494		if (ret == -ENOENT)
3495			ret = 0;
3496		goto out;
3497	}
3498
3499	/* Different inode, no need to delay the rename of sctx->cur_ino */
3500	if (right_gen != left_gen) {
3501		ret = 0;
3502		goto out;
3503	}
3504
3505	wdm = get_waiting_dir_move(sctx, di_key.objectid);
3506	if (wdm && !wdm->orphanized) {
3507		ret = add_pending_dir_move(sctx,
3508					   sctx->cur_ino,
3509					   sctx->cur_inode_gen,
3510					   di_key.objectid,
3511					   &sctx->new_refs,
3512					   &sctx->deleted_refs,
3513					   is_orphan);
3514		if (!ret)
3515			ret = 1;
3516	}
3517out:
3518	btrfs_free_path(path);
3519	return ret;
3520}
3521
3522/*
3523 * Check if ino ino1 is an ancestor of inode ino2 in the given root.
3524 * Return 1 if true, 0 if false and < 0 on error.
3525 */
3526static int is_ancestor(struct btrfs_root *root,
3527		       const u64 ino1,
3528		       const u64 ino1_gen,
3529		       const u64 ino2,
3530		       struct fs_path *fs_path)
3531{
3532	u64 ino = ino2;
3533
3534	while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3535		int ret;
3536		u64 parent;
3537		u64 parent_gen;
3538
3539		fs_path_reset(fs_path);
3540		ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3541		if (ret < 0) {
3542			if (ret == -ENOENT && ino == ino2)
3543				ret = 0;
3544			return ret;
3545		}
3546		if (parent == ino1)
3547			return parent_gen == ino1_gen ? 1 : 0;
3548		ino = parent;
3549	}
3550	return 0;
3551}
3552
3553static int wait_for_parent_move(struct send_ctx *sctx,
3554				struct recorded_ref *parent_ref,
3555				const bool is_orphan)
3556{
3557	int ret = 0;
3558	u64 ino = parent_ref->dir;
3559	u64 parent_ino_before, parent_ino_after;
3560	struct fs_path *path_before = NULL;
3561	struct fs_path *path_after = NULL;
3562	int len1, len2;
3563
3564	path_after = fs_path_alloc();
3565	path_before = fs_path_alloc();
3566	if (!path_after || !path_before) {
3567		ret = -ENOMEM;
3568		goto out;
3569	}
3570
3571	/*
3572	 * Our current directory inode may not yet be renamed/moved because some
3573	 * ancestor (immediate or not) has to be renamed/moved first. So find if
3574	 * such ancestor exists and make sure our own rename/move happens after
3575	 * that ancestor is processed to avoid path build infinite loops (done
3576	 * at get_cur_path()).
3577	 */
3578	while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3579		if (is_waiting_for_move(sctx, ino)) {
3580			/*
3581			 * If the current inode is an ancestor of ino in the
3582			 * parent root, we need to delay the rename of the
3583			 * current inode, otherwise don't delayed the rename
3584			 * because we can end up with a circular dependency
3585			 * of renames, resulting in some directories never
3586			 * getting the respective rename operations issued in
3587			 * the send stream or getting into infinite path build
3588			 * loops.
3589			 */
3590			ret = is_ancestor(sctx->parent_root,
3591					  sctx->cur_ino, sctx->cur_inode_gen,
3592					  ino, path_before);
3593			if (ret)
3594				break;
3595		}
3596
3597		fs_path_reset(path_before);
3598		fs_path_reset(path_after);
3599
3600		ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
3601				    NULL, path_after);
3602		if (ret < 0)
3603			goto out;
3604		ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
3605				    NULL, path_before);
3606		if (ret < 0 && ret != -ENOENT) {
3607			goto out;
3608		} else if (ret == -ENOENT) {
3609			ret = 0;
3610			break;
3611		}
3612
3613		len1 = fs_path_len(path_before);
3614		len2 = fs_path_len(path_after);
3615		if (ino > sctx->cur_ino &&
3616		    (parent_ino_before != parent_ino_after || len1 != len2 ||
3617		     memcmp(path_before->start, path_after->start, len1))) {
3618			ret = 1;
3619			break;
3620		}
3621		ino = parent_ino_after;
3622	}
3623
3624out:
3625	fs_path_free(path_before);
3626	fs_path_free(path_after);
3627
3628	if (ret == 1) {
3629		ret = add_pending_dir_move(sctx,
3630					   sctx->cur_ino,
3631					   sctx->cur_inode_gen,
3632					   ino,
3633					   &sctx->new_refs,
3634					   &sctx->deleted_refs,
3635					   is_orphan);
3636		if (!ret)
3637			ret = 1;
3638	}
3639
3640	return ret;
3641}
3642
3643/*
3644 * This does all the move/link/unlink/rmdir magic.
3645 */
3646static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
3647{
3648	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
3649	int ret = 0;
3650	struct recorded_ref *cur;
3651	struct recorded_ref *cur2;
3652	struct list_head check_dirs;
3653	struct fs_path *valid_path = NULL;
3654	u64 ow_inode = 0;
3655	u64 ow_gen;
3656	int did_overwrite = 0;
3657	int is_orphan = 0;
3658	u64 last_dir_ino_rm = 0;
3659	bool can_rename = true;
3660
3661	btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
3662
3663	/*
3664	 * This should never happen as the root dir always has the same ref
3665	 * which is always '..'
3666	 */
3667	BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID);
3668	INIT_LIST_HEAD(&check_dirs);
3669
3670	valid_path = fs_path_alloc();
3671	if (!valid_path) {
3672		ret = -ENOMEM;
3673		goto out;
3674	}
3675
3676	/*
3677	 * First, check if the first ref of the current inode was overwritten
3678	 * before. If yes, we know that the current inode was already orphanized
3679	 * and thus use the orphan name. If not, we can use get_cur_path to
3680	 * get the path of the first ref as it would like while receiving at
3681	 * this point in time.
3682	 * New inodes are always orphan at the beginning, so force to use the
3683	 * orphan name in this case.
3684	 * The first ref is stored in valid_path and will be updated if it
3685	 * gets moved around.
3686	 */
3687	if (!sctx->cur_inode_new) {
3688		ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
3689				sctx->cur_inode_gen);
3690		if (ret < 0)
3691			goto out;
3692		if (ret)
3693			did_overwrite = 1;
3694	}
3695	if (sctx->cur_inode_new || did_overwrite) {
3696		ret = gen_unique_name(sctx, sctx->cur_ino,
3697				sctx->cur_inode_gen, valid_path);
3698		if (ret < 0)
3699			goto out;
3700		is_orphan = 1;
3701	} else {
3702		ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
3703				valid_path);
3704		if (ret < 0)
3705			goto out;
3706	}
3707
3708	list_for_each_entry(cur, &sctx->new_refs, list) {
3709		/*
3710		 * We may have refs where the parent directory does not exist
3711		 * yet. This happens if the parent directories inum is higher
3712		 * the the current inum. To handle this case, we create the
3713		 * parent directory out of order. But we need to check if this
3714		 * did already happen before due to other refs in the same dir.
3715		 */
3716		ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
3717		if (ret < 0)
3718			goto out;
3719		if (ret == inode_state_will_create) {
3720			ret = 0;
3721			/*
3722			 * First check if any of the current inodes refs did
3723			 * already create the dir.
3724			 */
3725			list_for_each_entry(cur2, &sctx->new_refs, list) {
3726				if (cur == cur2)
3727					break;
3728				if (cur2->dir == cur->dir) {
3729					ret = 1;
3730					break;
3731				}
3732			}
3733
3734			/*
3735			 * If that did not happen, check if a previous inode
3736			 * did already create the dir.
3737			 */
3738			if (!ret)
3739				ret = did_create_dir(sctx, cur->dir);
3740			if (ret < 0)
3741				goto out;
3742			if (!ret) {
3743				ret = send_create_inode(sctx, cur->dir);
3744				if (ret < 0)
3745					goto out;
3746			}
3747		}
3748
3749		/*
3750		 * Check if this new ref would overwrite the first ref of
3751		 * another unprocessed inode. If yes, orphanize the
3752		 * overwritten inode. If we find an overwritten ref that is
3753		 * not the first ref, simply unlink it.
3754		 */
3755		ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
3756				cur->name, cur->name_len,
3757				&ow_inode, &ow_gen);
3758		if (ret < 0)
3759			goto out;
3760		if (ret) {
3761			ret = is_first_ref(sctx->parent_root,
3762					   ow_inode, cur->dir, cur->name,
3763					   cur->name_len);
3764			if (ret < 0)
3765				goto out;
3766			if (ret) {
3767				struct name_cache_entry *nce;
3768				struct waiting_dir_move *wdm;
3769
3770				ret = orphanize_inode(sctx, ow_inode, ow_gen,
3771						cur->full_path);
3772				if (ret < 0)
3773					goto out;
3774
3775				/*
3776				 * If ow_inode has its rename operation delayed
3777				 * make sure that its orphanized name is used in
3778				 * the source path when performing its rename
3779				 * operation.
3780				 */
3781				if (is_waiting_for_move(sctx, ow_inode)) {
3782					wdm = get_waiting_dir_move(sctx,
3783								   ow_inode);
3784					ASSERT(wdm);
3785					wdm->orphanized = true;
3786				}
3787
3788				/*
3789				 * Make sure we clear our orphanized inode's
3790				 * name from the name cache. This is because the
3791				 * inode ow_inode might be an ancestor of some
3792				 * other inode that will be orphanized as well
3793				 * later and has an inode number greater than
3794				 * sctx->send_progress. We need to prevent
3795				 * future name lookups from using the old name
3796				 * and get instead the orphan name.
3797				 */
3798				nce = name_cache_search(sctx, ow_inode, ow_gen);
3799				if (nce) {
3800					name_cache_delete(sctx, nce);
3801					kfree(nce);
3802				}
3803
3804				/*
3805				 * ow_inode might currently be an ancestor of
3806				 * cur_ino, therefore compute valid_path (the
3807				 * current path of cur_ino) again because it
3808				 * might contain the pre-orphanization name of
3809				 * ow_inode, which is no longer valid.
3810				 */
3811				fs_path_reset(valid_path);
3812				ret = get_cur_path(sctx, sctx->cur_ino,
3813					   sctx->cur_inode_gen, valid_path);
3814				if (ret < 0)
3815					goto out;
3816			} else {
3817				ret = send_unlink(sctx, cur->full_path);
3818				if (ret < 0)
3819					goto out;
3820			}
3821		}
3822
3823		if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
3824			ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
3825			if (ret < 0)
3826				goto out;
3827			if (ret == 1) {
3828				can_rename = false;
3829				*pending_move = 1;
3830			}
3831		}
3832
3833		if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
3834		    can_rename) {
3835			ret = wait_for_parent_move(sctx, cur, is_orphan);
3836			if (ret < 0)
3837				goto out;
3838			if (ret == 1) {
3839				can_rename = false;
3840				*pending_move = 1;
3841			}
3842		}
3843
3844		/*
3845		 * link/move the ref to the new place. If we have an orphan
3846		 * inode, move it and update valid_path. If not, link or move
3847		 * it depending on the inode mode.
3848		 */
3849		if (is_orphan && can_rename) {
3850			ret = send_rename(sctx, valid_path, cur->full_path);
3851			if (ret < 0)
3852				goto out;
3853			is_orphan = 0;
3854			ret = fs_path_copy(valid_path, cur->full_path);
3855			if (ret < 0)
3856				goto out;
3857		} else if (can_rename) {
3858			if (S_ISDIR(sctx->cur_inode_mode)) {
3859				/*
3860				 * Dirs can't be linked, so move it. For moved
3861				 * dirs, we always have one new and one deleted
3862				 * ref. The deleted ref is ignored later.
3863				 */
3864				ret = send_rename(sctx, valid_path,
3865						  cur->full_path);
3866				if (!ret)
3867					ret = fs_path_copy(valid_path,
3868							   cur->full_path);
3869				if (ret < 0)
3870					goto out;
3871			} else {
3872				ret = send_link(sctx, cur->full_path,
3873						valid_path);
3874				if (ret < 0)
3875					goto out;
3876			}
3877		}
3878		ret = dup_ref(cur, &check_dirs);
3879		if (ret < 0)
3880			goto out;
3881	}
3882
3883	if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
3884		/*
3885		 * Check if we can already rmdir the directory. If not,
3886		 * orphanize it. For every dir item inside that gets deleted
3887		 * later, we do this check again and rmdir it then if possible.
3888		 * See the use of check_dirs for more details.
3889		 */
3890		ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen,
3891				sctx->cur_ino);
3892		if (ret < 0)
3893			goto out;
3894		if (ret) {
3895			ret = send_rmdir(sctx, valid_path);
3896			if (ret < 0)
3897				goto out;
3898		} else if (!is_orphan) {
3899			ret = orphanize_inode(sctx, sctx->cur_ino,
3900					sctx->cur_inode_gen, valid_path);
3901			if (ret < 0)
3902				goto out;
3903			is_orphan = 1;
3904		}
3905
3906		list_for_each_entry(cur, &sctx->deleted_refs, list) {
3907			ret = dup_ref(cur, &check_dirs);
3908			if (ret < 0)
3909				goto out;
3910		}
3911	} else if (S_ISDIR(sctx->cur_inode_mode) &&
3912		   !list_empty(&sctx->deleted_refs)) {
3913		/*
3914		 * We have a moved dir. Add the old parent to check_dirs
3915		 */
3916		cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
3917				list);
3918		ret = dup_ref(cur, &check_dirs);
3919		if (ret < 0)
3920			goto out;
3921	} else if (!S_ISDIR(sctx->cur_inode_mode)) {
3922		/*
3923		 * We have a non dir inode. Go through all deleted refs and
3924		 * unlink them if they were not already overwritten by other
3925		 * inodes.
3926		 */
3927		list_for_each_entry(cur, &sctx->deleted_refs, list) {
3928			ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
3929					sctx->cur_ino, sctx->cur_inode_gen,
3930					cur->name, cur->name_len);
3931			if (ret < 0)
3932				goto out;
3933			if (!ret) {
3934				ret = send_unlink(sctx, cur->full_path);
3935				if (ret < 0)
3936					goto out;
3937			}
3938			ret = dup_ref(cur, &check_dirs);
3939			if (ret < 0)
3940				goto out;
3941		}
3942		/*
3943		 * If the inode is still orphan, unlink the orphan. This may
3944		 * happen when a previous inode did overwrite the first ref
3945		 * of this inode and no new refs were added for the current
3946		 * inode. Unlinking does not mean that the inode is deleted in
3947		 * all cases. There may still be links to this inode in other
3948		 * places.
3949		 */
3950		if (is_orphan) {
3951			ret = send_unlink(sctx, valid_path);
3952			if (ret < 0)
3953				goto out;
3954		}
3955	}
3956
3957	/*
3958	 * We did collect all parent dirs where cur_inode was once located. We
3959	 * now go through all these dirs and check if they are pending for
3960	 * deletion and if it's finally possible to perform the rmdir now.
3961	 * We also update the inode stats of the parent dirs here.
3962	 */
3963	list_for_each_entry(cur, &check_dirs, list) {
3964		/*
3965		 * In case we had refs into dirs that were not processed yet,
3966		 * we don't need to do the utime and rmdir logic for these dirs.
3967		 * The dir will be processed later.
3968		 */
3969		if (cur->dir > sctx->cur_ino)
3970			continue;
3971
3972		ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
3973		if (ret < 0)
3974			goto out;
3975
3976		if (ret == inode_state_did_create ||
3977		    ret == inode_state_no_change) {
3978			/* TODO delayed utimes */
3979			ret = send_utimes(sctx, cur->dir, cur->dir_gen);
3980			if (ret < 0)
3981				goto out;
3982		} else if (ret == inode_state_did_delete &&
3983			   cur->dir != last_dir_ino_rm) {
3984			ret = can_rmdir(sctx, cur->dir, cur->dir_gen,
3985					sctx->cur_ino);
3986			if (ret < 0)
3987				goto out;
3988			if (ret) {
3989				ret = get_cur_path(sctx, cur->dir,
3990						   cur->dir_gen, valid_path);
3991				if (ret < 0)
3992					goto out;
3993				ret = send_rmdir(sctx, valid_path);
3994				if (ret < 0)
3995					goto out;
3996				last_dir_ino_rm = cur->dir;
3997			}
3998		}
3999	}
4000
4001	ret = 0;
4002
4003out:
4004	__free_recorded_refs(&check_dirs);
4005	free_recorded_refs(sctx);
4006	fs_path_free(valid_path);
4007	return ret;
4008}
4009
4010static int record_ref(struct btrfs_root *root, int num, u64 dir, int index,
4011		      struct fs_path *name, void *ctx, struct list_head *refs)
4012{
4013	int ret = 0;
4014	struct send_ctx *sctx = ctx;
4015	struct fs_path *p;
4016	u64 gen;
4017
4018	p = fs_path_alloc();
4019	if (!p)
4020		return -ENOMEM;
4021
4022	ret = get_inode_info(root, dir, NULL, &gen, NULL, NULL,
4023			NULL, NULL);
4024	if (ret < 0)
4025		goto out;
4026
4027	ret = get_cur_path(sctx, dir, gen, p);
4028	if (ret < 0)
4029		goto out;
4030	ret = fs_path_add_path(p, name);
4031	if (ret < 0)
4032		goto out;
4033
4034	ret = __record_ref(refs, dir, gen, p);
4035
4036out:
4037	if (ret)
4038		fs_path_free(p);
4039	return ret;
4040}
4041
4042static int __record_new_ref(int num, u64 dir, int index,
4043			    struct fs_path *name,
4044			    void *ctx)
4045{
4046	struct send_ctx *sctx = ctx;
4047	return record_ref(sctx->send_root, num, dir, index, name,
4048			  ctx, &sctx->new_refs);
4049}
4050
4051
4052static int __record_deleted_ref(int num, u64 dir, int index,
4053				struct fs_path *name,
4054				void *ctx)
4055{
4056	struct send_ctx *sctx = ctx;
4057	return record_ref(sctx->parent_root, num, dir, index, name,
4058			  ctx, &sctx->deleted_refs);
4059}
4060
4061static int record_new_ref(struct send_ctx *sctx)
4062{
4063	int ret;
4064
4065	ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4066				sctx->cmp_key, 0, __record_new_ref, sctx);
4067	if (ret < 0)
4068		goto out;
4069	ret = 0;
4070
4071out:
4072	return ret;
4073}
4074
4075static int record_deleted_ref(struct send_ctx *sctx)
4076{
4077	int ret;
4078
4079	ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4080				sctx->cmp_key, 0, __record_deleted_ref, sctx);
4081	if (ret < 0)
4082		goto out;
4083	ret = 0;
4084
4085out:
4086	return ret;
4087}
4088
4089struct find_ref_ctx {
4090	u64 dir;
4091	u64 dir_gen;
4092	struct btrfs_root *root;
4093	struct fs_path *name;
4094	int found_idx;
4095};
4096
4097static int __find_iref(int num, u64 dir, int index,
4098		       struct fs_path *name,
4099		       void *ctx_)
4100{
4101	struct find_ref_ctx *ctx = ctx_;
4102	u64 dir_gen;
4103	int ret;
4104
4105	if (dir == ctx->dir && fs_path_len(name) == fs_path_len(ctx->name) &&
4106	    strncmp(name->start, ctx->name->start, fs_path_len(name)) == 0) {
4107		/*
4108		 * To avoid doing extra lookups we'll only do this if everything
4109		 * else matches.
4110		 */
4111		ret = get_inode_info(ctx->root, dir, NULL, &dir_gen, NULL,
4112				     NULL, NULL, NULL);
4113		if (ret)
4114			return ret;
4115		if (dir_gen != ctx->dir_gen)
4116			return 0;
4117		ctx->found_idx = num;
4118		return 1;
4119	}
4120	return 0;
4121}
4122
4123static int find_iref(struct btrfs_root *root,
4124		     struct btrfs_path *path,
4125		     struct btrfs_key *key,
4126		     u64 dir, u64 dir_gen, struct fs_path *name)
4127{
4128	int ret;
4129	struct find_ref_ctx ctx;
4130
4131	ctx.dir = dir;
4132	ctx.name = name;
4133	ctx.dir_gen = dir_gen;
4134	ctx.found_idx = -1;
4135	ctx.root = root;
4136
4137	ret = iterate_inode_ref(root, path, key, 0, __find_iref, &ctx);
4138	if (ret < 0)
4139		return ret;
4140
4141	if (ctx.found_idx == -1)
4142		return -ENOENT;
4143
4144	return ctx.found_idx;
4145}
4146
4147static int __record_changed_new_ref(int num, u64 dir, int index,
4148				    struct fs_path *name,
4149				    void *ctx)
4150{
4151	u64 dir_gen;
4152	int ret;
4153	struct send_ctx *sctx = ctx;
4154
4155	ret = get_inode_info(sctx->send_root, dir, NULL, &dir_gen, NULL,
4156			     NULL, NULL, NULL);
4157	if (ret)
4158		return ret;
4159
4160	ret = find_iref(sctx->parent_root, sctx->right_path,
4161			sctx->cmp_key, dir, dir_gen, name);
4162	if (ret == -ENOENT)
4163		ret = __record_new_ref(num, dir, index, name, sctx);
4164	else if (ret > 0)
4165		ret = 0;
4166
4167	return ret;
4168}
4169
4170static int __record_changed_deleted_ref(int num, u64 dir, int index,
4171					struct fs_path *name,
4172					void *ctx)
4173{
4174	u64 dir_gen;
4175	int ret;
4176	struct send_ctx *sctx = ctx;
4177
4178	ret = get_inode_info(sctx->parent_root, dir, NULL, &dir_gen, NULL,
4179			     NULL, NULL, NULL);
4180	if (ret)
4181		return ret;
4182
4183	ret = find_iref(sctx->send_root, sctx->left_path, sctx->cmp_key,
4184			dir, dir_gen, name);
4185	if (ret == -ENOENT)
4186		ret = __record_deleted_ref(num, dir, index, name, sctx);
4187	else if (ret > 0)
4188		ret = 0;
4189
4190	return ret;
4191}
4192
4193static int record_changed_ref(struct send_ctx *sctx)
4194{
4195	int ret = 0;
4196
4197	ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4198			sctx->cmp_key, 0, __record_changed_new_ref, sctx);
4199	if (ret < 0)
4200		goto out;
4201	ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4202			sctx->cmp_key, 0, __record_changed_deleted_ref, sctx);
4203	if (ret < 0)
4204		goto out;
4205	ret = 0;
4206
4207out:
4208	return ret;
4209}
4210
4211/*
4212 * Record and process all refs at once. Needed when an inode changes the
4213 * generation number, which means that it was deleted and recreated.
4214 */
4215static int process_all_refs(struct send_ctx *sctx,
4216			    enum btrfs_compare_tree_result cmd)
4217{
4218	int ret;
4219	struct btrfs_root *root;
4220	struct btrfs_path *path;
4221	struct btrfs_key key;
4222	struct btrfs_key found_key;
4223	struct extent_buffer *eb;
4224	int slot;
4225	iterate_inode_ref_t cb;
4226	int pending_move = 0;
4227
4228	path = alloc_path_for_send();
4229	if (!path)
4230		return -ENOMEM;
4231
4232	if (cmd == BTRFS_COMPARE_TREE_NEW) {
4233		root = sctx->send_root;
4234		cb = __record_new_ref;
4235	} else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4236		root = sctx->parent_root;
4237		cb = __record_deleted_ref;
4238	} else {
4239		btrfs_err(sctx->send_root->fs_info,
4240				"Wrong command %d in process_all_refs", cmd);
4241		ret = -EINVAL;
4242		goto out;
4243	}
4244
4245	key.objectid = sctx->cmp_key->objectid;
4246	key.type = BTRFS_INODE_REF_KEY;
4247	key.offset = 0;
4248	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4249	if (ret < 0)
4250		goto out;
4251
4252	while (1) {
4253		eb = path->nodes[0];
4254		slot = path->slots[0];
4255		if (slot >= btrfs_header_nritems(eb)) {
4256			ret = btrfs_next_leaf(root, path);
4257			if (ret < 0)
4258				goto out;
4259			else if (ret > 0)
4260				break;
4261			continue;
4262		}
4263
4264		btrfs_item_key_to_cpu(eb, &found_key, slot);
4265
4266		if (found_key.objectid != key.objectid ||
4267		    (found_key.type != BTRFS_INODE_REF_KEY &&
4268		     found_key.type != BTRFS_INODE_EXTREF_KEY))
4269			break;
4270
4271		ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4272		if (ret < 0)
4273			goto out;
4274
4275		path->slots[0]++;
4276	}
4277	btrfs_release_path(path);
4278
4279	/*
4280	 * We don't actually care about pending_move as we are simply
4281	 * re-creating this inode and will be rename'ing it into place once we
4282	 * rename the parent directory.
4283	 */
4284	ret = process_recorded_refs(sctx, &pending_move);
4285out:
4286	btrfs_free_path(path);
4287	return ret;
4288}
4289
4290static int send_set_xattr(struct send_ctx *sctx,
4291			  struct fs_path *path,
4292			  const char *name, int name_len,
4293			  const char *data, int data_len)
4294{
4295	int ret = 0;
4296
4297	ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4298	if (ret < 0)
4299		goto out;
4300
4301	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4302	TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4303	TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4304
4305	ret = send_cmd(sctx);
4306
4307tlv_put_failure:
4308out:
4309	return ret;
4310}
4311
4312static int send_remove_xattr(struct send_ctx *sctx,
4313			  struct fs_path *path,
4314			  const char *name, int name_len)
4315{
4316	int ret = 0;
4317
4318	ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4319	if (ret < 0)
4320		goto out;
4321
4322	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4323	TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4324
4325	ret = send_cmd(sctx);
4326
4327tlv_put_failure:
4328out:
4329	return ret;
4330}
4331
4332static int __process_new_xattr(int num, struct btrfs_key *di_key,
4333			       const char *name, int name_len,
4334			       const char *data, int data_len,
4335			       u8 type, void *ctx)
4336{
4337	int ret;
4338	struct send_ctx *sctx = ctx;
4339	struct fs_path *p;
4340	struct posix_acl_xattr_header dummy_acl;
4341
4342	p = fs_path_alloc();
4343	if (!p)
4344		return -ENOMEM;
4345
4346	/*
4347	 * This hack is needed because empty acls are stored as zero byte
4348	 * data in xattrs. Problem with that is, that receiving these zero byte
4349	 * acls will fail later. To fix this, we send a dummy acl list that
4350	 * only contains the version number and no entries.
4351	 */
4352	if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4353	    !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4354		if (data_len == 0) {
4355			dummy_acl.a_version =
4356					cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4357			data = (char *)&dummy_acl;
4358			data_len = sizeof(dummy_acl);
4359		}
4360	}
4361
4362	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4363	if (ret < 0)
4364		goto out;
4365
4366	ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4367
4368out:
4369	fs_path_free(p);
4370	return ret;
4371}
4372
4373static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4374				   const char *name, int name_len,
4375				   const char *data, int data_len,
4376				   u8 type, void *ctx)
4377{
4378	int ret;
4379	struct send_ctx *sctx = ctx;
4380	struct fs_path *p;
4381
4382	p = fs_path_alloc();
4383	if (!p)
4384		return -ENOMEM;
4385
4386	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4387	if (ret < 0)
4388		goto out;
4389
4390	ret = send_remove_xattr(sctx, p, name, name_len);
4391
4392out:
4393	fs_path_free(p);
4394	return ret;
4395}
4396
4397static int process_new_xattr(struct send_ctx *sctx)
4398{
4399	int ret = 0;
4400
4401	ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4402			       sctx->cmp_key, __process_new_xattr, sctx);
4403
4404	return ret;
4405}
4406
4407static int process_deleted_xattr(struct send_ctx *sctx)
4408{
4409	return iterate_dir_item(sctx->parent_root, sctx->right_path,
4410				sctx->cmp_key, __process_deleted_xattr, sctx);
4411}
4412
4413struct find_xattr_ctx {
4414	const char *name;
4415	int name_len;
4416	int found_idx;
4417	char *found_data;
4418	int found_data_len;
4419};
4420
4421static int __find_xattr(int num, struct btrfs_key *di_key,
4422			const char *name, int name_len,
4423			const char *data, int data_len,
4424			u8 type, void *vctx)
4425{
4426	struct find_xattr_ctx *ctx = vctx;
4427
4428	if (name_len == ctx->name_len &&
4429	    strncmp(name, ctx->name, name_len) == 0) {
4430		ctx->found_idx = num;
4431		ctx->found_data_len = data_len;
4432		ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
4433		if (!ctx->found_data)
4434			return -ENOMEM;
4435		return 1;
4436	}
4437	return 0;
4438}
4439
4440static int find_xattr(struct btrfs_root *root,
4441		      struct btrfs_path *path,
4442		      struct btrfs_key *key,
4443		      const char *name, int name_len,
4444		      char **data, int *data_len)
4445{
4446	int ret;
4447	struct find_xattr_ctx ctx;
4448
4449	ctx.name = name;
4450	ctx.name_len = name_len;
4451	ctx.found_idx = -1;
4452	ctx.found_data = NULL;
4453	ctx.found_data_len = 0;
4454
4455	ret = iterate_dir_item(root, path, key, __find_xattr, &ctx);
4456	if (ret < 0)
4457		return ret;
4458
4459	if (ctx.found_idx == -1)
4460		return -ENOENT;
4461	if (data) {
4462		*data = ctx.found_data;
4463		*data_len = ctx.found_data_len;
4464	} else {
4465		kfree(ctx.found_data);
4466	}
4467	return ctx.found_idx;
4468}
4469
4470
4471static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
4472				       const char *name, int name_len,
4473				       const char *data, int data_len,
4474				       u8 type, void *ctx)
4475{
4476	int ret;
4477	struct send_ctx *sctx = ctx;
4478	char *found_data = NULL;
4479	int found_data_len  = 0;
4480
4481	ret = find_xattr(sctx->parent_root, sctx->right_path,
4482			 sctx->cmp_key, name, name_len, &found_data,
4483			 &found_data_len);
4484	if (ret == -ENOENT) {
4485		ret = __process_new_xattr(num, di_key, name, name_len, data,
4486				data_len, type, ctx);
4487	} else if (ret >= 0) {
4488		if (data_len != found_data_len ||
4489		    memcmp(data, found_data, data_len)) {
4490			ret = __process_new_xattr(num, di_key, name, name_len,
4491					data, data_len, type, ctx);
4492		} else {
4493			ret = 0;
4494		}
4495	}
4496
4497	kfree(found_data);
4498	return ret;
4499}
4500
4501static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
4502					   const char *name, int name_len,
4503					   const char *data, int data_len,
4504					   u8 type, void *ctx)
4505{
4506	int ret;
4507	struct send_ctx *sctx = ctx;
4508
4509	ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
4510			 name, name_len, NULL, NULL);
4511	if (ret == -ENOENT)
4512		ret = __process_deleted_xattr(num, di_key, name, name_len, data,
4513				data_len, type, ctx);
4514	else if (ret >= 0)
4515		ret = 0;
4516
4517	return ret;
4518}
4519
4520static int process_changed_xattr(struct send_ctx *sctx)
4521{
4522	int ret = 0;
4523
4524	ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4525			sctx->cmp_key, __process_changed_new_xattr, sctx);
4526	if (ret < 0)
4527		goto out;
4528	ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
4529			sctx->cmp_key, __process_changed_deleted_xattr, sctx);
4530
4531out:
4532	return ret;
4533}
4534
4535static int process_all_new_xattrs(struct send_ctx *sctx)
4536{
4537	int ret;
4538	struct btrfs_root *root;
4539	struct btrfs_path *path;
4540	struct btrfs_key key;
4541	struct btrfs_key found_key;
4542	struct extent_buffer *eb;
4543	int slot;
4544
4545	path = alloc_path_for_send();
4546	if (!path)
4547		return -ENOMEM;
4548
4549	root = sctx->send_root;
4550
4551	key.objectid = sctx->cmp_key->objectid;
4552	key.type = BTRFS_XATTR_ITEM_KEY;
4553	key.offset = 0;
4554	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4555	if (ret < 0)
4556		goto out;
4557
4558	while (1) {
4559		eb = path->nodes[0];
4560		slot = path->slots[0];
4561		if (slot >= btrfs_header_nritems(eb)) {
4562			ret = btrfs_next_leaf(root, path);
4563			if (ret < 0) {
4564				goto out;
4565			} else if (ret > 0) {
4566				ret = 0;
4567				break;
4568			}
4569			continue;
4570		}
4571
4572		btrfs_item_key_to_cpu(eb, &found_key, slot);
4573		if (found_key.objectid != key.objectid ||
4574		    found_key.type != key.type) {
4575			ret = 0;
4576			goto out;
4577		}
4578
4579		ret = iterate_dir_item(root, path, &found_key,
4580				       __process_new_xattr, sctx);
4581		if (ret < 0)
4582			goto out;
4583
4584		path->slots[0]++;
4585	}
4586
4587out:
4588	btrfs_free_path(path);
4589	return ret;
4590}
4591
4592static ssize_t fill_read_buf(struct send_ctx *sctx, u64 offset, u32 len)
4593{
4594	struct btrfs_root *root = sctx->send_root;
4595	struct btrfs_fs_info *fs_info = root->fs_info;
4596	struct inode *inode;
4597	struct page *page;
4598	char *addr;
4599	struct btrfs_key key;
4600	pgoff_t index = offset >> PAGE_SHIFT;
4601	pgoff_t last_index;
4602	unsigned pg_offset = offset & ~PAGE_MASK;
4603	ssize_t ret = 0;
4604
4605	key.objectid = sctx->cur_ino;
4606	key.type = BTRFS_INODE_ITEM_KEY;
4607	key.offset = 0;
4608
4609	inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4610	if (IS_ERR(inode))
4611		return PTR_ERR(inode);
4612
4613	if (offset + len > i_size_read(inode)) {
4614		if (offset > i_size_read(inode))
4615			len = 0;
4616		else
4617			len = offset - i_size_read(inode);
4618	}
4619	if (len == 0)
4620		goto out;
4621
4622	last_index = (offset + len - 1) >> PAGE_SHIFT;
4623
4624	/* initial readahead */
4625	memset(&sctx->ra, 0, sizeof(struct file_ra_state));
4626	file_ra_state_init(&sctx->ra, inode->i_mapping);
4627	btrfs_force_ra(inode->i_mapping, &sctx->ra, NULL, index,
4628		       last_index - index + 1);
4629
4630	while (index <= last_index) {
4631		unsigned cur_len = min_t(unsigned, len,
4632					 PAGE_SIZE - pg_offset);
4633		page = find_or_create_page(inode->i_mapping, index, GFP_KERNEL);
4634		if (!page) {
4635			ret = -ENOMEM;
4636			break;
4637		}
4638
4639		if (!PageUptodate(page)) {
4640			btrfs_readpage(NULL, page);
4641			lock_page(page);
4642			if (!PageUptodate(page)) {
4643				unlock_page(page);
4644				put_page(page);
4645				ret = -EIO;
4646				break;
4647			}
4648		}
4649
4650		addr = kmap(page);
4651		memcpy(sctx->read_buf + ret, addr + pg_offset, cur_len);
4652		kunmap(page);
4653		unlock_page(page);
4654		put_page(page);
4655		index++;
4656		pg_offset = 0;
4657		len -= cur_len;
4658		ret += cur_len;
4659	}
4660out:
4661	iput(inode);
4662	return ret;
4663}
4664
4665/*
4666 * Read some bytes from the current inode/file and send a write command to
4667 * user space.
4668 */
4669static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
4670{
4671	struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
4672	int ret = 0;
4673	struct fs_path *p;
4674	ssize_t num_read = 0;
4675
4676	p = fs_path_alloc();
4677	if (!p)
4678		return -ENOMEM;
4679
4680	btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
4681
4682	num_read = fill_read_buf(sctx, offset, len);
4683	if (num_read <= 0) {
4684		if (num_read < 0)
4685			ret = num_read;
4686		goto out;
4687	}
4688
4689	ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
4690	if (ret < 0)
4691		goto out;
4692
4693	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4694	if (ret < 0)
4695		goto out;
4696
4697	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4698	TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4699	TLV_PUT(sctx, BTRFS_SEND_A_DATA, sctx->read_buf, num_read);
4700
4701	ret = send_cmd(sctx);
4702
4703tlv_put_failure:
4704out:
4705	fs_path_free(p);
4706	if (ret < 0)
4707		return ret;
4708	return num_read;
4709}
4710
4711/*
4712 * Send a clone command to user space.
4713 */
4714static int send_clone(struct send_ctx *sctx,
4715		      u64 offset, u32 len,
4716		      struct clone_root *clone_root)
4717{
4718	int ret = 0;
4719	struct fs_path *p;
4720	u64 gen;
4721
4722	btrfs_debug(sctx->send_root->fs_info,
4723		    "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
4724		    offset, len, clone_root->root->objectid, clone_root->ino,
4725		    clone_root->offset);
4726
4727	p = fs_path_alloc();
4728	if (!p)
4729		return -ENOMEM;
4730
4731	ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
4732	if (ret < 0)
4733		goto out;
4734
4735	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4736	if (ret < 0)
4737		goto out;
4738
4739	TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4740	TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
4741	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4742
4743	if (clone_root->root == sctx->send_root) {
4744		ret = get_inode_info(sctx->send_root, clone_root->ino, NULL,
4745				&gen, NULL, NULL, NULL, NULL);
4746		if (ret < 0)
4747			goto out;
4748		ret = get_cur_path(sctx, clone_root->ino, gen, p);
4749	} else {
4750		ret = get_inode_path(clone_root->root, clone_root->ino, p);
4751	}
4752	if (ret < 0)
4753		goto out;
4754
4755	/*
4756	 * If the parent we're using has a received_uuid set then use that as
4757	 * our clone source as that is what we will look for when doing a
4758	 * receive.
4759	 *
4760	 * This covers the case that we create a snapshot off of a received
4761	 * subvolume and then use that as the parent and try to receive on a
4762	 * different host.
4763	 */
4764	if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
4765		TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4766			     clone_root->root->root_item.received_uuid);
4767	else
4768		TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4769			     clone_root->root->root_item.uuid);
4770	TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
4771		    le64_to_cpu(clone_root->root->root_item.ctransid));
4772	TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
4773	TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
4774			clone_root->offset);
4775
4776	ret = send_cmd(sctx);
4777
4778tlv_put_failure:
4779out:
4780	fs_path_free(p);
4781	return ret;
4782}
4783
4784/*
4785 * Send an update extent command to user space.
4786 */
4787static int send_update_extent(struct send_ctx *sctx,
4788			      u64 offset, u32 len)
4789{
4790	int ret = 0;
4791	struct fs_path *p;
4792
4793	p = fs_path_alloc();
4794	if (!p)
4795		return -ENOMEM;
4796
4797	ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
4798	if (ret < 0)
4799		goto out;
4800
4801	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4802	if (ret < 0)
4803		goto out;
4804
4805	TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4806	TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4807	TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
4808
4809	ret = send_cmd(sctx);
4810
4811tlv_put_failure:
4812out:
4813	fs_path_free(p);
4814	return ret;
4815}
4816
4817static int send_hole(struct send_ctx *sctx, u64 end)
4818{
4819	struct fs_path *p = NULL;
4820	u64 offset = sctx->cur_inode_last_extent;
4821	u64 len;
4822	int ret = 0;
4823
4824	p = fs_path_alloc();
4825	if (!p)
4826		return -ENOMEM;
4827	ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4828	if (ret < 0)
4829		goto tlv_put_failure;
4830	memset(sctx->read_buf, 0, BTRFS_SEND_READ_SIZE);
4831	while (offset < end) {
4832		len = min_t(u64, end - offset, BTRFS_SEND_READ_SIZE);
4833
4834		ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
4835		if (ret < 0)
4836			break;
4837		TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4838		TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4839		TLV_PUT(sctx, BTRFS_SEND_A_DATA, sctx->read_buf, len);
4840		ret = send_cmd(sctx);
4841		if (ret < 0)
4842			break;
4843		offset += len;
4844	}
4845tlv_put_failure:
4846	fs_path_free(p);
4847	return ret;
4848}
4849
4850static int send_extent_data(struct send_ctx *sctx,
4851			    const u64 offset,
4852			    const u64 len)
4853{
4854	u64 sent = 0;
4855
4856	if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
4857		return send_update_extent(sctx, offset, len);
4858
4859	while (sent < len) {
4860		u64 size = len - sent;
4861		int ret;
4862
4863		if (size > BTRFS_SEND_READ_SIZE)
4864			size = BTRFS_SEND_READ_SIZE;
4865		ret = send_write(sctx, offset + sent, size);
4866		if (ret < 0)
4867			return ret;
4868		if (!ret)
4869			break;
4870		sent += ret;
4871	}
4872	return 0;
4873}
4874
4875static int clone_range(struct send_ctx *sctx,
4876		       struct clone_root *clone_root,
4877		       const u64 disk_byte,
4878		       u64 data_offset,
4879		       u64 offset,
4880		       u64 len)
4881{
4882	struct btrfs_path *path;
4883	struct btrfs_key key;
4884	int ret;
4885
4886	path = alloc_path_for_send();
4887	if (!path)
4888		return -ENOMEM;
4889
4890	/*
4891	 * We can't send a clone operation for the entire range if we find
4892	 * extent items in the respective range in the source file that
4893	 * refer to different extents or if we find holes.
4894	 * So check for that and do a mix of clone and regular write/copy
4895	 * operations if needed.
4896	 *
4897	 * Example:
4898	 *
4899	 * mkfs.btrfs -f /dev/sda
4900	 * mount /dev/sda /mnt
4901	 * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
4902	 * cp --reflink=always /mnt/foo /mnt/bar
4903	 * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
4904	 * btrfs subvolume snapshot -r /mnt /mnt/snap
4905	 *
4906	 * If when we send the snapshot and we are processing file bar (which
4907	 * has a higher inode number than foo) we blindly send a clone operation
4908	 * for the [0, 100K[ range from foo to bar, the receiver ends up getting
4909	 * a file bar that matches the content of file foo - iow, doesn't match
4910	 * the content from bar in the original filesystem.
4911	 */
4912	key.objectid = clone_root->ino;
4913	key.type = BTRFS_EXTENT_DATA_KEY;
4914	key.offset = clone_root->offset;
4915	ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
4916	if (ret < 0)
4917		goto out;
4918	if (ret > 0 && path->slots[0] > 0) {
4919		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
4920		if (key.objectid == clone_root->ino &&
4921		    key.type == BTRFS_EXTENT_DATA_KEY)
4922			path->slots[0]--;
4923	}
4924
4925	while (true) {
4926		struct extent_buffer *leaf = path->nodes[0];
4927		int slot = path->slots[0];
4928		struct btrfs_file_extent_item *ei;
4929		u8 type;
4930		u64 ext_len;
4931		u64 clone_len;
4932
4933		if (slot >= btrfs_header_nritems(leaf)) {
4934			ret = btrfs_next_leaf(clone_root->root, path);
4935			if (ret < 0)
4936				goto out;
4937			else if (ret > 0)
4938				break;
4939			continue;
4940		}
4941
4942		btrfs_item_key_to_cpu(leaf, &key, slot);
4943
4944		/*
4945		 * We might have an implicit trailing hole (NO_HOLES feature
4946		 * enabled). We deal with it after leaving this loop.
4947		 */
4948		if (key.objectid != clone_root->ino ||
4949		    key.type != BTRFS_EXTENT_DATA_KEY)
4950			break;
4951
4952		ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4953		type = btrfs_file_extent_type(leaf, ei);
4954		if (type == BTRFS_FILE_EXTENT_INLINE) {
4955			ext_len = btrfs_file_extent_inline_len(leaf, slot, ei);
4956			ext_len = PAGE_ALIGN(ext_len);
4957		} else {
4958			ext_len = btrfs_file_extent_num_bytes(leaf, ei);
4959		}
4960
4961		if (key.offset + ext_len <= clone_root->offset)
4962			goto next;
4963
4964		if (key.offset > clone_root->offset) {
4965			/* Implicit hole, NO_HOLES feature enabled. */
4966			u64 hole_len = key.offset - clone_root->offset;
4967
4968			if (hole_len > len)
4969				hole_len = len;
4970			ret = send_extent_data(sctx, offset, hole_len);
4971			if (ret < 0)
4972				goto out;
4973
4974			len -= hole_len;
4975			if (len == 0)
4976				break;
4977			offset += hole_len;
4978			clone_root->offset += hole_len;
4979			data_offset += hole_len;
4980		}
4981
4982		if (key.offset >= clone_root->offset + len)
4983			break;
4984
4985		clone_len = min_t(u64, ext_len, len);
4986
4987		if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
4988		    btrfs_file_extent_offset(leaf, ei) == data_offset)
4989			ret = send_clone(sctx, offset, clone_len, clone_root);
4990		else
4991			ret = send_extent_data(sctx, offset, clone_len);
4992
4993		if (ret < 0)
4994			goto out;
4995
4996		len -= clone_len;
4997		if (len == 0)
4998			break;
4999		offset += clone_len;
5000		clone_root->offset += clone_len;
5001		data_offset += clone_len;
5002next:
5003		path->slots[0]++;
5004	}
5005
5006	if (len > 0)
5007		ret = send_extent_data(sctx, offset, len);
5008	else
5009		ret = 0;
5010out:
5011	btrfs_free_path(path);
5012	return ret;
5013}
5014
5015static int send_write_or_clone(struct send_ctx *sctx,
5016			       struct btrfs_path *path,
5017			       struct btrfs_key *key,
5018			       struct clone_root *clone_root)
5019{
5020	int ret = 0;
5021	struct btrfs_file_extent_item *ei;
5022	u64 offset = key->offset;
5023	u64 len;
5024	u8 type;
5025	u64 bs = sctx->send_root->fs_info->sb->s_blocksize;
5026
5027	ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5028			struct btrfs_file_extent_item);
5029	type = btrfs_file_extent_type(path->nodes[0], ei);
5030	if (type == BTRFS_FILE_EXTENT_INLINE) {
5031		len = btrfs_file_extent_inline_len(path->nodes[0],
5032						   path->slots[0], ei);
5033		/*
5034		 * it is possible the inline item won't cover the whole page,
5035		 * but there may be items after this page.  Make
5036		 * sure to send the whole thing
5037		 */
5038		len = PAGE_ALIGN(len);
5039	} else {
5040		len = btrfs_file_extent_num_bytes(path->nodes[0], ei);
5041	}
5042
5043	if (offset + len > sctx->cur_inode_size)
5044		len = sctx->cur_inode_size - offset;
5045	if (len == 0) {
5046		ret = 0;
5047		goto out;
5048	}
5049
5050	if (clone_root && IS_ALIGNED(offset + len, bs)) {
5051		u64 disk_byte;
5052		u64 data_offset;
5053
5054		disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
5055		data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
5056		ret = clone_range(sctx, clone_root, disk_byte, data_offset,
5057				  offset, len);
5058	} else {
5059		ret = send_extent_data(sctx, offset, len);
5060	}
5061out:
5062	return ret;
5063}
5064
5065static int is_extent_unchanged(struct send_ctx *sctx,
5066			       struct btrfs_path *left_path,
5067			       struct btrfs_key *ekey)
5068{
5069	int ret = 0;
5070	struct btrfs_key key;
5071	struct btrfs_path *path = NULL;
5072	struct extent_buffer *eb;
5073	int slot;
5074	struct btrfs_key found_key;
5075	struct btrfs_file_extent_item *ei;
5076	u64 left_disknr;
5077	u64 right_disknr;
5078	u64 left_offset;
5079	u64 right_offset;
5080	u64 left_offset_fixed;
5081	u64 left_len;
5082	u64 right_len;
5083	u64 left_gen;
5084	u64 right_gen;
5085	u8 left_type;
5086	u8 right_type;
5087
5088	path = alloc_path_for_send();
5089	if (!path)
5090		return -ENOMEM;
5091
5092	eb = left_path->nodes[0];
5093	slot = left_path->slots[0];
5094	ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5095	left_type = btrfs_file_extent_type(eb, ei);
5096
5097	if (left_type != BTRFS_FILE_EXTENT_REG) {
5098		ret = 0;
5099		goto out;
5100	}
5101	left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5102	left_len = btrfs_file_extent_num_bytes(eb, ei);
5103	left_offset = btrfs_file_extent_offset(eb, ei);
5104	left_gen = btrfs_file_extent_generation(eb, ei);
5105
5106	/*
5107	 * Following comments will refer to these graphics. L is the left
5108	 * extents which we are checking at the moment. 1-8 are the right
5109	 * extents that we iterate.
5110	 *
5111	 *       |-----L-----|
5112	 * |-1-|-2a-|-3-|-4-|-5-|-6-|
5113	 *
5114	 *       |-----L-----|
5115	 * |--1--|-2b-|...(same as above)
5116	 *
5117	 * Alternative situation. Happens on files where extents got split.
5118	 *       |-----L-----|
5119	 * |-----------7-----------|-6-|
5120	 *
5121	 * Alternative situation. Happens on files which got larger.
5122	 *       |-----L-----|
5123	 * |-8-|
5124	 * Nothing follows after 8.
5125	 */
5126
5127	key.objectid = ekey->objectid;
5128	key.type = BTRFS_EXTENT_DATA_KEY;
5129	key.offset = ekey->offset;
5130	ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
5131	if (ret < 0)
5132		goto out;
5133	if (ret) {
5134		ret = 0;
5135		goto out;
5136	}
5137
5138	/*
5139	 * Handle special case where the right side has no extents at all.
5140	 */
5141	eb = path->nodes[0];
5142	slot = path->slots[0];
5143	btrfs_item_key_to_cpu(eb, &found_key, slot);
5144	if (found_key.objectid != key.objectid ||
5145	    found_key.type != key.type) {
5146		/* If we're a hole then just pretend nothing changed */
5147		ret = (left_disknr) ? 0 : 1;
5148		goto out;
5149	}
5150
5151	/*
5152	 * We're now on 2a, 2b or 7.
5153	 */
5154	key = found_key;
5155	while (key.offset < ekey->offset + left_len) {
5156		ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5157		right_type = btrfs_file_extent_type(eb, ei);
5158		if (right_type != BTRFS_FILE_EXTENT_REG) {
5159			ret = 0;
5160			goto out;
5161		}
5162
5163		right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5164		right_len = btrfs_file_extent_num_bytes(eb, ei);
5165		right_offset = btrfs_file_extent_offset(eb, ei);
5166		right_gen = btrfs_file_extent_generation(eb, ei);
5167
5168		/*
5169		 * Are we at extent 8? If yes, we know the extent is changed.
5170		 * This may only happen on the first iteration.
5171		 */
5172		if (found_key.offset + right_len <= ekey->offset) {
5173			/* If we're a hole just pretend nothing changed */
5174			ret = (left_disknr) ? 0 : 1;
5175			goto out;
5176		}
5177
5178		left_offset_fixed = left_offset;
5179		if (key.offset < ekey->offset) {
5180			/* Fix the right offset for 2a and 7. */
5181			right_offset += ekey->offset - key.offset;
5182		} else {
5183			/* Fix the left offset for all behind 2a and 2b */
5184			left_offset_fixed += key.offset - ekey->offset;
5185		}
5186
5187		/*
5188		 * Check if we have the same extent.
5189		 */
5190		if (left_disknr != right_disknr ||
5191		    left_offset_fixed != right_offset ||
5192		    left_gen != right_gen) {
5193			ret = 0;
5194			goto out;
5195		}
5196
5197		/*
5198		 * Go to the next extent.
5199		 */
5200		ret = btrfs_next_item(sctx->parent_root, path);
5201		if (ret < 0)
5202			goto out;
5203		if (!ret) {
5204			eb = path->nodes[0];
5205			slot = path->slots[0];
5206			btrfs_item_key_to_cpu(eb, &found_key, slot);
5207		}
5208		if (ret || found_key.objectid != key.objectid ||
5209		    found_key.type != key.type) {
5210			key.offset += right_len;
5211			break;
5212		}
5213		if (found_key.offset != key.offset + right_len) {
5214			ret = 0;
5215			goto out;
5216		}
5217		key = found_key;
5218	}
5219
5220	/*
5221	 * We're now behind the left extent (treat as unchanged) or at the end
5222	 * of the right side (treat as changed).
5223	 */
5224	if (key.offset >= ekey->offset + left_len)
5225		ret = 1;
5226	else
5227		ret = 0;
5228
5229
5230out:
5231	btrfs_free_path(path);
5232	return ret;
5233}
5234
5235static int get_last_extent(struct send_ctx *sctx, u64 offset)
5236{
5237	struct btrfs_path *path;
5238	struct btrfs_root *root = sctx->send_root;
5239	struct btrfs_file_extent_item *fi;
5240	struct btrfs_key key;
5241	u64 extent_end;
5242	u8 type;
5243	int ret;
5244
5245	path = alloc_path_for_send();
5246	if (!path)
5247		return -ENOMEM;
5248
5249	sctx->cur_inode_last_extent = 0;
5250
5251	key.objectid = sctx->cur_ino;
5252	key.type = BTRFS_EXTENT_DATA_KEY;
5253	key.offset = offset;
5254	ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
5255	if (ret < 0)
5256		goto out;
5257	ret = 0;
5258	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
5259	if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
5260		goto out;
5261
5262	fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
5263			    struct btrfs_file_extent_item);
5264	type = btrfs_file_extent_type(path->nodes[0], fi);
5265	if (type == BTRFS_FILE_EXTENT_INLINE) {
5266		u64 size = btrfs_file_extent_inline_len(path->nodes[0],
5267							path->slots[0], fi);
5268		extent_end = ALIGN(key.offset + size,
5269				   sctx->send_root->fs_info->sectorsize);
5270	} else {
5271		extent_end = key.offset +
5272			btrfs_file_extent_num_bytes(path->nodes[0], fi);
5273	}
5274	sctx->cur_inode_last_extent = extent_end;
5275out:
5276	btrfs_free_path(path);
5277	return ret;
5278}
5279
5280static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
5281			   struct btrfs_key *key)
5282{
5283	struct btrfs_file_extent_item *fi;
5284	u64 extent_end;
5285	u8 type;
5286	int ret = 0;
5287
5288	if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
5289		return 0;
5290
5291	if (sctx->cur_inode_last_extent == (u64)-1) {
5292		ret = get_last_extent(sctx, key->offset - 1);
5293		if (ret)
5294			return ret;
5295	}
5296
5297	fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
5298			    struct btrfs_file_extent_item);
5299	type = btrfs_file_extent_type(path->nodes[0], fi);
5300	if (type == BTRFS_FILE_EXTENT_INLINE) {
5301		u64 size = btrfs_file_extent_inline_len(path->nodes[0],
5302							path->slots[0], fi);
5303		extent_end = ALIGN(key->offset + size,
5304				   sctx->send_root->fs_info->sectorsize);
5305	} else {
5306		extent_end = key->offset +
5307			btrfs_file_extent_num_bytes(path->nodes[0], fi);
5308	}
5309
5310	if (path->slots[0] == 0 &&
5311	    sctx->cur_inode_last_extent < key->offset) {
5312		/*
5313		 * We might have skipped entire leafs that contained only
5314		 * file extent items for our current inode. These leafs have
5315		 * a generation number smaller (older) than the one in the
5316		 * current leaf and the leaf our last extent came from, and
5317		 * are located between these 2 leafs.
5318		 */
5319		ret = get_last_extent(sctx, key->offset - 1);
5320		if (ret)
5321			return ret;
5322	}
5323
5324	if (sctx->cur_inode_last_extent < key->offset)
5325		ret = send_hole(sctx, key->offset);
5326	sctx->cur_inode_last_extent = extent_end;
5327	return ret;
5328}
5329
5330static int process_extent(struct send_ctx *sctx,
5331			  struct btrfs_path *path,
5332			  struct btrfs_key *key)
5333{
5334	struct clone_root *found_clone = NULL;
5335	int ret = 0;
5336
5337	if (S_ISLNK(sctx->cur_inode_mode))
5338		return 0;
5339
5340	if (sctx->parent_root && !sctx->cur_inode_new) {
5341		ret = is_extent_unchanged(sctx, path, key);
5342		if (ret < 0)
5343			goto out;
5344		if (ret) {
5345			ret = 0;
5346			goto out_hole;
5347		}
5348	} else {
5349		struct btrfs_file_extent_item *ei;
5350		u8 type;
5351
5352		ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5353				    struct btrfs_file_extent_item);
5354		type = btrfs_file_extent_type(path->nodes[0], ei);
5355		if (type == BTRFS_FILE_EXTENT_PREALLOC ||
5356		    type == BTRFS_FILE_EXTENT_REG) {
5357			/*
5358			 * The send spec does not have a prealloc command yet,
5359			 * so just leave a hole for prealloc'ed extents until
5360			 * we have enough commands queued up to justify rev'ing
5361			 * the send spec.
5362			 */
5363			if (type == BTRFS_FILE_EXTENT_PREALLOC) {
5364				ret = 0;
5365				goto out;
5366			}
5367
5368			/* Have a hole, just skip it. */
5369			if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
5370				ret = 0;
5371				goto out;
5372			}
5373		}
5374	}
5375
5376	ret = find_extent_clone(sctx, path, key->objectid, key->offset,
5377			sctx->cur_inode_size, &found_clone);
5378	if (ret != -ENOENT && ret < 0)
5379		goto out;
5380
5381	ret = send_write_or_clone(sctx, path, key, found_clone);
5382	if (ret)
5383		goto out;
5384out_hole:
5385	ret = maybe_send_hole(sctx, path, key);
5386out:
5387	return ret;
5388}
5389
5390static int process_all_extents(struct send_ctx *sctx)
5391{
5392	int ret;
5393	struct btrfs_root *root;
5394	struct btrfs_path *path;
5395	struct btrfs_key key;
5396	struct btrfs_key found_key;
5397	struct extent_buffer *eb;
5398	int slot;
5399
5400	root = sctx->send_root;
5401	path = alloc_path_for_send();
5402	if (!path)
5403		return -ENOMEM;
5404
5405	key.objectid = sctx->cmp_key->objectid;
5406	key.type = BTRFS_EXTENT_DATA_KEY;
5407	key.offset = 0;
5408	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5409	if (ret < 0)
5410		goto out;
5411
5412	while (1) {
5413		eb = path->nodes[0];
5414		slot = path->slots[0];
5415
5416		if (slot >= btrfs_header_nritems(eb)) {
5417			ret = btrfs_next_leaf(root, path);
5418			if (ret < 0) {
5419				goto out;
5420			} else if (ret > 0) {
5421				ret = 0;
5422				break;
5423			}
5424			continue;
5425		}
5426
5427		btrfs_item_key_to_cpu(eb, &found_key, slot);
5428
5429		if (found_key.objectid != key.objectid ||
5430		    found_key.type != key.type) {
5431			ret = 0;
5432			goto out;
5433		}
5434
5435		ret = process_extent(sctx, path, &found_key);
5436		if (ret < 0)
5437			goto out;
5438
5439		path->slots[0]++;
5440	}
5441
5442out:
5443	btrfs_free_path(path);
5444	return ret;
5445}
5446
5447static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
5448					   int *pending_move,
5449					   int *refs_processed)
5450{
5451	int ret = 0;
5452
5453	if (sctx->cur_ino == 0)
5454		goto out;
5455	if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
5456	    sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
5457		goto out;
5458	if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
5459		goto out;
5460
5461	ret = process_recorded_refs(sctx, pending_move);
5462	if (ret < 0)
5463		goto out;
5464
5465	*refs_processed = 1;
5466out:
5467	return ret;
5468}
5469
5470static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
5471{
5472	int ret = 0;
5473	u64 left_mode;
5474	u64 left_uid;
5475	u64 left_gid;
5476	u64 right_mode;
5477	u64 right_uid;
5478	u64 right_gid;
5479	int need_chmod = 0;
5480	int need_chown = 0;
5481	int pending_move = 0;
5482	int refs_processed = 0;
5483
5484	ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
5485					      &refs_processed);
5486	if (ret < 0)
5487		goto out;
5488
5489	/*
5490	 * We have processed the refs and thus need to advance send_progress.
5491	 * Now, calls to get_cur_xxx will take the updated refs of the current
5492	 * inode into account.
5493	 *
5494	 * On the other hand, if our current inode is a directory and couldn't
5495	 * be moved/renamed because its parent was renamed/moved too and it has
5496	 * a higher inode number, we can only move/rename our current inode
5497	 * after we moved/renamed its parent. Therefore in this case operate on
5498	 * the old path (pre move/rename) of our current inode, and the
5499	 * move/rename will be performed later.
5500	 */
5501	if (refs_processed && !pending_move)
5502		sctx->send_progress = sctx->cur_ino + 1;
5503
5504	if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
5505		goto out;
5506	if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
5507		goto out;
5508
5509	ret = get_inode_info(sctx->send_root, sctx->cur_ino, NULL, NULL,
5510			&left_mode, &left_uid, &left_gid, NULL);
5511	if (ret < 0)
5512		goto out;
5513
5514	if (!sctx->parent_root || sctx->cur_inode_new) {
5515		need_chown = 1;
5516		if (!S_ISLNK(sctx->cur_inode_mode))
5517			need_chmod = 1;
5518	} else {
5519		ret = get_inode_info(sctx->parent_root, sctx->cur_ino,
5520				NULL, NULL, &right_mode, &right_uid,
5521				&right_gid, NULL);
5522		if (ret < 0)
5523			goto out;
5524
5525		if (left_uid != right_uid || left_gid != right_gid)
5526			need_chown = 1;
5527		if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
5528			need_chmod = 1;
5529	}
5530
5531	if (S_ISREG(sctx->cur_inode_mode)) {
5532		if (need_send_hole(sctx)) {
5533			if (sctx->cur_inode_last_extent == (u64)-1 ||
5534			    sctx->cur_inode_last_extent <
5535			    sctx->cur_inode_size) {
5536				ret = get_last_extent(sctx, (u64)-1);
5537				if (ret)
5538					goto out;
5539			}
5540			if (sctx->cur_inode_last_extent <
5541			    sctx->cur_inode_size) {
5542				ret = send_hole(sctx, sctx->cur_inode_size);
5543				if (ret)
5544					goto out;
5545			}
5546		}
5547		ret = send_truncate(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5548				sctx->cur_inode_size);
5549		if (ret < 0)
5550			goto out;
5551	}
5552
5553	if (need_chown) {
5554		ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5555				left_uid, left_gid);
5556		if (ret < 0)
5557			goto out;
5558	}
5559	if (need_chmod) {
5560		ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5561				left_mode);
5562		if (ret < 0)
5563			goto out;
5564	}
5565
5566	/*
5567	 * If other directory inodes depended on our current directory
5568	 * inode's move/rename, now do their move/rename operations.
5569	 */
5570	if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
5571		ret = apply_children_dir_moves(sctx);
5572		if (ret)
5573			goto out;
5574		/*
5575		 * Need to send that every time, no matter if it actually
5576		 * changed between the two trees as we have done changes to
5577		 * the inode before. If our inode is a directory and it's
5578		 * waiting to be moved/renamed, we will send its utimes when
5579		 * it's moved/renamed, therefore we don't need to do it here.
5580		 */
5581		sctx->send_progress = sctx->cur_ino + 1;
5582		ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
5583		if (ret < 0)
5584			goto out;
5585	}
5586
5587out:
5588	return ret;
5589}
5590
5591static int changed_inode(struct send_ctx *sctx,
5592			 enum btrfs_compare_tree_result result)
5593{
5594	int ret = 0;
5595	struct btrfs_key *key = sctx->cmp_key;
5596	struct btrfs_inode_item *left_ii = NULL;
5597	struct btrfs_inode_item *right_ii = NULL;
5598	u64 left_gen = 0;
5599	u64 right_gen = 0;
5600
5601	sctx->cur_ino = key->objectid;
5602	sctx->cur_inode_new_gen = 0;
5603	sctx->cur_inode_last_extent = (u64)-1;
5604
5605	/*
5606	 * Set send_progress to current inode. This will tell all get_cur_xxx
5607	 * functions that the current inode's refs are not updated yet. Later,
5608	 * when process_recorded_refs is finished, it is set to cur_ino + 1.
5609	 */
5610	sctx->send_progress = sctx->cur_ino;
5611
5612	if (result == BTRFS_COMPARE_TREE_NEW ||
5613	    result == BTRFS_COMPARE_TREE_CHANGED) {
5614		left_ii = btrfs_item_ptr(sctx->left_path->nodes[0],
5615				sctx->left_path->slots[0],
5616				struct btrfs_inode_item);
5617		left_gen = btrfs_inode_generation(sctx->left_path->nodes[0],
5618				left_ii);
5619	} else {
5620		right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
5621				sctx->right_path->slots[0],
5622				struct btrfs_inode_item);
5623		right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
5624				right_ii);
5625	}
5626	if (result == BTRFS_COMPARE_TREE_CHANGED) {
5627		right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
5628				sctx->right_path->slots[0],
5629				struct btrfs_inode_item);
5630
5631		right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
5632				right_ii);
5633
5634		/*
5635		 * The cur_ino = root dir case is special here. We can't treat
5636		 * the inode as deleted+reused because it would generate a
5637		 * stream that tries to delete/mkdir the root dir.
5638		 */
5639		if (left_gen != right_gen &&
5640		    sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
5641			sctx->cur_inode_new_gen = 1;
5642	}
5643
5644	if (result == BTRFS_COMPARE_TREE_NEW) {
5645		sctx->cur_inode_gen = left_gen;
5646		sctx->cur_inode_new = 1;
5647		sctx->cur_inode_deleted = 0;
5648		sctx->cur_inode_size = btrfs_inode_size(
5649				sctx->left_path->nodes[0], left_ii);
5650		sctx->cur_inode_mode = btrfs_inode_mode(
5651				sctx->left_path->nodes[0], left_ii);
5652		sctx->cur_inode_rdev = btrfs_inode_rdev(
5653				sctx->left_path->nodes[0], left_ii);
5654		if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
5655			ret = send_create_inode_if_needed(sctx);
5656	} else if (result == BTRFS_COMPARE_TREE_DELETED) {
5657		sctx->cur_inode_gen = right_gen;
5658		sctx->cur_inode_new = 0;
5659		sctx->cur_inode_deleted = 1;
5660		sctx->cur_inode_size = btrfs_inode_size(
5661				sctx->right_path->nodes[0], right_ii);
5662		sctx->cur_inode_mode = btrfs_inode_mode(
5663				sctx->right_path->nodes[0], right_ii);
5664	} else if (result == BTRFS_COMPARE_TREE_CHANGED) {
5665		/*
5666		 * We need to do some special handling in case the inode was
5667		 * reported as changed with a changed generation number. This
5668		 * means that the original inode was deleted and new inode
5669		 * reused the same inum. So we have to treat the old inode as
5670		 * deleted and the new one as new.
5671		 */
5672		if (sctx->cur_inode_new_gen) {
5673			/*
5674			 * First, process the inode as if it was deleted.
5675			 */
5676			sctx->cur_inode_gen = right_gen;
5677			sctx->cur_inode_new = 0;
5678			sctx->cur_inode_deleted = 1;
5679			sctx->cur_inode_size = btrfs_inode_size(
5680					sctx->right_path->nodes[0], right_ii);
5681			sctx->cur_inode_mode = btrfs_inode_mode(
5682					sctx->right_path->nodes[0], right_ii);
5683			ret = process_all_refs(sctx,
5684					BTRFS_COMPARE_TREE_DELETED);
5685			if (ret < 0)
5686				goto out;
5687
5688			/*
5689			 * Now process the inode as if it was new.
5690			 */
5691			sctx->cur_inode_gen = left_gen;
5692			sctx->cur_inode_new = 1;
5693			sctx->cur_inode_deleted = 0;
5694			sctx->cur_inode_size = btrfs_inode_size(
5695					sctx->left_path->nodes[0], left_ii);
5696			sctx->cur_inode_mode = btrfs_inode_mode(
5697					sctx->left_path->nodes[0], left_ii);
5698			sctx->cur_inode_rdev = btrfs_inode_rdev(
5699					sctx->left_path->nodes[0], left_ii);
5700			ret = send_create_inode_if_needed(sctx);
5701			if (ret < 0)
5702				goto out;
5703
5704			ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW);
5705			if (ret < 0)
5706				goto out;
5707			/*
5708			 * Advance send_progress now as we did not get into
5709			 * process_recorded_refs_if_needed in the new_gen case.
5710			 */
5711			sctx->send_progress = sctx->cur_ino + 1;
5712
5713			/*
5714			 * Now process all extents and xattrs of the inode as if
5715			 * they were all new.
5716			 */
5717			ret = process_all_extents(sctx);
5718			if (ret < 0)
5719				goto out;
5720			ret = process_all_new_xattrs(sctx);
5721			if (ret < 0)
5722				goto out;
5723		} else {
5724			sctx->cur_inode_gen = left_gen;
5725			sctx->cur_inode_new = 0;
5726			sctx->cur_inode_new_gen = 0;
5727			sctx->cur_inode_deleted = 0;
5728			sctx->cur_inode_size = btrfs_inode_size(
5729					sctx->left_path->nodes[0], left_ii);
5730			sctx->cur_inode_mode = btrfs_inode_mode(
5731					sctx->left_path->nodes[0], left_ii);
5732		}
5733	}
5734
5735out:
5736	return ret;
5737}
5738
5739/*
5740 * We have to process new refs before deleted refs, but compare_trees gives us
5741 * the new and deleted refs mixed. To fix this, we record the new/deleted refs
5742 * first and later process them in process_recorded_refs.
5743 * For the cur_inode_new_gen case, we skip recording completely because
5744 * changed_inode did already initiate processing of refs. The reason for this is
5745 * that in this case, compare_tree actually compares the refs of 2 different
5746 * inodes. To fix this, process_all_refs is used in changed_inode to handle all
5747 * refs of the right tree as deleted and all refs of the left tree as new.
5748 */
5749static int changed_ref(struct send_ctx *sctx,
5750		       enum btrfs_compare_tree_result result)
5751{
5752	int ret = 0;
5753
5754	if (sctx->cur_ino != sctx->cmp_key->objectid) {
5755		inconsistent_snapshot_error(sctx, result, "reference");
5756		return -EIO;
5757	}
5758
5759	if (!sctx->cur_inode_new_gen &&
5760	    sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) {
5761		if (result == BTRFS_COMPARE_TREE_NEW)
5762			ret = record_new_ref(sctx);
5763		else if (result == BTRFS_COMPARE_TREE_DELETED)
5764			ret = record_deleted_ref(sctx);
5765		else if (result == BTRFS_COMPARE_TREE_CHANGED)
5766			ret = record_changed_ref(sctx);
5767	}
5768
5769	return ret;
5770}
5771
5772/*
5773 * Process new/deleted/changed xattrs. We skip processing in the
5774 * cur_inode_new_gen case because changed_inode did already initiate processing
5775 * of xattrs. The reason is the same as in changed_ref
5776 */
5777static int changed_xattr(struct send_ctx *sctx,
5778			 enum btrfs_compare_tree_result result)
5779{
5780	int ret = 0;
5781
5782	if (sctx->cur_ino != sctx->cmp_key->objectid) {
5783		inconsistent_snapshot_error(sctx, result, "xattr");
5784		return -EIO;
5785	}
5786
5787	if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
5788		if (result == BTRFS_COMPARE_TREE_NEW)
5789			ret = process_new_xattr(sctx);
5790		else if (result == BTRFS_COMPARE_TREE_DELETED)
5791			ret = process_deleted_xattr(sctx);
5792		else if (result == BTRFS_COMPARE_TREE_CHANGED)
5793			ret = process_changed_xattr(sctx);
5794	}
5795
5796	return ret;
5797}
5798
5799/*
5800 * Process new/deleted/changed extents. We skip processing in the
5801 * cur_inode_new_gen case because changed_inode did already initiate processing
5802 * of extents. The reason is the same as in changed_ref
5803 */
5804static int changed_extent(struct send_ctx *sctx,
5805			  enum btrfs_compare_tree_result result)
5806{
5807	int ret = 0;
5808
5809	if (sctx->cur_ino != sctx->cmp_key->objectid) {
5810
5811		if (result == BTRFS_COMPARE_TREE_CHANGED) {
5812			struct extent_buffer *leaf_l;
5813			struct extent_buffer *leaf_r;
5814			struct btrfs_file_extent_item *ei_l;
5815			struct btrfs_file_extent_item *ei_r;
5816
5817			leaf_l = sctx->left_path->nodes[0];
5818			leaf_r = sctx->right_path->nodes[0];
5819			ei_l = btrfs_item_ptr(leaf_l,
5820					      sctx->left_path->slots[0],
5821					      struct btrfs_file_extent_item);
5822			ei_r = btrfs_item_ptr(leaf_r,
5823					      sctx->right_path->slots[0],
5824					      struct btrfs_file_extent_item);
5825
5826			/*
5827			 * We may have found an extent item that has changed
5828			 * only its disk_bytenr field and the corresponding
5829			 * inode item was not updated. This case happens due to
5830			 * very specific timings during relocation when a leaf
5831			 * that contains file extent items is COWed while
5832			 * relocation is ongoing and its in the stage where it
5833			 * updates data pointers. So when this happens we can
5834			 * safely ignore it since we know it's the same extent,
5835			 * but just at different logical and physical locations
5836			 * (when an extent is fully replaced with a new one, we
5837			 * know the generation number must have changed too,
5838			 * since snapshot creation implies committing the current
5839			 * transaction, and the inode item must have been updated
5840			 * as well).
5841			 * This replacement of the disk_bytenr happens at
5842			 * relocation.c:replace_file_extents() through
5843			 * relocation.c:btrfs_reloc_cow_block().
5844			 */
5845			if (btrfs_file_extent_generation(leaf_l, ei_l) ==
5846			    btrfs_file_extent_generation(leaf_r, ei_r) &&
5847			    btrfs_file_extent_ram_bytes(leaf_l, ei_l) ==
5848			    btrfs_file_extent_ram_bytes(leaf_r, ei_r) &&
5849			    btrfs_file_extent_compression(leaf_l, ei_l) ==
5850			    btrfs_file_extent_compression(leaf_r, ei_r) &&
5851			    btrfs_file_extent_encryption(leaf_l, ei_l) ==
5852			    btrfs_file_extent_encryption(leaf_r, ei_r) &&
5853			    btrfs_file_extent_other_encoding(leaf_l, ei_l) ==
5854			    btrfs_file_extent_other_encoding(leaf_r, ei_r) &&
5855			    btrfs_file_extent_type(leaf_l, ei_l) ==
5856			    btrfs_file_extent_type(leaf_r, ei_r) &&
5857			    btrfs_file_extent_disk_bytenr(leaf_l, ei_l) !=
5858			    btrfs_file_extent_disk_bytenr(leaf_r, ei_r) &&
5859			    btrfs_file_extent_disk_num_bytes(leaf_l, ei_l) ==
5860			    btrfs_file_extent_disk_num_bytes(leaf_r, ei_r) &&
5861			    btrfs_file_extent_offset(leaf_l, ei_l) ==
5862			    btrfs_file_extent_offset(leaf_r, ei_r) &&
5863			    btrfs_file_extent_num_bytes(leaf_l, ei_l) ==
5864			    btrfs_file_extent_num_bytes(leaf_r, ei_r))
5865				return 0;
5866		}
5867
5868		inconsistent_snapshot_error(sctx, result, "extent");
5869		return -EIO;
5870	}
5871
5872	if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
5873		if (result != BTRFS_COMPARE_TREE_DELETED)
5874			ret = process_extent(sctx, sctx->left_path,
5875					sctx->cmp_key);
5876	}
5877
5878	return ret;
5879}
5880
5881static int dir_changed(struct send_ctx *sctx, u64 dir)
5882{
5883	u64 orig_gen, new_gen;
5884	int ret;
5885
5886	ret = get_inode_info(sctx->send_root, dir, NULL, &new_gen, NULL, NULL,
5887			     NULL, NULL);
5888	if (ret)
5889		return ret;
5890
5891	ret = get_inode_info(sctx->parent_root, dir, NULL, &orig_gen, NULL,
5892			     NULL, NULL, NULL);
5893	if (ret)
5894		return ret;
5895
5896	return (orig_gen != new_gen) ? 1 : 0;
5897}
5898
5899static int compare_refs(struct send_ctx *sctx, struct btrfs_path *path,
5900			struct btrfs_key *key)
5901{
5902	struct btrfs_inode_extref *extref;
5903	struct extent_buffer *leaf;
5904	u64 dirid = 0, last_dirid = 0;
5905	unsigned long ptr;
5906	u32 item_size;
5907	u32 cur_offset = 0;
5908	int ref_name_len;
5909	int ret = 0;
5910
5911	/* Easy case, just check this one dirid */
5912	if (key->type == BTRFS_INODE_REF_KEY) {
5913		dirid = key->offset;
5914
5915		ret = dir_changed(sctx, dirid);
5916		goto out;
5917	}
5918
5919	leaf = path->nodes[0];
5920	item_size = btrfs_item_size_nr(leaf, path->slots[0]);
5921	ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
5922	while (cur_offset < item_size) {
5923		extref = (struct btrfs_inode_extref *)(ptr +
5924						       cur_offset);
5925		dirid = btrfs_inode_extref_parent(leaf, extref);
5926		ref_name_len = btrfs_inode_extref_name_len(leaf, extref);
5927		cur_offset += ref_name_len + sizeof(*extref);
5928		if (dirid == last_dirid)
5929			continue;
5930		ret = dir_changed(sctx, dirid);
5931		if (ret)
5932			break;
5933		last_dirid = dirid;
5934	}
5935out:
5936	return ret;
5937}
5938
5939/*
5940 * Updates compare related fields in sctx and simply forwards to the actual
5941 * changed_xxx functions.
5942 */
5943static int changed_cb(struct btrfs_root *left_root,
5944		      struct btrfs_root *right_root,
5945		      struct btrfs_path *left_path,
5946		      struct btrfs_path *right_path,
5947		      struct btrfs_key *key,
5948		      enum btrfs_compare_tree_result result,
5949		      void *ctx)
5950{
5951	int ret = 0;
5952	struct send_ctx *sctx = ctx;
5953
5954	if (result == BTRFS_COMPARE_TREE_SAME) {
5955		if (key->type == BTRFS_INODE_REF_KEY ||
5956		    key->type == BTRFS_INODE_EXTREF_KEY) {
5957			ret = compare_refs(sctx, left_path, key);
5958			if (!ret)
5959				return 0;
5960			if (ret < 0)
5961				return ret;
5962		} else if (key->type == BTRFS_EXTENT_DATA_KEY) {
5963			return maybe_send_hole(sctx, left_path, key);
5964		} else {
5965			return 0;
5966		}
5967		result = BTRFS_COMPARE_TREE_CHANGED;
5968		ret = 0;
5969	}
5970
5971	sctx->left_path = left_path;
5972	sctx->right_path = right_path;
5973	sctx->cmp_key = key;
5974
5975	ret = finish_inode_if_needed(sctx, 0);
5976	if (ret < 0)
5977		goto out;
5978
5979	/* Ignore non-FS objects */
5980	if (key->objectid == BTRFS_FREE_INO_OBJECTID ||
5981	    key->objectid == BTRFS_FREE_SPACE_OBJECTID)
5982		goto out;
5983
5984	if (key->type == BTRFS_INODE_ITEM_KEY)
5985		ret = changed_inode(sctx, result);
5986	else if (key->type == BTRFS_INODE_REF_KEY ||
5987		 key->type == BTRFS_INODE_EXTREF_KEY)
5988		ret = changed_ref(sctx, result);
5989	else if (key->type == BTRFS_XATTR_ITEM_KEY)
5990		ret = changed_xattr(sctx, result);
5991	else if (key->type == BTRFS_EXTENT_DATA_KEY)
5992		ret = changed_extent(sctx, result);
5993
5994out:
5995	return ret;
5996}
5997
5998static int full_send_tree(struct send_ctx *sctx)
5999{
6000	int ret;
6001	struct btrfs_root *send_root = sctx->send_root;
6002	struct btrfs_key key;
6003	struct btrfs_key found_key;
6004	struct btrfs_path *path;
6005	struct extent_buffer *eb;
6006	int slot;
6007
6008	path = alloc_path_for_send();
6009	if (!path)
6010		return -ENOMEM;
6011
6012	key.objectid = BTRFS_FIRST_FREE_OBJECTID;
6013	key.type = BTRFS_INODE_ITEM_KEY;
6014	key.offset = 0;
6015
6016	ret = btrfs_search_slot_for_read(send_root, &key, path, 1, 0);
6017	if (ret < 0)
6018		goto out;
6019	if (ret)
6020		goto out_finish;
6021
6022	while (1) {
6023		eb = path->nodes[0];
6024		slot = path->slots[0];
6025		btrfs_item_key_to_cpu(eb, &found_key, slot);
6026
6027		ret = changed_cb(send_root, NULL, path, NULL,
6028				&found_key, BTRFS_COMPARE_TREE_NEW, sctx);
6029		if (ret < 0)
6030			goto out;
6031
6032		key.objectid = found_key.objectid;
6033		key.type = found_key.type;
6034		key.offset = found_key.offset + 1;
6035
6036		ret = btrfs_next_item(send_root, path);
6037		if (ret < 0)
6038			goto out;
6039		if (ret) {
6040			ret  = 0;
6041			break;
6042		}
6043	}
6044
6045out_finish:
6046	ret = finish_inode_if_needed(sctx, 1);
6047
6048out:
6049	btrfs_free_path(path);
6050	return ret;
6051}
6052
6053static int send_subvol(struct send_ctx *sctx)
6054{
6055	int ret;
6056
6057	if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_STREAM_HEADER)) {
6058		ret = send_header(sctx);
6059		if (ret < 0)
6060			goto out;
6061	}
6062
6063	ret = send_subvol_begin(sctx);
6064	if (ret < 0)
6065		goto out;
6066
6067	if (sctx->parent_root) {
6068		ret = btrfs_compare_trees(sctx->send_root, sctx->parent_root,
6069				changed_cb, sctx);
6070		if (ret < 0)
6071			goto out;
6072		ret = finish_inode_if_needed(sctx, 1);
6073		if (ret < 0)
6074			goto out;
6075	} else {
6076		ret = full_send_tree(sctx);
6077		if (ret < 0)
6078			goto out;
6079	}
6080
6081out:
6082	free_recorded_refs(sctx);
6083	return ret;
6084}
6085
6086/*
6087 * If orphan cleanup did remove any orphans from a root, it means the tree
6088 * was modified and therefore the commit root is not the same as the current
6089 * root anymore. This is a problem, because send uses the commit root and
6090 * therefore can see inode items that don't exist in the current root anymore,
6091 * and for example make calls to btrfs_iget, which will do tree lookups based
6092 * on the current root and not on the commit root. Those lookups will fail,
6093 * returning a -ESTALE error, and making send fail with that error. So make
6094 * sure a send does not see any orphans we have just removed, and that it will
6095 * see the same inodes regardless of whether a transaction commit happened
6096 * before it started (meaning that the commit root will be the same as the
6097 * current root) or not.
6098 */
6099static int ensure_commit_roots_uptodate(struct send_ctx *sctx)
6100{
6101	int i;
6102	struct btrfs_trans_handle *trans = NULL;
6103
6104again:
6105	if (sctx->parent_root &&
6106	    sctx->parent_root->node != sctx->parent_root->commit_root)
6107		goto commit_trans;
6108
6109	for (i = 0; i < sctx->clone_roots_cnt; i++)
6110		if (sctx->clone_roots[i].root->node !=
6111		    sctx->clone_roots[i].root->commit_root)
6112			goto commit_trans;
6113
6114	if (trans)
6115		return btrfs_end_transaction(trans);
6116
6117	return 0;
6118
6119commit_trans:
6120	/* Use any root, all fs roots will get their commit roots updated. */
6121	if (!trans) {
6122		trans = btrfs_join_transaction(sctx->send_root);
6123		if (IS_ERR(trans))
6124			return PTR_ERR(trans);
6125		goto again;
6126	}
6127
6128	return btrfs_commit_transaction(trans);
6129}
6130
6131static void btrfs_root_dec_send_in_progress(struct btrfs_root* root)
6132{
6133	spin_lock(&root->root_item_lock);
6134	root->send_in_progress--;
6135	/*
6136	 * Not much left to do, we don't know why it's unbalanced and
6137	 * can't blindly reset it to 0.
6138	 */
6139	if (root->send_in_progress < 0)
6140		btrfs_err(root->fs_info,
6141			  "send_in_progres unbalanced %d root %llu",
6142			  root->send_in_progress, root->root_key.objectid);
6143	spin_unlock(&root->root_item_lock);
6144}
6145
6146long btrfs_ioctl_send(struct file *mnt_file, void __user *arg_)
6147{
6148	int ret = 0;
6149	struct btrfs_root *send_root = BTRFS_I(file_inode(mnt_file))->root;
6150	struct btrfs_fs_info *fs_info = send_root->fs_info;
6151	struct btrfs_root *clone_root;
6152	struct btrfs_ioctl_send_args *arg = NULL;
6153	struct btrfs_key key;
6154	struct send_ctx *sctx = NULL;
6155	u32 i;
6156	u64 *clone_sources_tmp = NULL;
6157	int clone_sources_to_rollback = 0;
6158	unsigned alloc_size;
6159	int sort_clone_roots = 0;
6160	int index;
6161
6162	if (!capable(CAP_SYS_ADMIN))
6163		return -EPERM;
6164
6165	/*
6166	 * The subvolume must remain read-only during send, protect against
6167	 * making it RW. This also protects against deletion.
6168	 */
6169	spin_lock(&send_root->root_item_lock);
6170	send_root->send_in_progress++;
6171	spin_unlock(&send_root->root_item_lock);
6172
6173	/*
6174	 * This is done when we lookup the root, it should already be complete
6175	 * by the time we get here.
6176	 */
6177	WARN_ON(send_root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE);
6178
6179	/*
6180	 * Userspace tools do the checks and warn the user if it's
6181	 * not RO.
6182	 */
6183	if (!btrfs_root_readonly(send_root)) {
6184		ret = -EPERM;
6185		goto out;
6186	}
6187
6188	arg = memdup_user(arg_, sizeof(*arg));
6189	if (IS_ERR(arg)) {
6190		ret = PTR_ERR(arg);
6191		arg = NULL;
6192		goto out;
6193	}
6194
6195	if (arg->clone_sources_count >
6196	    ULLONG_MAX / sizeof(*arg->clone_sources)) {
6197		ret = -EINVAL;
6198		goto out;
6199	}
6200
6201	if (!access_ok(VERIFY_READ, arg->clone_sources,
6202			sizeof(*arg->clone_sources) *
6203			arg->clone_sources_count)) {
6204		ret = -EFAULT;
6205		goto out;
6206	}
6207
6208	if (arg->flags & ~BTRFS_SEND_FLAG_MASK) {
6209		ret = -EINVAL;
6210		goto out;
6211	}
6212
6213	sctx = kzalloc(sizeof(struct send_ctx), GFP_KERNEL);
6214	if (!sctx) {
6215		ret = -ENOMEM;
6216		goto out;
6217	}
6218
6219	INIT_LIST_HEAD(&sctx->new_refs);
6220	INIT_LIST_HEAD(&sctx->deleted_refs);
6221	INIT_RADIX_TREE(&sctx->name_cache, GFP_KERNEL);
6222	INIT_LIST_HEAD(&sctx->name_cache_list);
6223
6224	sctx->flags = arg->flags;
6225
6226	sctx->send_filp = fget(arg->send_fd);
6227	if (!sctx->send_filp) {
6228		ret = -EBADF;
6229		goto out;
6230	}
6231
6232	sctx->send_root = send_root;
6233	/*
6234	 * Unlikely but possible, if the subvolume is marked for deletion but
6235	 * is slow to remove the directory entry, send can still be started
6236	 */
6237	if (btrfs_root_dead(sctx->send_root)) {
6238		ret = -EPERM;
6239		goto out;
6240	}
6241
6242	sctx->clone_roots_cnt = arg->clone_sources_count;
6243
6244	sctx->send_max_size = BTRFS_SEND_BUF_SIZE;
6245	sctx->send_buf = kmalloc(sctx->send_max_size, GFP_KERNEL | __GFP_NOWARN);
6246	if (!sctx->send_buf) {
6247		sctx->send_buf = vmalloc(sctx->send_max_size);
6248		if (!sctx->send_buf) {
6249			ret = -ENOMEM;
6250			goto out;
6251		}
6252	}
6253
6254	sctx->read_buf = kmalloc(BTRFS_SEND_READ_SIZE, GFP_KERNEL | __GFP_NOWARN);
6255	if (!sctx->read_buf) {
6256		sctx->read_buf = vmalloc(BTRFS_SEND_READ_SIZE);
6257		if (!sctx->read_buf) {
6258			ret = -ENOMEM;
6259			goto out;
6260		}
6261	}
6262
6263	sctx->pending_dir_moves = RB_ROOT;
6264	sctx->waiting_dir_moves = RB_ROOT;
6265	sctx->orphan_dirs = RB_ROOT;
6266
6267	alloc_size = sizeof(struct clone_root) * (arg->clone_sources_count + 1);
6268
6269	sctx->clone_roots = kzalloc(alloc_size, GFP_KERNEL | __GFP_NOWARN);
6270	if (!sctx->clone_roots) {
6271		sctx->clone_roots = vzalloc(alloc_size);
6272		if (!sctx->clone_roots) {
6273			ret = -ENOMEM;
6274			goto out;
6275		}
6276	}
6277
6278	alloc_size = arg->clone_sources_count * sizeof(*arg->clone_sources);
6279
6280	if (arg->clone_sources_count) {
6281		clone_sources_tmp = kmalloc(alloc_size, GFP_KERNEL | __GFP_NOWARN);
6282		if (!clone_sources_tmp) {
6283			clone_sources_tmp = vmalloc(alloc_size);
6284			if (!clone_sources_tmp) {
6285				ret = -ENOMEM;
6286				goto out;
6287			}
6288		}
6289
6290		ret = copy_from_user(clone_sources_tmp, arg->clone_sources,
6291				alloc_size);
6292		if (ret) {
6293			ret = -EFAULT;
6294			goto out;
6295		}
6296
6297		for (i = 0; i < arg->clone_sources_count; i++) {
6298			key.objectid = clone_sources_tmp[i];
6299			key.type = BTRFS_ROOT_ITEM_KEY;
6300			key.offset = (u64)-1;
6301
6302			index = srcu_read_lock(&fs_info->subvol_srcu);
6303
6304			clone_root = btrfs_read_fs_root_no_name(fs_info, &key);
6305			if (IS_ERR(clone_root)) {
6306				srcu_read_unlock(&fs_info->subvol_srcu, index);
6307				ret = PTR_ERR(clone_root);
6308				goto out;
6309			}
6310			spin_lock(&clone_root->root_item_lock);
6311			if (!btrfs_root_readonly(clone_root) ||
6312			    btrfs_root_dead(clone_root)) {
6313				spin_unlock(&clone_root->root_item_lock);
6314				srcu_read_unlock(&fs_info->subvol_srcu, index);
6315				ret = -EPERM;
6316				goto out;
6317			}
6318			clone_root->send_in_progress++;
6319			spin_unlock(&clone_root->root_item_lock);
6320			srcu_read_unlock(&fs_info->subvol_srcu, index);
6321
6322			sctx->clone_roots[i].root = clone_root;
6323			clone_sources_to_rollback = i + 1;
6324		}
6325		kvfree(clone_sources_tmp);
6326		clone_sources_tmp = NULL;
6327	}
6328
6329	if (arg->parent_root) {
6330		key.objectid = arg->parent_root;
6331		key.type = BTRFS_ROOT_ITEM_KEY;
6332		key.offset = (u64)-1;
6333
6334		index = srcu_read_lock(&fs_info->subvol_srcu);
6335
6336		sctx->parent_root = btrfs_read_fs_root_no_name(fs_info, &key);
6337		if (IS_ERR(sctx->parent_root)) {
6338			srcu_read_unlock(&fs_info->subvol_srcu, index);
6339			ret = PTR_ERR(sctx->parent_root);
6340			goto out;
6341		}
6342
6343		spin_lock(&sctx->parent_root->root_item_lock);
6344		sctx->parent_root->send_in_progress++;
6345		if (!btrfs_root_readonly(sctx->parent_root) ||
6346				btrfs_root_dead(sctx->parent_root)) {
6347			spin_unlock(&sctx->parent_root->root_item_lock);
6348			srcu_read_unlock(&fs_info->subvol_srcu, index);
6349			ret = -EPERM;
6350			goto out;
6351		}
6352		spin_unlock(&sctx->parent_root->root_item_lock);
6353
6354		srcu_read_unlock(&fs_info->subvol_srcu, index);
6355	}
6356
6357	/*
6358	 * Clones from send_root are allowed, but only if the clone source
6359	 * is behind the current send position. This is checked while searching
6360	 * for possible clone sources.
6361	 */
6362	sctx->clone_roots[sctx->clone_roots_cnt++].root = sctx->send_root;
6363
6364	/* We do a bsearch later */
6365	sort(sctx->clone_roots, sctx->clone_roots_cnt,
6366			sizeof(*sctx->clone_roots), __clone_root_cmp_sort,
6367			NULL);
6368	sort_clone_roots = 1;
6369
6370	ret = ensure_commit_roots_uptodate(sctx);
6371	if (ret)
6372		goto out;
6373
6374	current->journal_info = BTRFS_SEND_TRANS_STUB;
6375	ret = send_subvol(sctx);
6376	current->journal_info = NULL;
6377	if (ret < 0)
6378		goto out;
6379
6380	if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_END_CMD)) {
6381		ret = begin_cmd(sctx, BTRFS_SEND_C_END);
6382		if (ret < 0)
6383			goto out;
6384		ret = send_cmd(sctx);
6385		if (ret < 0)
6386			goto out;
6387	}
6388
6389out:
6390	WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->pending_dir_moves));
6391	while (sctx && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)) {
6392		struct rb_node *n;
6393		struct pending_dir_move *pm;
6394
6395		n = rb_first(&sctx->pending_dir_moves);
6396		pm = rb_entry(n, struct pending_dir_move, node);
6397		while (!list_empty(&pm->list)) {
6398			struct pending_dir_move *pm2;
6399
6400			pm2 = list_first_entry(&pm->list,
6401					       struct pending_dir_move, list);
6402			free_pending_move(sctx, pm2);
6403		}
6404		free_pending_move(sctx, pm);
6405	}
6406
6407	WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves));
6408	while (sctx && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) {
6409		struct rb_node *n;
6410		struct waiting_dir_move *dm;
6411
6412		n = rb_first(&sctx->waiting_dir_moves);
6413		dm = rb_entry(n, struct waiting_dir_move, node);
6414		rb_erase(&dm->node, &sctx->waiting_dir_moves);
6415		kfree(dm);
6416	}
6417
6418	WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->orphan_dirs));
6419	while (sctx && !RB_EMPTY_ROOT(&sctx->orphan_dirs)) {
6420		struct rb_node *n;
6421		struct orphan_dir_info *odi;
6422
6423		n = rb_first(&sctx->orphan_dirs);
6424		odi = rb_entry(n, struct orphan_dir_info, node);
6425		free_orphan_dir_info(sctx, odi);
6426	}
6427
6428	if (sort_clone_roots) {
6429		for (i = 0; i < sctx->clone_roots_cnt; i++)
6430			btrfs_root_dec_send_in_progress(
6431					sctx->clone_roots[i].root);
6432	} else {
6433		for (i = 0; sctx && i < clone_sources_to_rollback; i++)
6434			btrfs_root_dec_send_in_progress(
6435					sctx->clone_roots[i].root);
6436
6437		btrfs_root_dec_send_in_progress(send_root);
6438	}
6439	if (sctx && !IS_ERR_OR_NULL(sctx->parent_root))
6440		btrfs_root_dec_send_in_progress(sctx->parent_root);
6441
6442	kfree(arg);
6443	kvfree(clone_sources_tmp);
6444
6445	if (sctx) {
6446		if (sctx->send_filp)
6447			fput(sctx->send_filp);
6448
6449		kvfree(sctx->clone_roots);
6450		kvfree(sctx->send_buf);
6451		kvfree(sctx->read_buf);
6452
6453		name_cache_free(sctx);
6454
6455		kfree(sctx);
6456	}
6457
6458	return ret;
6459}