// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (C) 2012 Alexander Block.  All rights reserved.
 */

#include <linux/bsearch.h>
#include <linux/fs.h>
#include <linux/file.h>
#include <linux/sort.h>
#include <linux/mount.h>
#include <linux/xattr.h>
#include <linux/posix_acl_xattr.h>
#include <linux/radix-tree.h>
#include <linux/vmalloc.h>
#include <linux/string.h>
#include <linux/compat.h>
#include <linux/crc32c.h>

#include "send.h"
#include "backref.h"
#include "locking.h"
#include "disk-io.h"
#include "btrfs_inode.h"
#include "transaction.h"
#include "compression.h"
#include "xattr.h"

/*
 * Maximum number of references an extent can have in order for us to attempt to
 * issue clone operations instead of write operations. This currently exists to
 * avoid hitting limitations of the backreference walking code (taking a lot of
 * time and using too much memory for extents with large number of references).
 */
#define SEND_MAX_EXTENT_REFS    64

/*
 * A fs_path is a helper to dynamically build path names with unknown size.
 * It reallocates the internal buffer on demand.
 * It allows fast adding of path elements on the right side (normal path) and
 * fast adding to the left side (reversed path). A reversed path can also be
 * unreversed if needed.
 */
struct fs_path {
        union {
                struct {
                        char *start;
                        char *end;

                        char *buf;
                        unsigned short buf_len:15;
                        unsigned short reversed:1;
                        char inline_buf[];
                };
                /*
                 * Average path length does not exceed 200 bytes, we'll have
                 * better packing in the slab and higher chance to satisfy
                 * a allocation later during send.
                 */
                char pad[256];
        };
};
#define FS_PATH_INLINE_SIZE \
        (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))


/* reused for each extent */
struct clone_root {
        struct btrfs_root *root;
        u64 ino;
        u64 offset;

        u64 found_refs;
};

#define SEND_CTX_MAX_NAME_CACHE_SIZE 128
#define SEND_CTX_NAME_CACHE_CLEAN_SIZE (SEND_CTX_MAX_NAME_CACHE_SIZE * 2)

struct send_ctx {
        struct file *send_filp;
        loff_t send_off;
        char *send_buf;
        u32 send_size;
        u32 send_max_size;
        u64 total_send_size;
        u64 cmd_send_size[BTRFS_SEND_C_MAX + 1];
        u64 flags;      /* 'flags' member of btrfs_ioctl_send_args is u64 */

        struct btrfs_root *send_root;
        struct btrfs_root *parent_root;
        struct clone_root *clone_roots;
        int clone_roots_cnt;

        /* current state of the compare_tree call */
        struct btrfs_path *left_path;
        struct btrfs_path *right_path;
        struct btrfs_key *cmp_key;

        /*
         * infos of the currently processed inode. In case of deleted inodes,
         * these are the values from the deleted inode.
         */
        u64 cur_ino;
        u64 cur_inode_gen;
        int cur_inode_new;
        int cur_inode_new_gen;
        int cur_inode_deleted;
        u64 cur_inode_size;
        u64 cur_inode_mode;
        u64 cur_inode_rdev;
        u64 cur_inode_last_extent;
        u64 cur_inode_next_write_offset;
        bool ignore_cur_inode;

        u64 send_progress;

        struct list_head new_refs;
        struct list_head deleted_refs;

        struct radix_tree_root name_cache;
        struct list_head name_cache_list;
        int name_cache_size;

        struct file_ra_state ra;

        /*
         * We process inodes by their increasing order, so if before an
         * incremental send we reverse the parent/child relationship of
         * directories such that a directory with a lower inode number was
         * the parent of a directory with a higher inode number, and the one
         * becoming the new parent got renamed too, we can't rename/move the
         * directory with lower inode number when we finish processing it - we
         * must process the directory with higher inode number first, then
         * rename/move it and then rename/move the directory with lower inode
         * number. Example follows.
         *
         * Tree state when the first send was performed:
         *
         * .
         * |-- a                   (ino 257)
         *     |-- b               (ino 258)
         *         |
         *         |
         *         |-- c           (ino 259)
         *         |   |-- d       (ino 260)
         *         |
         *         |-- c2          (ino 261)
         *
         * Tree state when the second (incremental) send is performed:
         *
         * .
         * |-- a                   (ino 257)
         *     |-- b               (ino 258)
         *         |-- c2          (ino 261)
         *             |-- d2      (ino 260)
         *                 |-- cc  (ino 259)
         *
         * The sequence of steps that lead to the second state was:
         *
         * mv /a/b/c/d /a/b/c2/d2
         * mv /a/b/c /a/b/c2/d2/cc
         *
         * "c" has lower inode number, but we can't move it (2nd mv operation)
         * before we move "d", which has higher inode number.
         *
         * So we just memorize which move/rename operations must be performed
         * later when their respective parent is processed and moved/renamed.
         */

        /* Indexed by parent directory inode number. */
        struct rb_root pending_dir_moves;

        /*
         * Reverse index, indexed by the inode number of a directory that
         * is waiting for the move/rename of its immediate parent before its
         * own move/rename can be performed.
         */
        struct rb_root waiting_dir_moves;

        /*
         * A directory that is going to be rm'ed might have a child directory
         * which is in the pending directory moves index above. In this case,
         * the directory can only be removed after the move/rename of its child
         * is performed. Example:
         *
         * Parent snapshot:
         *
         * .                        (ino 256)
         * |-- a/                   (ino 257)
         *     |-- b/               (ino 258)
         *         |-- c/           (ino 259)
         *         |   |-- x/       (ino 260)
         *         |
         *         |-- y/           (ino 261)
         *
         * Send snapshot:
         *
         * .                        (ino 256)
         * |-- a/                   (ino 257)
         *     |-- b/               (ino 258)
         *         |-- YY/          (ino 261)
         *              |-- x/      (ino 260)
         *
         * Sequence of steps that lead to the send snapshot:
         * rm -f /a/b/c/foo.txt
         * mv /a/b/y /a/b/YY
         * mv /a/b/c/x /a/b/YY
         * rmdir /a/b/c
         *
         * When the child is processed, its move/rename is delayed until its
         * parent is processed (as explained above), but all other operations
         * like update utimes, chown, chgrp, etc, are performed and the paths
         * that it uses for those operations must use the orphanized name of
         * its parent (the directory we're going to rm later), so we need to
         * memorize that name.
         *
         * Indexed by the inode number of the directory to be deleted.
         */
        struct rb_root orphan_dirs;
};

struct pending_dir_move {
        struct rb_node node;
        struct list_head list;
        u64 parent_ino;
        u64 ino;
        u64 gen;
        struct list_head update_refs;
};

struct waiting_dir_move {
        struct rb_node node;
        u64 ino;
        /*
         * There might be some directory that could not be removed because it
         * was waiting for this directory inode to be moved first. Therefore
         * after this directory is moved, we can try to rmdir the ino rmdir_ino.
         */
        u64 rmdir_ino;
        u64 rmdir_gen;
        bool orphanized;
};

struct orphan_dir_info {
        struct rb_node node;
        u64 ino;
        u64 gen;
        u64 last_dir_index_offset;
};

struct name_cache_entry {
        struct list_head list;
        /*
         * radix_tree has only 32bit entries but we need to handle 64bit inums.
         * We use the lower 32bit of the 64bit inum to store it in the tree. If
         * more then one inum would fall into the same entry, we use radix_list
         * to store the additional entries. radix_list is also used to store
         * entries where two entries have the same inum but different
         * generations.
         */
        struct list_head radix_list;
        u64 ino;
        u64 gen;
        u64 parent_ino;
        u64 parent_gen;
        int ret;
        int need_later_update;
        int name_len;
        char name[];
};

#define ADVANCE                                                 1
#define ADVANCE_ONLY_NEXT                                       -1

enum btrfs_compare_tree_result {
        BTRFS_COMPARE_TREE_NEW,
        BTRFS_COMPARE_TREE_DELETED,
        BTRFS_COMPARE_TREE_CHANGED,
        BTRFS_COMPARE_TREE_SAME,
};

__cold
static void inconsistent_snapshot_error(struct send_ctx *sctx,
                                        enum btrfs_compare_tree_result result,
                                        const char *what)
{
        const char *result_string;

        switch (result) {
        case BTRFS_COMPARE_TREE_NEW:
                result_string = "new";
                break;
        case BTRFS_COMPARE_TREE_DELETED:
                result_string = "deleted";
                break;
        case BTRFS_COMPARE_TREE_CHANGED:
                result_string = "updated";
                break;
        case BTRFS_COMPARE_TREE_SAME:
                ASSERT(0);
                result_string = "unchanged";
                break;
        default:
                ASSERT(0);
                result_string = "unexpected";
        }

        btrfs_err(sctx->send_root->fs_info,
                  "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
                  result_string, what, sctx->cmp_key->objectid,
                  sctx->send_root->root_key.objectid,
                  (sctx->parent_root ?
                   sctx->parent_root->root_key.objectid : 0));
}

static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);

static struct waiting_dir_move *
get_waiting_dir_move(struct send_ctx *sctx, u64 ino);

static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen);

static int need_send_hole(struct send_ctx *sctx)
{
        return (sctx->parent_root && !sctx->cur_inode_new &&
                !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
                S_ISREG(sctx->cur_inode_mode));
}

static void fs_path_reset(struct fs_path *p)
{
        if (p->reversed) {
                p->start = p->buf + p->buf_len - 1;
                p->end = p->start;
                *p->start = 0;
        } else {
                p->start = p->buf;
                p->end = p->start;
                *p->start = 0;
        }
}

static struct fs_path *fs_path_alloc(void)
{
        struct fs_path *p;

        p = kmalloc(sizeof(*p), GFP_KERNEL);
        if (!p)
                return NULL;
        p->reversed = 0;
        p->buf = p->inline_buf;
        p->buf_len = FS_PATH_INLINE_SIZE;
        fs_path_reset(p);
        return p;
}

static struct fs_path *fs_path_alloc_reversed(void)
{
        struct fs_path *p;

        p = fs_path_alloc();
        if (!p)
                return NULL;
        p->reversed = 1;
        fs_path_reset(p);
        return p;
}

static void fs_path_free(struct fs_path *p)
{
        if (!p)
                return;
        if (p->buf != p->inline_buf)
                kfree(p->buf);
        kfree(p);
}

static int fs_path_len(struct fs_path *p)
{
        return p->end - p->start;
}

static int fs_path_ensure_buf(struct fs_path *p, int len)
{
        char *tmp_buf;
        int path_len;
        int old_buf_len;

        len++;

        if (p->buf_len >= len)
                return 0;

        if (len > PATH_MAX) {
                WARN_ON(1);
                return -ENOMEM;
        }

        path_len = p->end - p->start;
        old_buf_len = p->buf_len;

        /*
         * First time the inline_buf does not suffice
         */
        if (p->buf == p->inline_buf) {
                tmp_buf = kmalloc(len, GFP_KERNEL);
                if (tmp_buf)
                        memcpy(tmp_buf, p->buf, old_buf_len);
        } else {
                tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
        }
        if (!tmp_buf)
                return -ENOMEM;
        p->buf = tmp_buf;
        /*
         * The real size of the buffer is bigger, this will let the fast path
         * happen most of the time
         */
        p->buf_len = ksize(p->buf);

        if (p->reversed) {
                tmp_buf = p->buf + old_buf_len - path_len - 1;
                p->end = p->buf + p->buf_len - 1;
                p->start = p->end - path_len;
                memmove(p->start, tmp_buf, path_len + 1);
        } else {
                p->start = p->buf;
                p->end = p->start + path_len;
        }
        return 0;
}

static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
                                   char **prepared)
{
        int ret;
        int new_len;

        new_len = p->end - p->start + name_len;
        if (p->start != p->end)
                new_len++;
        ret = fs_path_ensure_buf(p, new_len);
        if (ret < 0)
                goto out;

        if (p->reversed) {
                if (p->start != p->end)
                        *--p->start = '/';
                p->start -= name_len;
                *prepared = p->start;
        } else {
                if (p->start != p->end)
                        *p->end++ = '/';
                *prepared = p->end;
                p->end += name_len;
                *p->end = 0;
        }

out:
        return ret;
}

static int fs_path_add(struct fs_path *p, const char *name, int name_len)
{
        int ret;
        char *prepared;

        ret = fs_path_prepare_for_add(p, name_len, &prepared);
        if (ret < 0)
                goto out;
        memcpy(prepared, name, name_len);

out:
        return ret;
}

static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
{
        int ret;
        char *prepared;

        ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
        if (ret < 0)
                goto out;
        memcpy(prepared, p2->start, p2->end - p2->start);

out:
        return ret;
}

static int fs_path_add_from_extent_buffer(struct fs_path *p,
                                          struct extent_buffer *eb,
                                          unsigned long off, int len)
{
        int ret;
        char *prepared;

        ret = fs_path_prepare_for_add(p, len, &prepared);
        if (ret < 0)
                goto out;

        read_extent_buffer(eb, prepared, off, len);

out:
        return ret;
}

static int fs_path_copy(struct fs_path *p, struct fs_path *from)
{
        int ret;

        p->reversed = from->reversed;
        fs_path_reset(p);

        ret = fs_path_add_path(p, from);

        return ret;
}


static void fs_path_unreverse(struct fs_path *p)
{
        char *tmp;
        int len;

        if (!p->reversed)
                return;

        tmp = p->start;
        len = p->end - p->start;
        p->start = p->buf;
        p->end = p->start + len;
        memmove(p->start, tmp, len + 1);
        p->reversed = 0;
}

static struct btrfs_path *alloc_path_for_send(void)
{
        struct btrfs_path *path;

        path = btrfs_alloc_path();
        if (!path)
                return NULL;
        path->search_commit_root = 1;
        path->skip_locking = 1;
        path->need_commit_sem = 1;
        return path;
}

static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
{
        int ret;
        u32 pos = 0;

        while (pos < len) {
                ret = kernel_write(filp, buf + pos, len - pos, off);
                /* TODO handle that correctly */
                /*if (ret == -ERESTARTSYS) {
                        continue;
                }*/
                if (ret < 0)
                        return ret;
                if (ret == 0) {
                        return -EIO;
                }
                pos += ret;
        }

        return 0;
}

static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
{
        struct btrfs_tlv_header *hdr;
        int total_len = sizeof(*hdr) + len;
        int left = sctx->send_max_size - sctx->send_size;

        if (unlikely(left < total_len))
                return -EOVERFLOW;

        hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
        put_unaligned_le16(attr, &hdr->tlv_type);
        put_unaligned_le16(len, &hdr->tlv_len);
        memcpy(hdr + 1, data, len);
        sctx->send_size += total_len;

        return 0;
}

#define TLV_PUT_DEFINE_INT(bits) \
        static int tlv_put_u##bits(struct send_ctx *sctx,               \
                        u##bits attr, u##bits value)                    \
        {                                                               \
                __le##bits __tmp = cpu_to_le##bits(value);              \
                return tlv_put(sctx, attr, &__tmp, sizeof(__tmp));      \
        }

TLV_PUT_DEFINE_INT(64)

static int tlv_put_string(struct send_ctx *sctx, u16 attr,
                          const char *str, int len)
{
        if (len == -1)
                len = strlen(str);
        return tlv_put(sctx, attr, str, len);
}

static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
                        const u8 *uuid)
{
        return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
}

static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
                                  struct extent_buffer *eb,
                                  struct btrfs_timespec *ts)
{
        struct btrfs_timespec bts;
        read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
        return tlv_put(sctx, attr, &bts, sizeof(bts));
}


#define TLV_PUT(sctx, attrtype, data, attrlen) \
        do { \
                ret = tlv_put(sctx, attrtype, data, attrlen); \
                if (ret < 0) \
                        goto tlv_put_failure; \
        } while (0)

#define TLV_PUT_INT(sctx, attrtype, bits, value) \
        do { \
                ret = tlv_put_u##bits(sctx, attrtype, value); \
                if (ret < 0) \
                        goto tlv_put_failure; \
        } while (0)

#define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
#define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
#define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
#define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
#define TLV_PUT_STRING(sctx, attrtype, str, len) \
        do { \
                ret = tlv_put_string(sctx, attrtype, str, len); \
                if (ret < 0) \
                        goto tlv_put_failure; \
        } while (0)
#define TLV_PUT_PATH(sctx, attrtype, p) \
        do { \
                ret = tlv_put_string(sctx, attrtype, p->start, \
                        p->end - p->start); \
                if (ret < 0) \
                        goto tlv_put_failure; \
        } while(0)
#define TLV_PUT_UUID(sctx, attrtype, uuid) \
        do { \
                ret = tlv_put_uuid(sctx, attrtype, uuid); \
                if (ret < 0) \
                        goto tlv_put_failure; \
        } while (0)
#define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
        do { \
                ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
                if (ret < 0) \
                        goto tlv_put_failure; \
        } while (0)

static int send_header(struct send_ctx *sctx)
{
        struct btrfs_stream_header hdr;

        strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
        hdr.version = cpu_to_le32(BTRFS_SEND_STREAM_VERSION);

        return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
                                        &sctx->send_off);
}

/*
 * For each command/item we want to send to userspace, we call this function.
 */
static int begin_cmd(struct send_ctx *sctx, int cmd)
{
        struct btrfs_cmd_header *hdr;

        if (WARN_ON(!sctx->send_buf))
                return -EINVAL;

        BUG_ON(sctx->send_size);

        sctx->send_size += sizeof(*hdr);
        hdr = (struct btrfs_cmd_header *)sctx->send_buf;
        put_unaligned_le16(cmd, &hdr->cmd);

        return 0;
}

static int send_cmd(struct send_ctx *sctx)
{
        int ret;
        struct btrfs_cmd_header *hdr;
        u32 crc;

        hdr = (struct btrfs_cmd_header *)sctx->send_buf;
        put_unaligned_le32(sctx->send_size - sizeof(*hdr), &hdr->len);
        put_unaligned_le32(0, &hdr->crc);

        crc = btrfs_crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
        put_unaligned_le32(crc, &hdr->crc);

        ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
                                        &sctx->send_off);

        sctx->total_send_size += sctx->send_size;
        sctx->cmd_send_size[get_unaligned_le16(&hdr->cmd)] += sctx->send_size;
        sctx->send_size = 0;

        return ret;
}

/*
 * Sends a move instruction to user space
 */
static int send_rename(struct send_ctx *sctx,
                     struct fs_path *from, struct fs_path *to)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret;

        btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);

        ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
        if (ret < 0)
                goto out;

        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);

        ret = send_cmd(sctx);

tlv_put_failure:
out:
        return ret;
}

/*
 * Sends a link instruction to user space
 */
static int send_link(struct send_ctx *sctx,
                     struct fs_path *path, struct fs_path *lnk)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret;

        btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);

        ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
        if (ret < 0)
                goto out;

        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);

        ret = send_cmd(sctx);

tlv_put_failure:
out:
        return ret;
}

/*
 * Sends an unlink instruction to user space
 */
static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret;

        btrfs_debug(fs_info, "send_unlink %s", path->start);

        ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
        if (ret < 0)
                goto out;

        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);

        ret = send_cmd(sctx);

tlv_put_failure:
out:
        return ret;
}

/*
 * Sends a rmdir instruction to user space
 */
static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret;

        btrfs_debug(fs_info, "send_rmdir %s", path->start);

        ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
        if (ret < 0)
                goto out;

        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);

        ret = send_cmd(sctx);

tlv_put_failure:
out:
        return ret;
}

/*
 * Helper function to retrieve some fields from an inode item.
 */
static int __get_inode_info(struct btrfs_root *root, struct btrfs_path *path,
                          u64 ino, u64 *size, u64 *gen, u64 *mode, u64 *uid,
                          u64 *gid, u64 *rdev)
{
        int ret;
        struct btrfs_inode_item *ii;
        struct btrfs_key key;

        key.objectid = ino;
        key.type = BTRFS_INODE_ITEM_KEY;
        key.offset = 0;
        ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
        if (ret) {
                if (ret > 0)
                        ret = -ENOENT;
                return ret;
        }

        ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
                        struct btrfs_inode_item);
        if (size)
                *size = btrfs_inode_size(path->nodes[0], ii);
        if (gen)
                *gen = btrfs_inode_generation(path->nodes[0], ii);
        if (mode)
                *mode = btrfs_inode_mode(path->nodes[0], ii);
        if (uid)
                *uid = btrfs_inode_uid(path->nodes[0], ii);
        if (gid)
                *gid = btrfs_inode_gid(path->nodes[0], ii);
        if (rdev)
                *rdev = btrfs_inode_rdev(path->nodes[0], ii);

        return ret;
}

static int get_inode_info(struct btrfs_root *root,
                          u64 ino, u64 *size, u64 *gen,
                          u64 *mode, u64 *uid, u64 *gid,
                          u64 *rdev)
{
        struct btrfs_path *path;
        int ret;

        path = alloc_path_for_send();
        if (!path)
                return -ENOMEM;
        ret = __get_inode_info(root, path, ino, size, gen, mode, uid, gid,
                               rdev);
        btrfs_free_path(path);
        return ret;
}

typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
                                   struct fs_path *p,
                                   void *ctx);

/*
 * Helper function to iterate the entries in ONE btrfs_inode_ref or
 * btrfs_inode_extref.
 * The iterate callback may return a non zero value to stop iteration. This can
 * be a negative value for error codes or 1 to simply stop it.
 *
 * path must point to the INODE_REF or INODE_EXTREF when called.
 */
static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
                             struct btrfs_key *found_key, int resolve,
                             iterate_inode_ref_t iterate, void *ctx)
{
        struct extent_buffer *eb = path->nodes[0];
        struct btrfs_item *item;
        struct btrfs_inode_ref *iref;
        struct btrfs_inode_extref *extref;
        struct btrfs_path *tmp_path;
        struct fs_path *p;
        u32 cur = 0;
        u32 total;
        int slot = path->slots[0];
        u32 name_len;
        char *start;
        int ret = 0;
        int num = 0;
        int index;
        u64 dir;
        unsigned long name_off;
        unsigned long elem_size;
        unsigned long ptr;

        p = fs_path_alloc_reversed();
        if (!p)
                return -ENOMEM;

        tmp_path = alloc_path_for_send();
        if (!tmp_path) {
                fs_path_free(p);
                return -ENOMEM;
        }


        if (found_key->type == BTRFS_INODE_REF_KEY) {
                ptr = (unsigned long)btrfs_item_ptr(eb, slot,
                                                    struct btrfs_inode_ref);
                item = btrfs_item_nr(slot);
                total = btrfs_item_size(eb, item);
                elem_size = sizeof(*iref);
        } else {
                ptr = btrfs_item_ptr_offset(eb, slot);
                total = btrfs_item_size_nr(eb, slot);
                elem_size = sizeof(*extref);
        }

        while (cur < total) {
                fs_path_reset(p);

                if (found_key->type == BTRFS_INODE_REF_KEY) {
                        iref = (struct btrfs_inode_ref *)(ptr + cur);
                        name_len = btrfs_inode_ref_name_len(eb, iref);
                        name_off = (unsigned long)(iref + 1);
                        index = btrfs_inode_ref_index(eb, iref);
                        dir = found_key->offset;
                } else {
                        extref = (struct btrfs_inode_extref *)(ptr + cur);
                        name_len = btrfs_inode_extref_name_len(eb, extref);
                        name_off = (unsigned long)&extref->name;
                        index = btrfs_inode_extref_index(eb, extref);
                        dir = btrfs_inode_extref_parent(eb, extref);
                }

                if (resolve) {
                        start = btrfs_ref_to_path(root, tmp_path, name_len,
                                                  name_off, eb, dir,
                                                  p->buf, p->buf_len);
                        if (IS_ERR(start)) {
                                ret = PTR_ERR(start);
                                goto out;
                        }
                        if (start < p->buf) {
                                /* overflow , try again with larger buffer */
                                ret = fs_path_ensure_buf(p,
                                                p->buf_len + p->buf - start);
                                if (ret < 0)
                                        goto out;
                                start = btrfs_ref_to_path(root, tmp_path,
                                                          name_len, name_off,
                                                          eb, dir,
                                                          p->buf, p->buf_len);
                                if (IS_ERR(start)) {
                                        ret = PTR_ERR(start);
                                        goto out;
                                }
                                BUG_ON(start < p->buf);
                        }
                        p->start = start;
                } else {
                        ret = fs_path_add_from_extent_buffer(p, eb, name_off,
                                                             name_len);
                        if (ret < 0)
                                goto out;
                }

                cur += elem_size + name_len;
                ret = iterate(num, dir, index, p, ctx);
                if (ret)
                        goto out;
                num++;
        }

out:
        btrfs_free_path(tmp_path);
        fs_path_free(p);
        return ret;
}

typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
                                  const char *name, int name_len,
                                  const char *data, int data_len,
                                  u8 type, void *ctx);

/*
 * Helper function to iterate the entries in ONE btrfs_dir_item.
 * The iterate callback may return a non zero value to stop iteration. This can
 * be a negative value for error codes or 1 to simply stop it.

 *
 * path must point to the dir item when called.
 */
static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
                            iterate_dir_item_t iterate, void *ctx)
{
        int ret = 0;
        struct extent_buffer *eb;
        struct btrfs_item *item;
        struct btrfs_dir_item *di;
        struct btrfs_key di_key;
        char *buf = NULL;
        int buf_len;
        u32 name_len;
        u32 data_len;
        u32 cur;
        u32 len;
        u32 total;
        int slot;
        int num;
        u8 type;

        /*
         * Start with a small buffer (1 page). If later we end up needing more
         * space, which can happen for xattrs on a fs with a leaf size greater
         * then the page size, attempt to increase the buffer. Typically xattr
         * values are small.
         */
        buf_len = PATH_MAX;
        buf = kmalloc(buf_len, GFP_KERNEL);
        if (!buf) {
                ret = -ENOMEM;
                goto out;
        }

        eb = path->nodes[0];
        slot = path->slots[0];
        item = btrfs_item_nr(slot);
        di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
        cur = 0;
        len = 0;
        total = btrfs_item_size(eb, item);

        num = 0;
        while (cur < total) {
                name_len = btrfs_dir_name_len(eb, di);
                data_len = btrfs_dir_data_len(eb, di);
                type = btrfs_dir_type(eb, di);
                btrfs_dir_item_key_to_cpu(eb, di, &di_key);

                if (type == BTRFS_FT_XATTR) {
                        if (name_len > XATTR_NAME_MAX) {
                                ret = -ENAMETOOLONG;
                                goto out;
                        }
                        if (name_len + data_len >
                                        BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
                                ret = -E2BIG;
                                goto out;
                        }
                } else {
                        /*
                         * Path too long
                         */
                        if (name_len + data_len > PATH_MAX) {
                                ret = -ENAMETOOLONG;
                                goto out;
                        }
                }

                if (name_len + data_len > buf_len) {
                        buf_len = name_len + data_len;
                        if (is_vmalloc_addr(buf)) {
                                vfree(buf);
                                buf = NULL;
                        } else {
                                char *tmp = krealloc(buf, buf_len,
                                                GFP_KERNEL | __GFP_NOWARN);

                                if (!tmp)
                                        kfree(buf);
                                buf = tmp;
                        }
                        if (!buf) {
                                buf = kvmalloc(buf_len, GFP_KERNEL);
                                if (!buf) {
                                        ret = -ENOMEM;
                                        goto out;
                                }
                        }
                }

                read_extent_buffer(eb, buf, (unsigned long)(di + 1),
                                name_len + data_len);

                len = sizeof(*di) + name_len + data_len;
                di = (struct btrfs_dir_item *)((char *)di + len);
                cur += len;

                ret = iterate(num, &di_key, buf, name_len, buf + name_len,
                                data_len, type, ctx);
                if (ret < 0)
                        goto out;
                if (ret) {
                        ret = 0;
                        goto out;
                }

                num++;
        }

out:
        kvfree(buf);
        return ret;
}

static int __copy_first_ref(int num, u64 dir, int index,
                            struct fs_path *p, void *ctx)
{
        int ret;
        struct fs_path *pt = ctx;

        ret = fs_path_copy(pt, p);
        if (ret < 0)
                return ret;

        /* we want the first only */
        return 1;
}

/*
 * Retrieve the first path of an inode. If an inode has more then one
 * ref/hardlink, this is ignored.
 */
static int get_inode_path(struct btrfs_root *root,
                          u64 ino, struct fs_path *path)
{
        int ret;
        struct btrfs_key key, found_key;
        struct btrfs_path *p;

        p = alloc_path_for_send();
        if (!p)
                return -ENOMEM;

        fs_path_reset(path);

        key.objectid = ino;
        key.type = BTRFS_INODE_REF_KEY;
        key.offset = 0;

        ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
        if (ret < 0)
                goto out;
        if (ret) {
                ret = 1;
                goto out;
        }
        btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
        if (found_key.objectid != ino ||
            (found_key.type != BTRFS_INODE_REF_KEY &&
             found_key.type != BTRFS_INODE_EXTREF_KEY)) {
                ret = -ENOENT;
                goto out;
        }

        ret = iterate_inode_ref(root, p, &found_key, 1,
                                __copy_first_ref, path);
        if (ret < 0)
                goto out;
        ret = 0;

out:
        btrfs_free_path(p);
        return ret;
}

struct backref_ctx {
        struct send_ctx *sctx;

        /* number of total found references */
        u64 found;

        /*
         * used for clones found in send_root. clones found behind cur_objectid
         * and cur_offset are not considered as allowed clones.
         */
        u64 cur_objectid;
        u64 cur_offset;

        /* may be truncated in case it's the last extent in a file */
        u64 extent_len;

        /* Just to check for bugs in backref resolving */
        int found_itself;
};

static int __clone_root_cmp_bsearch(const void *key, const void *elt)
{
        u64 root = (u64)(uintptr_t)key;
        struct clone_root *cr = (struct clone_root *)elt;

        if (root < cr->root->root_key.objectid)
                return -1;
        if (root > cr->root->root_key.objectid)
                return 1;
        return 0;
}

static int __clone_root_cmp_sort(const void *e1, const void *e2)
{
        struct clone_root *cr1 = (struct clone_root *)e1;
        struct clone_root *cr2 = (struct clone_root *)e2;

        if (cr1->root->root_key.objectid < cr2->root->root_key.objectid)
                return -1;
        if (cr1->root->root_key.objectid > cr2->root->root_key.objectid)
                return 1;
        return 0;
}

/*
 * Called for every backref that is found for the current extent.
 * Results are collected in sctx->clone_roots->ino/offset/found_refs
 */
static int __iterate_backrefs(u64 ino, u64 offset, u64 root, void *ctx_)
{
        struct backref_ctx *bctx = ctx_;
        struct clone_root *found;

        /* First check if the root is in the list of accepted clone sources */
        found = bsearch((void *)(uintptr_t)root, bctx->sctx->clone_roots,
                        bctx->sctx->clone_roots_cnt,
                        sizeof(struct clone_root),
                        __clone_root_cmp_bsearch);
        if (!found)
                return 0;

        if (found->root == bctx->sctx->send_root &&
            ino == bctx->cur_objectid &&
            offset == bctx->cur_offset) {
                bctx->found_itself = 1;
        }

        /*
         * Make sure we don't consider clones from send_root that are
         * behind the current inode/offset.
         */
        if (found->root == bctx->sctx->send_root) {
                /*
                 * If the source inode was not yet processed we can't issue a
                 * clone operation, as the source extent does not exist yet at
                 * the destination of the stream.
                 */
                if (ino > bctx->cur_objectid)
                        return 0;
                /*
                 * We clone from the inode currently being sent as long as the
                 * source extent is already processed, otherwise we could try
                 * to clone from an extent that does not exist yet at the
                 * destination of the stream.
                 */
                if (ino == bctx->cur_objectid &&
                    offset + bctx->extent_len >
                    bctx->sctx->cur_inode_next_write_offset)
                        return 0;
        }

        bctx->found++;
        found->found_refs++;
        if (ino < found->ino) {
                found->ino = ino;
                found->offset = offset;
        } else if (found->ino == ino) {
                /*
                 * same extent found more then once in the same file.
                 */
                if (found->offset > offset + bctx->extent_len)
                        found->offset = offset;
        }

        return 0;
}

/*
 * Given an inode, offset and extent item, it finds a good clone for a clone
 * instruction. Returns -ENOENT when none could be found. The function makes
 * sure that the returned clone is usable at the point where sending is at the
 * moment. This means, that no clones are accepted which lie behind the current
 * inode+offset.
 *
 * path must point to the extent item when called.
 */
static int find_extent_clone(struct send_ctx *sctx,
                             struct btrfs_path *path,
                             u64 ino, u64 data_offset,
                             u64 ino_size,
                             struct clone_root **found)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret;
        int extent_type;
        u64 logical;
        u64 disk_byte;
        u64 num_bytes;
        u64 extent_item_pos;
        u64 flags = 0;
        struct btrfs_file_extent_item *fi;
        struct extent_buffer *eb = path->nodes[0];
        struct backref_ctx *backref_ctx = NULL;
        struct clone_root *cur_clone_root;
        struct btrfs_key found_key;
        struct btrfs_path *tmp_path;
        struct btrfs_extent_item *ei;
        int compressed;
        u32 i;

        tmp_path = alloc_path_for_send();
        if (!tmp_path)
                return -ENOMEM;

        /* We only use this path under the commit sem */
        tmp_path->need_commit_sem = 0;

        backref_ctx = kmalloc(sizeof(*backref_ctx), GFP_KERNEL);
        if (!backref_ctx) {
                ret = -ENOMEM;
                goto out;
        }

        if (data_offset >= ino_size) {
                /*
                 * There may be extents that lie behind the file's size.
                 * I at least had this in combination with snapshotting while
                 * writing large files.
                 */
                ret = 0;
                goto out;
        }

        fi = btrfs_item_ptr(eb, path->slots[0],
                        struct btrfs_file_extent_item);
        extent_type = btrfs_file_extent_type(eb, fi);
        if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
                ret = -ENOENT;
                goto out;
        }
        compressed = btrfs_file_extent_compression(eb, fi);

        num_bytes = btrfs_file_extent_num_bytes(eb, fi);
        disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
        if (disk_byte == 0) {
                ret = -ENOENT;
                goto out;
        }
        logical = disk_byte + btrfs_file_extent_offset(eb, fi);

        down_read(&fs_info->commit_root_sem);
        ret = extent_from_logical(fs_info, disk_byte, tmp_path,
                                  &found_key, &flags);
        up_read(&fs_info->commit_root_sem);

        if (ret < 0)
                goto out;
        if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
                ret = -EIO;
                goto out;
        }

        ei = btrfs_item_ptr(tmp_path->nodes[0], tmp_path->slots[0],
                            struct btrfs_extent_item);
        /*
         * Backreference walking (iterate_extent_inodes() below) is currently
         * too expensive when an extent has a large number of references, both
         * in time spent and used memory. So for now just fallback to write
         * operations instead of clone operations when an extent has more than
         * a certain amount of references.
         */
        if (btrfs_extent_refs(tmp_path->nodes[0], ei) > SEND_MAX_EXTENT_REFS) {
                ret = -ENOENT;
                goto out;
        }
        btrfs_release_path(tmp_path);

        /*
         * Setup the clone roots.
         */
        for (i = 0; i < sctx->clone_roots_cnt; i++) {
                cur_clone_root = sctx->clone_roots + i;
                cur_clone_root->ino = (u64)-1;
                cur_clone_root->offset = 0;
                cur_clone_root->found_refs = 0;
        }

        backref_ctx->sctx = sctx;
        backref_ctx->found = 0;
        backref_ctx->cur_objectid = ino;
        backref_ctx->cur_offset = data_offset;
        backref_ctx->found_itself = 0;
        backref_ctx->extent_len = num_bytes;

        /*
         * The last extent of a file may be too large due to page alignment.
         * We need to adjust extent_len in this case so that the checks in
         * __iterate_backrefs work.
         */
        if (data_offset + num_bytes >= ino_size)
                backref_ctx->extent_len = ino_size - data_offset;

        /*
         * Now collect all backrefs.
         */
        if (compressed == BTRFS_COMPRESS_NONE)
                extent_item_pos = logical - found_key.objectid;
        else
                extent_item_pos = 0;
        ret = iterate_extent_inodes(fs_info, found_key.objectid,
                                    extent_item_pos, 1, __iterate_backrefs,
                                    backref_ctx, false);

        if (ret < 0)
                goto out;

        if (!backref_ctx->found_itself) {
                /* found a bug in backref code? */
                ret = -EIO;
                btrfs_err(fs_info,
                          "did not find backref in send_root. inode=%llu, offset=%llu, disk_byte=%llu found extent=%llu",
                          ino, data_offset, disk_byte, found_key.objectid);
                goto out;
        }

        btrfs_debug(fs_info,
                    "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
                    data_offset, ino, num_bytes, logical);

        if (!backref_ctx->found)
                btrfs_debug(fs_info, "no clones found");

        cur_clone_root = NULL;
        for (i = 0; i < sctx->clone_roots_cnt; i++) {
                if (sctx->clone_roots[i].found_refs) {
                        if (!cur_clone_root)
                                cur_clone_root = sctx->clone_roots + i;
                        else if (sctx->clone_roots[i].root == sctx->send_root)
                                /* prefer clones from send_root over others */
                                cur_clone_root = sctx->clone_roots + i;
                }

        }

        if (cur_clone_root) {
                *found = cur_clone_root;
                ret = 0;
        } else {
                ret = -ENOENT;
        }

out:
        btrfs_free_path(tmp_path);
        kfree(backref_ctx);
        return ret;
}

static int read_symlink(struct btrfs_root *root,
                        u64 ino,
                        struct fs_path *dest)
{
        int ret;
        struct btrfs_path *path;
        struct btrfs_key key;
        struct btrfs_file_extent_item *ei;
        u8 type;
        u8 compression;
        unsigned long off;
        int len;

        path = alloc_path_for_send();
        if (!path)
                return -ENOMEM;

        key.objectid = ino;
        key.type = BTRFS_EXTENT_DATA_KEY;
        key.offset = 0;
        ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
        if (ret < 0)
                goto out;
        if (ret) {
                /*
                 * An empty symlink inode. Can happen in rare error paths when
                 * creating a symlink (transaction committed before the inode
                 * eviction handler removed the symlink inode items and a crash
                 * happened in between or the subvol was snapshoted in between).
                 * Print an informative message to dmesg/syslog so that the user
                 * can delete the symlink.
                 */
                btrfs_err(root->fs_info,
                          "Found empty symlink inode %llu at root %llu",
                          ino, root->root_key.objectid);
                ret = -EIO;
                goto out;
        }

        ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
                        struct btrfs_file_extent_item);
        type = btrfs_file_extent_type(path->nodes[0], ei);
        compression = btrfs_file_extent_compression(path->nodes[0], ei);
        BUG_ON(type != BTRFS_FILE_EXTENT_INLINE);
        BUG_ON(compression);

        off = btrfs_file_extent_inline_start(ei);
        len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);

        ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);

out:
        btrfs_free_path(path);
        return ret;
}

/*
 * Helper function to generate a file name that is unique in the root of
 * send_root and parent_root. This is used to generate names for orphan inodes.
 */
static int gen_unique_name(struct send_ctx *sctx,
                           u64 ino, u64 gen,
                           struct fs_path *dest)
{
        int ret = 0;
        struct btrfs_path *path;
        struct btrfs_dir_item *di;
        char tmp[64];
        int len;
        u64 idx = 0;

        path = alloc_path_for_send();
        if (!path)
                return -ENOMEM;

        while (1) {
                len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
                                ino, gen, idx);
                ASSERT(len < sizeof(tmp));

                di = btrfs_lookup_dir_item(NULL, sctx->send_root,
                                path, BTRFS_FIRST_FREE_OBJECTID,
                                tmp, strlen(tmp), 0);
                btrfs_release_path(path);
                if (IS_ERR(di)) {
                        ret = PTR_ERR(di);
                        goto out;
                }
                if (di) {
                        /* not unique, try again */
                        idx++;
                        continue;
                }

                if (!sctx->parent_root) {
                        /* unique */
                        ret = 0;
                        break;
                }

                di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
                                path, BTRFS_FIRST_FREE_OBJECTID,
                                tmp, strlen(tmp), 0);
                btrfs_release_path(path);
                if (IS_ERR(di)) {
                        ret = PTR_ERR(di);
                        goto out;
                }
                if (di) {
                        /* not unique, try again */
                        idx++;
                        continue;
                }
                /* unique */
                break;
        }

        ret = fs_path_add(dest, tmp, strlen(tmp));

out:
        btrfs_free_path(path);
        return ret;
}

enum inode_state {
        inode_state_no_change,
        inode_state_will_create,
        inode_state_did_create,
        inode_state_will_delete,
        inode_state_did_delete,
};

static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen)
{
        int ret;
        int left_ret;
        int right_ret;
        u64 left_gen;
        u64 right_gen;

        ret = get_inode_info(sctx->send_root, ino, NULL, &left_gen, NULL, NULL,
                        NULL, NULL);
        if (ret < 0 && ret != -ENOENT)
                goto out;
        left_ret = ret;

        if (!sctx->parent_root) {
                right_ret = -ENOENT;
        } else {
                ret = get_inode_info(sctx->parent_root, ino, NULL, &right_gen,
                                NULL, NULL, NULL, NULL);
                if (ret < 0 && ret != -ENOENT)
                        goto out;
                right_ret = ret;
        }

        if (!left_ret && !right_ret) {
                if (left_gen == gen && right_gen == gen) {
                        ret = inode_state_no_change;
                } else if (left_gen == gen) {
                        if (ino < sctx->send_progress)
                                ret = inode_state_did_create;
                        else
                                ret = inode_state_will_create;
                } else if (right_gen == gen) {
                        if (ino < sctx->send_progress)
                                ret = inode_state_did_delete;
                        else
                                ret = inode_state_will_delete;
                } else  {
                        ret = -ENOENT;
                }
        } else if (!left_ret) {
                if (left_gen == gen) {
                        if (ino < sctx->send_progress)
                                ret = inode_state_did_create;
                        else
                                ret = inode_state_will_create;
                } else {
                        ret = -ENOENT;
                }
        } else if (!right_ret) {
                if (right_gen == gen) {
                        if (ino < sctx->send_progress)
                                ret = inode_state_did_delete;
                        else
                                ret = inode_state_will_delete;
                } else {
                        ret = -ENOENT;
                }
        } else {
                ret = -ENOENT;
        }

out:
        return ret;
}

static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen)
{
        int ret;

        if (ino == BTRFS_FIRST_FREE_OBJECTID)
                return 1;

        ret = get_cur_inode_state(sctx, ino, gen);
        if (ret < 0)
                goto out;

        if (ret == inode_state_no_change ||
            ret == inode_state_did_create ||
            ret == inode_state_will_delete)
                ret = 1;
        else
                ret = 0;

out:
        return ret;
}

/*
 * Helper function to lookup a dir item in a dir.
 */
static int lookup_dir_item_inode(struct btrfs_root *root,
                                 u64 dir, const char *name, int name_len,
                                 u64 *found_inode,
                                 u8 *found_type)
{
        int ret = 0;
        struct btrfs_dir_item *di;
        struct btrfs_key key;
        struct btrfs_path *path;

        path = alloc_path_for_send();
        if (!path)
                return -ENOMEM;

        di = btrfs_lookup_dir_item(NULL, root, path,
                        dir, name, name_len, 0);
        if (IS_ERR_OR_NULL(di)) {
                ret = di ? PTR_ERR(di) : -ENOENT;
                goto out;
        }
        btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
        if (key.type == BTRFS_ROOT_ITEM_KEY) {
                ret = -ENOENT;
                goto out;
        }
        *found_inode = key.objectid;
        *found_type = btrfs_dir_type(path->nodes[0], di);

out:
        btrfs_free_path(path);
        return ret;
}

/*
 * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
 * generation of the parent dir and the name of the dir entry.
 */
static int get_first_ref(struct btrfs_root *root, u64 ino,
                         u64 *dir, u64 *dir_gen, struct fs_path *name)
{
        int ret;
        struct btrfs_key key;
        struct btrfs_key found_key;
        struct btrfs_path *path;
        int len;
        u64 parent_dir;

        path = alloc_path_for_send();
        if (!path)
                return -ENOMEM;

        key.objectid = ino;
        key.type = BTRFS_INODE_REF_KEY;
        key.offset = 0;

        ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
        if (ret < 0)
                goto out;
        if (!ret)
                btrfs_item_key_to_cpu(path->nodes[0], &found_key,
                                path->slots[0]);
        if (ret || found_key.objectid != ino ||
            (found_key.type != BTRFS_INODE_REF_KEY &&
             found_key.type != BTRFS_INODE_EXTREF_KEY)) {
                ret = -ENOENT;
                goto out;
        }

        if (found_key.type == BTRFS_INODE_REF_KEY) {
                struct btrfs_inode_ref *iref;
                iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
                                      struct btrfs_inode_ref);
                len = btrfs_inode_ref_name_len(path->nodes[0], iref);
                ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
                                                     (unsigned long)(iref + 1),
                                                     len);
                parent_dir = found_key.offset;
        } else {
                struct btrfs_inode_extref *extref;
                extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
                                        struct btrfs_inode_extref);
                len = btrfs_inode_extref_name_len(path->nodes[0], extref);
                ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
                                        (unsigned long)&extref->name, len);
                parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
        }
        if (ret < 0)
                goto out;
        btrfs_release_path(path);

        if (dir_gen) {
                ret = get_inode_info(root, parent_dir, NULL, dir_gen, NULL,
                                     NULL, NULL, NULL);
                if (ret < 0)
                        goto out;
        }

        *dir = parent_dir;

out:
        btrfs_free_path(path);
        return ret;
}

static int is_first_ref(struct btrfs_root *root,
                        u64 ino, u64 dir,
                        const char *name, int name_len)
{
        int ret;
        struct fs_path *tmp_name;
        u64 tmp_dir;

        tmp_name = fs_path_alloc();
        if (!tmp_name)
                return -ENOMEM;

        ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
        if (ret < 0)
                goto out;

        if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
                ret = 0;
                goto out;
        }

        ret = !memcmp(tmp_name->start, name, name_len);

out:
        fs_path_free(tmp_name);
        return ret;
}

/*
 * Used by process_recorded_refs to determine if a new ref would overwrite an
 * already existing ref. In case it detects an overwrite, it returns the
 * inode/gen in who_ino/who_gen.
 * When an overwrite is detected, process_recorded_refs does proper orphanizing
 * to make sure later references to the overwritten inode are possible.
 * Orphanizing is however only required for the first ref of an inode.
 * process_recorded_refs does an additional is_first_ref check to see if
 * orphanizing is really required.
 */
static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
                              const char *name, int name_len,
                              u64 *who_ino, u64 *who_gen, u64 *who_mode)
{
        int ret = 0;
        u64 gen;
        u64 other_inode = 0;
        u8 other_type = 0;

        if (!sctx->parent_root)
                goto out;

        ret = is_inode_existent(sctx, dir, dir_gen);
        if (ret <= 0)
                goto out;

        /*
         * If we have a parent root we need to verify that the parent dir was
         * not deleted and then re-created, if it was then we have no overwrite
         * and we can just unlink this entry.
         */
        if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID) {
                ret = get_inode_info(sctx->parent_root, dir, NULL, &gen, NULL,
                                     NULL, NULL, NULL);
                if (ret < 0 && ret != -ENOENT)
                        goto out;
                if (ret) {
                        ret = 0;
                        goto out;
                }
                if (gen != dir_gen)
                        goto out;
        }

        ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
                        &other_inode, &other_type);
        if (ret < 0 && ret != -ENOENT)
                goto out;
        if (ret) {
                ret = 0;
                goto out;
        }

        /*
         * Check if the overwritten ref was already processed. If yes, the ref
         * was already unlinked/moved, so we can safely assume that we will not
         * overwrite anything at this point in time.
         */
        if (other_inode > sctx->send_progress ||
            is_waiting_for_move(sctx, other_inode)) {
                ret = get_inode_info(sctx->parent_root, other_inode, NULL,
                                who_gen, who_mode, NULL, NULL, NULL);
                if (ret < 0)
                        goto out;

                ret = 1;
                *who_ino = other_inode;
        } else {
                ret = 0;
        }

out:
        return ret;
}

/*
 * Checks if the ref was overwritten by an already processed inode. This is
 * used by __get_cur_name_and_parent to find out if the ref was orphanized and
 * thus the orphan name needs be used.
 * process_recorded_refs also uses it to avoid unlinking of refs that were
 * overwritten.
 */
static int did_overwrite_ref(struct send_ctx *sctx,
                            u64 dir, u64 dir_gen,
                            u64 ino, u64 ino_gen,
                            const char *name, int name_len)
{
        int ret = 0;
        u64 gen;
        u64 ow_inode;
        u8 other_type;

        if (!sctx->parent_root)
                goto out;

        ret = is_inode_existent(sctx, dir, dir_gen);
        if (ret <= 0)
                goto out;

        if (dir != BTRFS_FIRST_FREE_OBJECTID) {
                ret = get_inode_info(sctx->send_root, dir, NULL, &gen, NULL,
                                     NULL, NULL, NULL);
                if (ret < 0 && ret != -ENOENT)
                        goto out;
                if (ret) {
                        ret = 0;
                        goto out;
                }
                if (gen != dir_gen)
                        goto out;
        }

        /* check if the ref was overwritten by another ref */
        ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
                        &ow_inode, &other_type);
        if (ret < 0 && ret != -ENOENT)
                goto out;
        if (ret) {
                /* was never and will never be overwritten */
                ret = 0;
                goto out;
        }

        ret = get_inode_info(sctx->send_root, ow_inode, NULL, &gen, NULL, NULL,
                        NULL, NULL);
        if (ret < 0)
                goto out;

        if (ow_inode == ino && gen == ino_gen) {
                ret = 0;
                goto out;
        }

        /*
         * We know that it is or will be overwritten. Check this now.
         * The current inode being processed might have been the one that caused
         * inode 'ino' to be orphanized, therefore check if ow_inode matches
         * the current inode being processed.
         */
        if ((ow_inode < sctx->send_progress) ||
            (ino != sctx->cur_ino && ow_inode == sctx->cur_ino &&
             gen == sctx->cur_inode_gen))
                ret = 1;
        else
                ret = 0;

out:
        return ret;
}

/*
 * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
 * that got overwritten. This is used by process_recorded_refs to determine
 * if it has to use the path as returned by get_cur_path or the orphan name.
 */
static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
{
        int ret = 0;
        struct fs_path *name = NULL;
        u64 dir;
        u64 dir_gen;

        if (!sctx->parent_root)
                goto out;

        name = fs_path_alloc();
        if (!name)
                return -ENOMEM;

        ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
        if (ret < 0)
                goto out;

        ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
                        name->start, fs_path_len(name));

out:
        fs_path_free(name);
        return ret;
}

/*
 * Insert a name cache entry. On 32bit kernels the radix tree index is 32bit,
 * so we need to do some special handling in case we have clashes. This function
 * takes care of this with the help of name_cache_entry::radix_list.
 * In case of error, nce is kfreed.
 */
static int name_cache_insert(struct send_ctx *sctx,
                             struct name_cache_entry *nce)
{
        int ret = 0;
        struct list_head *nce_head;

        nce_head = radix_tree_lookup(&sctx->name_cache,
                        (unsigned long)nce->ino);
        if (!nce_head) {
                nce_head = kmalloc(sizeof(*nce_head), GFP_KERNEL);
                if (!nce_head) {
                        kfree(nce);
                        return -ENOMEM;
                }
                INIT_LIST_HEAD(nce_head);

                ret = radix_tree_insert(&sctx->name_cache, nce->ino, nce_head);
                if (ret < 0) {
                        kfree(nce_head);
                        kfree(nce);
                        return ret;
                }
        }
        list_add_tail(&nce->radix_list, nce_head);
        list_add_tail(&nce->list, &sctx->name_cache_list);
        sctx->name_cache_size++;

        return ret;
}

static void name_cache_delete(struct send_ctx *sctx,
                              struct name_cache_entry *nce)
{
        struct list_head *nce_head;

        nce_head = radix_tree_lookup(&sctx->name_cache,
                        (unsigned long)nce->ino);
        if (!nce_head) {
                btrfs_err(sctx->send_root->fs_info,
              "name_cache_delete lookup failed ino %llu cache size %d, leaking memory",
                        nce->ino, sctx->name_cache_size);
        }

        list_del(&nce->radix_list);
        list_del(&nce->list);
        sctx->name_cache_size--;

        /*
         * We may not get to the final release of nce_head if the lookup fails
         */
        if (nce_head && list_empty(nce_head)) {
                radix_tree_delete(&sctx->name_cache, (unsigned long)nce->ino);
                kfree(nce_head);
        }
}

static struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
                                                    u64 ino, u64 gen)
{
        struct list_head *nce_head;
        struct name_cache_entry *cur;

        nce_head = radix_tree_lookup(&sctx->name_cache, (unsigned long)ino);
        if (!nce_head)
                return NULL;

        list_for_each_entry(cur, nce_head, radix_list) {
                if (cur->ino == ino && cur->gen == gen)
                        return cur;
        }
        return NULL;
}

/*
 * Remove some entries from the beginning of name_cache_list.
 */
static void name_cache_clean_unused(struct send_ctx *sctx)
{
        struct name_cache_entry *nce;

        if (sctx->name_cache_size < SEND_CTX_NAME_CACHE_CLEAN_SIZE)
                return;

        while (sctx->name_cache_size > SEND_CTX_MAX_NAME_CACHE_SIZE) {
                nce = list_entry(sctx->name_cache_list.next,
                                struct name_cache_entry, list);
                name_cache_delete(sctx, nce);
                kfree(nce);
        }
}

static void name_cache_free(struct send_ctx *sctx)
{
        struct name_cache_entry *nce;

        while (!list_empty(&sctx->name_cache_list)) {
                nce = list_entry(sctx->name_cache_list.next,
                                struct name_cache_entry, list);
                name_cache_delete(sctx, nce);
                kfree(nce);
        }
}

/*
 * Used by get_cur_path for each ref up to the root.
 * Returns 0 if it succeeded.
 * Returns 1 if the inode is not existent or got overwritten. In that case, the
 * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
 * is returned, parent_ino/parent_gen are not guaranteed to be valid.
 * Returns <0 in case of error.
 */
static int __get_cur_name_and_parent(struct send_ctx *sctx,
                                     u64 ino, u64 gen,
                                     u64 *parent_ino,
                                     u64 *parent_gen,
                                     struct fs_path *dest)
{
        int ret;
        int nce_ret;
        struct name_cache_entry *nce = NULL;

        /*
         * First check if we already did a call to this function with the same
         * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
         * return the cached result.
         */
        nce = name_cache_search(sctx, ino, gen);
        if (nce) {
                if (ino < sctx->send_progress && nce->need_later_update) {
                        name_cache_delete(sctx, nce);
                        kfree(nce);
                        nce = NULL;
                } else {
                        /*
                         * Removes the entry from the list and adds it back to
                         * the end.  This marks the entry as recently used so
                         * that name_cache_clean_unused does not remove it.
                         */
                        list_move_tail(&nce->list, &sctx->name_cache_list);

                        *parent_ino = nce->parent_ino;
                        *parent_gen = nce->parent_gen;
                        ret = fs_path_add(dest, nce->name, nce->name_len);
                        if (ret < 0)
                                goto out;
                        ret = nce->ret;
                        goto out;
                }
        }

        /*
         * If the inode is not existent yet, add the orphan name and return 1.
         * This should only happen for the parent dir that we determine in
         * __record_new_ref
         */
        ret = is_inode_existent(sctx, ino, gen);
        if (ret < 0)
                goto out;

        if (!ret) {
                ret = gen_unique_name(sctx, ino, gen, dest);
                if (ret < 0)
                        goto out;
                ret = 1;
                goto out_cache;
        }

        /*
         * Depending on whether the inode was already processed or not, use
         * send_root or parent_root for ref lookup.
         */
        if (ino < sctx->send_progress)
                ret = get_first_ref(sctx->send_root, ino,
                                    parent_ino, parent_gen, dest);
        else
                ret = get_first_ref(sctx->parent_root, ino,
                                    parent_ino, parent_gen, dest);
        if (ret < 0)
                goto out;

        /*
         * Check if the ref was overwritten by an inode's ref that was processed
         * earlier. If yes, treat as orphan and return 1.
         */
        ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
                        dest->start, dest->end - dest->start);
        if (ret < 0)
                goto out;
        if (ret) {
                fs_path_reset(dest);
                ret = gen_unique_name(sctx, ino, gen, dest);
                if (ret < 0)
                        goto out;
                ret = 1;
        }

out_cache:
        /*
         * Store the result of the lookup in the name cache.
         */
        nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
        if (!nce) {
                ret = -ENOMEM;
                goto out;
        }

        nce->ino = ino;
        nce->gen = gen;
        nce->parent_ino = *parent_ino;
        nce->parent_gen = *parent_gen;
        nce->name_len = fs_path_len(dest);
        nce->ret = ret;
        strcpy(nce->name, dest->start);

        if (ino < sctx->send_progress)
                nce->need_later_update = 0;
        else
                nce->need_later_update = 1;

        nce_ret = name_cache_insert(sctx, nce);
        if (nce_ret < 0)
                ret = nce_ret;
        name_cache_clean_unused(sctx);

out:
        return ret;
}

/*
 * Magic happens here. This function returns the first ref to an inode as it
 * would look like while receiving the stream at this point in time.
 * We walk the path up to the root. For every inode in between, we check if it
 * was already processed/sent. If yes, we continue with the parent as found
 * in send_root. If not, we continue with the parent as found in parent_root.
 * If we encounter an inode that was deleted at this point in time, we use the
 * inodes "orphan" name instead of the real name and stop. Same with new inodes
 * that were not created yet and overwritten inodes/refs.
 *
 * When do we have orphan inodes:
 * 1. When an inode is freshly created and thus no valid refs are available yet
 * 2. When a directory lost all it's refs (deleted) but still has dir items
 *    inside which were not processed yet (pending for move/delete). If anyone
 *    tried to get the path to the dir items, it would get a path inside that
 *    orphan directory.
 * 3. When an inode is moved around or gets new links, it may overwrite the ref
 *    of an unprocessed inode. If in that case the first ref would be
 *    overwritten, the overwritten inode gets "orphanized". Later when we
 *    process this overwritten inode, it is restored at a new place by moving
 *    the orphan inode.
 *
 * sctx->send_progress tells this function at which point in time receiving
 * would be.
 */
static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
                        struct fs_path *dest)
{
        int ret = 0;
        struct fs_path *name = NULL;
        u64 parent_inode = 0;
        u64 parent_gen = 0;
        int stop = 0;

        name = fs_path_alloc();
        if (!name) {
                ret = -ENOMEM;
                goto out;
        }

        dest->reversed = 1;
        fs_path_reset(dest);

        while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
                struct waiting_dir_move *wdm;

                fs_path_reset(name);

                if (is_waiting_for_rm(sctx, ino, gen)) {
                        ret = gen_unique_name(sctx, ino, gen, name);
                        if (ret < 0)
                                goto out;
                        ret = fs_path_add_path(dest, name);
                        break;
                }

                wdm = get_waiting_dir_move(sctx, ino);
                if (wdm && wdm->orphanized) {
                        ret = gen_unique_name(sctx, ino, gen, name);
                        stop = 1;
                } else if (wdm) {
                        ret = get_first_ref(sctx->parent_root, ino,
                                            &parent_inode, &parent_gen, name);
                } else {
                        ret = __get_cur_name_and_parent(sctx, ino, gen,
                                                        &parent_inode,
                                                        &parent_gen, name);
                        if (ret)
                                stop = 1;
                }

                if (ret < 0)
                        goto out;

                ret = fs_path_add_path(dest, name);
                if (ret < 0)
                        goto out;

                ino = parent_inode;
                gen = parent_gen;
        }

out:
        fs_path_free(name);
        if (!ret)
                fs_path_unreverse(dest);
        return ret;
}

/*
 * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
 */
static int send_subvol_begin(struct send_ctx *sctx)
{
        int ret;
        struct btrfs_root *send_root = sctx->send_root;
        struct btrfs_root *parent_root = sctx->parent_root;
        struct btrfs_path *path;
        struct btrfs_key key;
        struct btrfs_root_ref *ref;
        struct extent_buffer *leaf;
        char *name = NULL;
        int namelen;

        path = btrfs_alloc_path();
        if (!path)
                return -ENOMEM;

        name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
        if (!name) {
                btrfs_free_path(path);
                return -ENOMEM;
        }

        key.objectid = send_root->root_key.objectid;
        key.type = BTRFS_ROOT_BACKREF_KEY;
        key.offset = 0;

        ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
                                &key, path, 1, 0);
        if (ret < 0)
                goto out;
        if (ret) {
                ret = -ENOENT;
                goto out;
        }

        leaf = path->nodes[0];
        btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
        if (key.type != BTRFS_ROOT_BACKREF_KEY ||
            key.objectid != send_root->root_key.objectid) {
                ret = -ENOENT;
                goto out;
        }
        ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
        namelen = btrfs_root_ref_name_len(leaf, ref);
        read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
        btrfs_release_path(path);

        if (parent_root) {
                ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
                if (ret < 0)
                        goto out;
        } else {
                ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
                if (ret < 0)
                        goto out;
        }

        TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);

        if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
                TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
                            sctx->send_root->root_item.received_uuid);
        else
                TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
                            sctx->send_root->root_item.uuid);

        TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
                    btrfs_root_ctransid(&sctx->send_root->root_item));
        if (parent_root) {
                if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
                        TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
                                     parent_root->root_item.received_uuid);
                else
                        TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
                                     parent_root->root_item.uuid);
                TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
                            btrfs_root_ctransid(&sctx->parent_root->root_item));
        }

        ret = send_cmd(sctx);

tlv_put_failure:
out:
        btrfs_free_path(path);
        kfree(name);
        return ret;
}

static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret = 0;
        struct fs_path *p;

        btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);

        p = fs_path_alloc();
        if (!p)
                return -ENOMEM;

        ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
        if (ret < 0)
                goto out;

        ret = get_cur_path(sctx, ino, gen, p);
        if (ret < 0)
                goto out;
        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
        TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);

        ret = send_cmd(sctx);

tlv_put_failure:
out:
        fs_path_free(p);
        return ret;
}

static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret = 0;
        struct fs_path *p;

        btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);

        p = fs_path_alloc();
        if (!p)
                return -ENOMEM;

        ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
        if (ret < 0)
                goto out;

        ret = get_cur_path(sctx, ino, gen, p);
        if (ret < 0)
                goto out;
        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
        TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);

        ret = send_cmd(sctx);

tlv_put_failure:
out:
        fs_path_free(p);
        return ret;
}

static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret = 0;
        struct fs_path *p;

        btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
                    ino, uid, gid);

        p = fs_path_alloc();
        if (!p)
                return -ENOMEM;

        ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
        if (ret < 0)
                goto out;

        ret = get_cur_path(sctx, ino, gen, p);
        if (ret < 0)
                goto out;
        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
        TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
        TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);

        ret = send_cmd(sctx);

tlv_put_failure:
out:
        fs_path_free(p);
        return ret;
}

static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret = 0;
        struct fs_path *p = NULL;
        struct btrfs_inode_item *ii;
        struct btrfs_path *path = NULL;
        struct extent_buffer *eb;
        struct btrfs_key key;
        int slot;

        btrfs_debug(fs_info, "send_utimes %llu", ino);

        p = fs_path_alloc();
        if (!p)
                return -ENOMEM;

        path = alloc_path_for_send();
        if (!path) {
                ret = -ENOMEM;
                goto out;
        }

        key.objectid = ino;
        key.type = BTRFS_INODE_ITEM_KEY;
        key.offset = 0;
        ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
        if (ret > 0)
                ret = -ENOENT;
        if (ret < 0)
                goto out;

        eb = path->nodes[0];
        slot = path->slots[0];
        ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);

        ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
        if (ret < 0)
                goto out;

        ret = get_cur_path(sctx, ino, gen, p);
        if (ret < 0)
                goto out;
        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
        TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
        TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
        TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
        /* TODO Add otime support when the otime patches get into upstream */

        ret = send_cmd(sctx);

tlv_put_failure:
out:
        fs_path_free(p);
        btrfs_free_path(path);
        return ret;
}

/*
 * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
 * a valid path yet because we did not process the refs yet. So, the inode
 * is created as orphan.
 */
static int send_create_inode(struct send_ctx *sctx, u64 ino)
{
        struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
        int ret = 0;
        struct fs_path *p;
        int cmd;
        u64 gen;
        u64 mode;
        u64 rdev;

        btrfs_debug(fs_info, "send_create_inode %llu", ino);

        p = fs_path_alloc();
        if (!p)
                return -ENOMEM;

        if (ino != sctx->cur_ino) {
                ret = get_inode_info(sctx->send_root, ino, NULL, &gen, &mode,
                                     NULL, NULL, &rdev);
                if (ret < 0)
                        goto out;
        } else {
                gen = sctx->cur_inode_gen;
                mode = sctx->cur_inode_mode;
                rdev = sctx->cur_inode_rdev;
        }

        if (S_ISREG(mode)) {
                cmd = BTRFS_SEND_C_MKFILE;
        } else if (S_ISDIR(mode)) {
                cmd = BTRFS_SEND_C_MKDIR;
        } else if (S_ISLNK(mode)) {
                cmd = BTRFS_SEND_C_SYMLINK;
        } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
                cmd = BTRFS_SEND_C_MKNOD;
        } else if (S_ISFIFO(mode)) {
                cmd = BTRFS_SEND_C_MKFIFO;
        } else if (S_ISSOCK(mode)) {
                cmd = BTRFS_SEND_C_MKSOCK;
        } else {
                btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
                                (int)(mode & S_IFMT));
                ret = -EOPNOTSUPP;
                goto out;
        }

        ret = begin_cmd(sctx, cmd);
        if (ret < 0)
                goto out;

        ret = gen_unique_name(sctx, ino, gen, p);
        if (ret < 0)
                goto out;

        TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
        TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);

        if (S_ISLNK(mode)) {
                fs_path_reset(p);
                ret = read_symlink(sctx->send_root, ino, p);
                if (ret < 0)
                        goto out;
                TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
        } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
                   S_ISFIFO(mode) || S_ISSOCK(mode)) {
                TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
                TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
        }

        ret = send_cmd(sctx);
        if (ret < 0)
                goto out;


tlv_put_failure:
out:
        fs_path_free(p);
        return ret;
}

/*
 * We need some special handling for inodes that get processed before the parent
 * directory got created. See process_recorded_refs for details.
 * This function does the check if we already created the dir out of order.
 */
static int did_create_dir(struct send_ctx *sctx, u64 dir)
{
        int ret = 0;
        struct btrfs_path *path = NULL;
        struct btrfs_key key;
        struct btrfs_key found_key;
        struct btrfs_key di_key;
        struct extent_buffer *eb;
        struct btrfs_dir_item *di;
        int slot;

        path = alloc_path_for_send();
        if (!path) {
                ret = -ENOMEM;
                goto out;
        }

        key.objectid = dir;
        key.type = BTRFS_DIR_INDEX_KEY;
        key.offset = 0;
        ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
        if (ret < 0)
                goto out;

        while (1) {
                eb = path->nodes[0];
                slot = path->slots[0];
                if (slot >= btrfs_header_nritems(eb)) {
                        ret = btrfs_next_leaf(sctx->send_root, path);
                        if (ret < 0) {
                                goto out;
                        } else if (ret > 0) {
                                ret = 0;
                                break;
                        }
                        continue;
                }

                btrfs_item_key_to_cpu(eb, &found_key, slot);
                if (found_key.objectid != key.objectid ||
                    found_key.type != key.type) {
                        ret = 0;
                        goto out;
                }

                di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
                btrfs_dir_item_key_to_cpu(eb, di, &di_key);

                if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
                    di_key.objectid < sctx->send_progress) {
                        ret = 1;
                        goto out;
                }

                path->slots[0]++;
        }

out:
        btrfs_free_path(path);
        return ret;
}

/*
 * Only creates the inode if it is:
 * 1. Not a directory
 * 2. Or a directory which was not created already due to out of order
 *    directories. See did_create_dir and process_recorded_refs for details.
 */
static int send_create_inode_if_needed(struct send_ctx *sctx)
{
        int ret;

        if (S_ISDIR(sctx->cur_inode_mode)) {
                ret = did_create_dir(sctx, sctx->cur_ino);
                if (ret < 0)
                        goto out;
                if (ret) {
                        ret = 0;
                        goto out;
                }
        }

        ret = send_create_inode(sctx, sctx->cur_ino);
        if (ret < 0)
                goto out;

out:
        return ret;
}

struct recorded_ref {
        struct list_head list;
        char *name;
        struct fs_path *full_path;
        u64 dir;
        u64 dir_gen;
        int name_len;
};

static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
{
        ref->full_path = path;
        ref->name = (char *)kbasename(ref->full_path->start);
        ref->name_len = ref->full_path->end - ref->name;
}

/*
 * We need to process new refs before deleted refs, but compare_tree gives us
 * everything mixed. So we first record all refs and later process them.
 * This function is a helper to record one ref.
 */
static int __record_ref(struct list_head *head, u64 dir,
                      u64 dir_gen, struct fs_path *path)
{
        struct recorded_ref *ref;

        ref = kmalloc(sizeof(*ref), GFP_KERNEL);
        if (!ref)
                return -ENOMEM;

        ref->dir = dir;
        ref->dir_gen = dir_gen;
        set_ref_path(ref, path);
        list_add_tail(&ref->list, head);
        return 0;
}

static int dup_ref(struct recorded_ref *ref, struct list_head *list)
{
        struct recorded_ref *new;

        new = kmalloc(sizeof(*ref), GFP_KERNEL);
        if (!new)
                return -ENOMEM;

        new->dir = ref->dir;
        new->dir_gen = ref->dir_gen;
        new->full_path = NULL;
        INIT_LIST_HEAD(&new->list);
        list_add_tail(&new->list, list);
        return 0;
}

static void __free_recorded_refs(struct list_head *head)
{
        struct recorded_ref *cur;

        while (!list_empty(head)) {
                cur = list_entry(head->next, struct recorded_ref, list);
                fs_path_free(cur->full_path);
                list_del(&cur->list);
                kfree(cur);
        }
}

static void free_recorded_refs(struct send_ctx *sctx)
{
        __free_recorded_refs(&sctx->new_refs);
        __free_recorded_refs(&sctx->deleted_refs);
}

/*
 * Renames/moves a file/dir to its orphan name. Used when the first
 * ref of an unprocessed inode gets overwritten and for all non empty
 * directories.
 */
static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
                          struct fs_path *path)
{
        int ret;
        struct fs_path *orphan;

        orphan = fs_path_alloc();
        if (!orphan)
                return -ENOMEM;

        ret = gen_unique_name(sctx, ino, gen, orphan);
        if (ret < 0)
                goto out;

        ret = send_rename(sctx, path, orphan);

out:
        fs_path_free(orphan);
        return ret;
}

static struct orphan_dir_info *add_orphan_dir_info(struct send_ctx *sctx,
                                                   u64 dir_ino, u64 dir_gen)
{
        struct rb_node **p = &sctx->orphan_dirs.rb_node;
        struct rb_node *parent = NULL;
        struct orphan_dir_info *entry, *odi;

        while (*p) {
                parent = *p;
                entry = rb_entry(parent, struct orphan_dir_info, node);
                if (dir_ino < entry->ino)
                        p = &(*p)->rb_left;
                else if (dir_ino > entry->ino)
                        p = &(*p)->rb_right;
                else if (dir_gen < entry->gen)
                        p = &(*p)->rb_left;
                else if (dir_gen > entry->gen)
                        p = &(*p)->rb_right;
                else
                        return entry;
        }

        odi = kmalloc(sizeof(*odi), GFP_KERNEL);
        if (!odi)
                return ERR_PTR(-ENOMEM);
        odi->ino = dir_ino;
        odi->gen = dir_gen;
        odi->last_dir_index_offset = 0;

        rb_link_node(&odi->node, parent, p);
        rb_insert_color(&odi->node, &sctx->orphan_dirs);
        return odi;
}

static struct orphan_dir_info *get_orphan_dir_info(struct send_ctx *sctx,
                                                   u64 dir_ino, u64 gen)
{
        struct rb_node *n = sctx->orphan_dirs.rb_node;
        struct orphan_dir_info *entry;

        while (n) {
                entry = rb_entry(n, struct orphan_dir_info, node);
                if (dir_ino < entry->ino)
                        n = n->rb_left;
                else if (dir_ino > entry->ino)
                        n = n->rb_right;
                else if (gen < entry->gen)
                        n = n->rb_left;
                else if (gen > entry->gen)
                        n = n->rb_right;
                else
                        return entry;
        }
        return NULL;
}

static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen)
{
        struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino, gen);

        return odi != NULL;
}

static void free_orphan_dir_info(struct send_ctx *sctx,
                                 struct orphan_dir_info *odi)
{
        if (!odi)
                return;
        rb_erase(&odi->node, &sctx->orphan_dirs);
        kfree(odi);
}

/*
 * Returns 1 if a directory can be removed at this point in time.
 * We check this by iterating all dir items and checking if the inode behind
 * the dir item was already processed.
 */
static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen,
                     u64 send_progress)
{
        int ret = 0;
        struct btrfs_root *root = sctx->parent_root;
        struct btrfs_path *path;
        struct btrfs_key key;
        struct btrfs_key found_key;
        struct btrfs_key loc;
        struct btrfs_dir_item *di;
        struct orphan_dir_info *odi = NULL;

        /*
         * Don't try to rmdir the top/root subvolume dir.
         */
        if (dir == BTRFS_FIRST_FREE_OBJECTID)
                return 0;

        path = alloc_path_for_send();
        if (!path)
                return -ENOMEM;

        key.objectid = dir;
        key.type = BTRFS_DIR_INDEX_KEY;
        key.offset = 0;

        odi = get_orphan_dir_info(sctx, dir, dir_gen);
        if (odi)
                key.offset = odi->last_dir_index_offset;

        ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
        if (ret < 0)
                goto out;

        while (1) {
                struct waiting_dir_move *dm;

                if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
                        ret = btrfs_next_leaf(root, path);
                        if (ret < 0)
                                goto out;
                        else if (ret > 0)
                                break;
                        continue;
                }
                btrfs_item_key_to_cpu(path->nodes[0], &found_key,
                                      path->slots[0]);
                if (found_key.objectid != key.objectid ||
                    found_key.type != key.type)
                        break;

                di = btrfs_item_ptr(path->nodes[0], path->slots[0],
                                struct btrfs_dir_item);
                btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);

                dm = get_waiting_dir_move(sctx, loc.objectid);
                if (dm) {
                        odi = add_orphan_dir_info(sctx, dir, dir_gen);
                        if (IS_ERR(odi)) {
                                ret = PTR_ERR(odi);
                                goto out;
                        }
                        odi->gen = dir_gen;
                        odi->last_dir_index_offset = found_key.offset;
                        dm->rmdir_ino = dir;
                        dm->rmdir_gen = dir_gen;
                        ret = 0;
                        goto out;
                }

                if (loc.objectid > send_progress) {
                        odi = add_orphan_dir_info(sctx, dir, dir_gen);
                        if (IS_ERR(odi)) {
                                ret = PTR_ERR(odi);
                                goto out;
                        }
                        odi->gen = dir_gen;
                        odi->last_dir_index_offset = found_key.offset;
                        ret = 0;
                        goto out;
                }