inode.c: fix initialization of i_streams

[wimlib] / src / dentry.c

/*
 * dentry.c - see description below
 */

/*
 * Copyright (C) 2012, 2013, 2014, 2015 Eric Biggers
 *
 * This file is free software; you can redistribute it and/or modify it under
 * the terms of the GNU Lesser General Public License as published by the Free
 * Software Foundation; either version 3 of the License, or (at your option) any
 * later version.
 *
 * This file is distributed in the hope that it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
 * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
 * details.
 *
 * You should have received a copy of the GNU Lesser General Public License
 * along with this file; if not, see http://www.gnu.org/licenses/.
 */

/*
 * This file contains logic to deal with WIM directory entries, or "dentries":
 *
 *  - Reading a dentry tree from a metadata resource in a WIM file
 *  - Writing a dentry tree to a metadata resource in a WIM file
 *  - Iterating through a tree of WIM dentries
 *  - Path lookup: translating a path into a WIM dentry or inode
 *  - Creating, modifying, and deleting WIM dentries
 *
 * Notes:
 *
 *  - A WIM file can contain multiple images, each of which has an independent
 *    tree of dentries.  "On disk", the dentry tree for an image is stored in
 *    the "metadata resource" for that image.
 *
 *  - Multiple dentries in an image may correspond to the same inode, or "file".
 *    When this occurs, it means that the file has multiple names, or "hard
 *    links".  A dentry is not a file, but rather the name of a file!
 *
 *  - Inodes are not represented explicitly in the WIM file format.  Instead,
 *    the metadata resource provides a "hard link group ID" for each dentry.
 *    wimlib handles pulling out actual inodes from this information, but this
 *    occurs in inode_fixup.c and not in this file.
 *
 *  - wimlib does not allow *directory* hard links, so a WIM image really does
 *    have a *tree* of dentries (and not an arbitrary graph of dentries).
 *
 *  - wimlib indexes dentries both case-insensitively and case-sensitively,
 *    allowing either behavior to be used for path lookup.
 *
 *  - Multiple dentries in a directory might have the same case-insensitive
 *    name.  But wimlib enforces that at most one dentry in a directory can have
 *    a given case-sensitive name.
 */

#ifdef HAVE_CONFIG_H
#  include "config.h"
#endif

#include <errno.h>

#include "wimlib/assert.h"
#include "wimlib/dentry.h"
#include "wimlib/inode.h"
#include "wimlib/encoding.h"
#include "wimlib/endianness.h"
#include "wimlib/metadata.h"
#include "wimlib/paths.h"

/* On-disk format of a WIM dentry (directory entry), located in the metadata
 * resource for a WIM image.  */
struct wim_dentry_on_disk {

        /* Length of this directory entry in bytes, not including any extra
         * stream entries.  Should be a multiple of 8 so that the following
         * dentry or extra stream entry is aligned on an 8-byte boundary.  (If
         * not, wimlib will round it up.)  It must be at least as long as the
         * fixed-length fields of the dentry (WIM_DENTRY_DISK_SIZE), plus the
         * lengths of the file name and/or short name if present, plus the size
         * of any "extra" data.
         *
         * It is also possible for this field to be 0.  This case indicates the
         * end of a list of sibling entries in a directory.  It also means the
         * real length is 8, because the dentry included only the length field,
         * but that takes up 8 bytes.  */
        le64 length;

        /* File attributes for the file or directory.  This is a bitwise OR of
         * the FILE_ATTRIBUTE_* constants and should correspond to the value
         * retrieved by GetFileAttributes() on Windows. */
        le32 attributes;

        /* A value that specifies the security descriptor for this file or
         * directory.  If -1, the file or directory has no security descriptor.
         * Otherwise, it is a 0-based index into the WIM image's table of
         * security descriptors (see: `struct wim_security_data') */
        sle32 security_id;

        /* Offset, in bytes, from the start of the uncompressed metadata
         * resource of this directory's child directory entries, or 0 if this
         * directory entry does not correspond to a directory or otherwise does
         * not have any children. */
        le64 subdir_offset;

        /* Reserved fields */
        le64 unused_1;
        le64 unused_2;

        /* Creation time, last access time, and last write time, in
         * 100-nanosecond intervals since 12:00 a.m UTC January 1, 1601.  They
         * should correspond to the times gotten by calling GetFileTime() on
         * Windows. */
        le64 creation_time;
        le64 last_access_time;
        le64 last_write_time;

        /*
         * Usually this is the SHA-1 message digest of the file's "contents"
         * (the unnamed data stream).
         *
         * If the file has FILE_ATTRIBUTE_REPARSE_POINT set, then this is
         * instead usually the SHA-1 message digest of the uncompressed reparse
         * point data.
         *
         * However, there are some special rules that need to be applied to
         * interpret this field correctly when extra stream entries are present.
         * See the code for details.
         */
        u8 default_hash[SHA1_HASH_SIZE];

        /* The format of the following data is not yet completely known and they
         * do not correspond to Microsoft's documentation.
         *
         * If this directory entry is for a reparse point (has
         * FILE_ATTRIBUTE_REPARSE_POINT set in the 'attributes' field), then the
         * version of the following fields containing the reparse tag is valid.
         * Furthermore, the field notated as not_rpfixed, as far as I can tell,
         * is supposed to be set to 1 if reparse point fixups (a.k.a. fixing the
         * targets of absolute symbolic links) were *not* done, and otherwise 0.
         *
         * If this directory entry is not for a reparse point, then the version
         * of the following fields containing the hard_link_group_id is valid.
         * All MS says about this field is that "If this file is part of a hard
         * link set, all the directory entries in the set will share the same
         * value in this field.".  However, more specifically I have observed
         * the following:
         *    - If the file is part of a hard link set of size 1, then the
         *    hard_link_group_id should be set to either 0, which is treated
         *    specially as indicating "not hardlinked", or any unique value.
         *    - The specific nonzero values used to identity hard link sets do
         *    not matter, as long as they are unique.
         *    - However, due to bugs in Microsoft's software, it is actually NOT
         *    guaranteed that directory entries that share the same hard link
         *    group ID are actually hard linked to each either.  See
         *    inode_fixup.c for the code that handles this.
         */
        union {
                struct {
                        le32 rp_unknown_1;
                        le32 reparse_tag;
                        le16 rp_unknown_2;
                        le16 not_rpfixed;
                } _packed_attribute reparse;
                struct {
                        le32 rp_unknown_1;
                        le64 hard_link_group_id;
                } _packed_attribute nonreparse;
        };

        /* Number of extra stream entries that directly follow this dentry
         * on-disk.  */
        le16 num_extra_streams;

        /* If nonzero, this is the length, in bytes, of this dentry's UTF-16LE
         * encoded short name (8.3 DOS-compatible name), excluding the null
         * terminator.  If zero, then the long name of this dentry does not have
         * a corresponding short name (but this does not exclude the possibility
         * that another dentry for the same file has a short name).  */
        le16 short_name_nbytes;

        /* If nonzero, this is the length, in bytes, of this dentry's UTF-16LE
         * encoded "long" name, excluding the null terminator.  If zero, then
         * this file has no long name.  The root dentry should not have a long
         * name, but all other dentries in the image should have long names.  */
        le16 file_name_nbytes;

        /* Beginning of optional, variable-length fields  */

        /* If file_name_nbytes != 0, the next field will be the UTF-16LE encoded
         * long file name.  This will be null-terminated, so the size of this
         * field will really be file_name_nbytes + 2.  */
        /*utf16lechar file_name[];*/

        /* If short_name_nbytes != 0, the next field will be the UTF-16LE
         * encoded short name.  This will be null-terminated, so the size of
         * this field will really be short_name_nbytes + 2.  */
        /*utf16lechar short_name[];*/

        /* If there is still space in the dentry (according to the 'length'
         * field) after 8-byte alignment, then the remaining space will be a
         * variable-length list of tagged metadata items.  See tagged_items.c
         * for more information.  */
        /* u8 tagged_items[] _aligned_attribute(8); */

} _packed_attribute;
        /* If num_extra_streams != 0, then there are that many extra stream
         * entries following the dentry, starting on the next 8-byte aligned
         * boundary.  They are not counted in the 'length' field of the dentry.
         */

/* On-disk format of an extra stream entry.  This represents an extra NTFS-style
 * "stream" associated with the file, such as a named data stream.  */
struct wim_extra_stream_entry_on_disk {

        /* Length of this extra stream entry, in bytes.  This includes all
         * fixed-length fields, plus the name and null terminator if present,
         * and any needed padding such that the length is a multiple of 8.  */
        le64 length;

        /* Reserved field  */
        le64 reserved;

        /* SHA-1 message digest of this stream's uncompressed data, or all
         * zeroes if this stream's data is of zero length.  */
        u8 hash[SHA1_HASH_SIZE];

        /* Length of this stream's name, in bytes and excluding the null
         * terminator; or 0 if this stream is unnamed.  */
        le16 name_nbytes;

        /* Stream name in UTF-16LE.  It is @name_nbytes bytes long, excluding
         * the null terminator.  There is a null terminator character if
         * @name_nbytes != 0; i.e., if this stream is named.  */
        utf16lechar name[];
} _packed_attribute;

static void
do_dentry_set_name(struct wim_dentry *dentry, utf16lechar *file_name,
                   size_t file_name_nbytes)
{
        FREE(dentry->file_name);
        dentry->file_name = file_name;
        dentry->file_name_nbytes = file_name_nbytes;

        if (dentry_has_short_name(dentry)) {
                FREE(dentry->short_name);
                dentry->short_name = NULL;
                dentry->short_name_nbytes = 0;
        }
}

/*
 * Set the name of a WIM dentry from a UTF-16LE string.
 *
 * This sets the long name of the dentry.  The short name will automatically be
 * removed, since it may not be appropriate for the new long name.
 *
 * The @name string need not be null-terminated, since its length is specified
 * in @name_nbytes.
 *
 * If @name_nbytes is 0, both the long and short names of the dentry will be
 * removed.
 *
 * Only use this function on unlinked dentries, since it doesn't update the name
 * indices.  For dentries that are currently linked into the tree, use
 * rename_wim_path().
 *
 * Returns 0 or WIMLIB_ERR_NOMEM.
 */
int
dentry_set_name_utf16le(struct wim_dentry *dentry, const utf16lechar *name,
                        size_t name_nbytes)
{
        utf16lechar *dup = NULL;

        if (name_nbytes) {
                dup = utf16le_dupz(name, name_nbytes);
                if (!dup)
                        return WIMLIB_ERR_NOMEM;
        }
        do_dentry_set_name(dentry, dup, name_nbytes);
        return 0;
}


/*
 * Set the name of a WIM dentry from a 'tchar' string.
 *
 * This sets the long name of the dentry.  The short name will automatically be
 * removed, since it may not be appropriate for the new long name.
 *
 * If @name is NULL or empty, both the long and short names of the dentry will
 * be removed.
 *
 * Only use this function on unlinked dentries, since it doesn't update the name
 * indices.  For dentries that are currently linked into the tree, use
 * rename_wim_path().
 *
 * Returns 0 or an error code resulting from a failed string conversion.
 */
int
dentry_set_name(struct wim_dentry *dentry, const tchar *name)
{
        utf16lechar *name_utf16le = NULL;
        size_t name_utf16le_nbytes = 0;
        int ret;

        if (name && *name) {
                ret = tstr_to_utf16le(name, tstrlen(name) * sizeof(tchar),
                                      &name_utf16le, &name_utf16le_nbytes);
                if (ret)
                        return ret;
        }

        do_dentry_set_name(dentry, name_utf16le, name_utf16le_nbytes);
        return 0;
}

/* Calculate the minimum unaligned length, in bytes, of an on-disk WIM dentry
 * that has names of the specified lengths.  (Zero length means the
 * corresponding name actually does not exist.)  The returned value excludes
 * tagged metadata items as well as any extra stream entries that may need to
 * follow the dentry.  */
static size_t
dentry_min_len_with_names(u16 file_name_nbytes, u16 short_name_nbytes)
{
        size_t length = sizeof(struct wim_dentry_on_disk);
        if (file_name_nbytes)
                length += (u32)file_name_nbytes + 2;
        if (short_name_nbytes)
                length += (u32)short_name_nbytes + 2;
        return length;
}


/* Return the length, in bytes, required for the specified stream on-disk, when
 * represented as an extra stream entry.  */
static size_t
stream_out_total_length(const struct wim_inode_stream *strm)
{
        /* Account for the fixed length portion  */
        size_t len = sizeof(struct wim_extra_stream_entry_on_disk);

        /* For named streams, account for the variable-length name.  */
        if (stream_is_named(strm))
                len += utf16le_len_bytes(strm->stream_name) + 2;

        /* Account for any necessary padding to the next 8-byte boundary.  */
        return (len + 7) & ~7;
}

/*
 * Calculate the total number of bytes that will be consumed when a dentry is
 * written.  This includes the fixed-length portion of the dentry, the name
 * fields, any tagged metadata items, and any extra stream entries.  This also
 * includes all alignment bytes.
 */
size_t
dentry_out_total_length(const struct wim_dentry *dentry)
{
        const struct wim_inode *inode = dentry->d_inode;
        size_t len;

        len = dentry_min_len_with_names(dentry->file_name_nbytes,
                                        dentry->short_name_nbytes);
        len = (len + 7) & ~7;

        if (inode->i_extra_size) {
                len += inode->i_extra_size;
                len = (len + 7) & ~7;
        }

        if (!(inode->i_attributes & FILE_ATTRIBUTE_ENCRYPTED)) {
                /*
                 * Extra stream entries:
                 *
                 * - Use one extra stream entry for each named data stream
                 * - Use one extra stream entry for the unnamed data stream when there is either:
                 *      - a reparse point stream
                 *      - at least one named data stream (for Windows PE bug workaround)
                 * - Use one extra stream entry for the reparse point stream if there is one
                 */
                bool have_named_data_stream = false;
                bool have_reparse_point_stream = false;
                for (unsigned i = 0; i < inode->i_num_streams; i++) {
                        const struct wim_inode_stream *strm = &inode->i_streams[i];
                        if (stream_is_named_data_stream(strm)) {
                                len += stream_out_total_length(strm);
                                have_named_data_stream = true;
                        } else if (strm->stream_type == STREAM_TYPE_REPARSE_POINT) {
                                wimlib_assert(inode->i_attributes & FILE_ATTRIBUTE_REPARSE_POINT);
                                have_reparse_point_stream = true;
                        }
                }

                if (have_named_data_stream || have_reparse_point_stream) {
                        if (have_reparse_point_stream)
                                len += (sizeof(struct wim_extra_stream_entry_on_disk) + 7) & ~7;
                        len += (sizeof(struct wim_extra_stream_entry_on_disk) + 7) & ~7;
                }
        }

        return len;
}

/* Internal version of for_dentry_in_tree() that omits the NULL check  */
static int
do_for_dentry_in_tree(struct wim_dentry *dentry,
                      int (*visitor)(struct wim_dentry *, void *), void *arg)
{
        int ret;
        struct wim_dentry *child;

        ret = (*visitor)(dentry, arg);
        if (unlikely(ret))
                return ret;

        for_dentry_child(child, dentry) {
                ret = do_for_dentry_in_tree(child, visitor, arg);
                if (unlikely(ret))
                        return ret;
        }
        return 0;
}

/* Internal version of for_dentry_in_tree_depth() that omits the NULL check  */
static int
do_for_dentry_in_tree_depth(struct wim_dentry *dentry,
                            int (*visitor)(struct wim_dentry *, void *), void *arg)
{
        int ret;
        struct wim_dentry *child;

        for_dentry_child_postorder(child, dentry) {
                ret = do_for_dentry_in_tree_depth(child, visitor, arg);
                if (unlikely(ret))
                        return ret;
        }
        return unlikely((*visitor)(dentry, arg));
}

/*
 * Call a function on all dentries in a tree.
 *
 * @arg will be passed as the second argument to each invocation of @visitor.
 *
 * This function does a pre-order traversal --- that is, a parent will be
 * visited before its children.  It also will visit siblings in order of
 * case-sensitive filename.  Equivalently, this function visits the entire tree
 * in the case-sensitive lexicographic order of the full paths.
 *
 * It is safe to pass NULL for @root, which means that the dentry tree is empty.
 * In this case, this function does nothing.
 *
 * @visitor must not modify the structure of the dentry tree during the
 * traversal.
 *
 * The return value will be 0 if all calls to @visitor returned 0.  Otherwise,
 * the return value will be the first nonzero value returned by @visitor.
 */
int
for_dentry_in_tree(struct wim_dentry *root,
                   int (*visitor)(struct wim_dentry *, void *), void *arg)
{
        if (unlikely(!root))
                return 0;
        return do_for_dentry_in_tree(root, visitor, arg);
}

/* Like for_dentry_in_tree(), but do a depth-first traversal of the dentry tree.
 * That is, the visitor function will be called on a dentry's children before
 * itself.  It will be safe to free a dentry when visiting it.  */
static int
for_dentry_in_tree_depth(struct wim_dentry *root,
                         int (*visitor)(struct wim_dentry *, void *), void *arg)
{
        if (unlikely(!root))
                return 0;
        return do_for_dentry_in_tree_depth(root, visitor, arg);
}

/*
 * Calculate the full path to @dentry within the WIM image, if not already done.
 *
 * The full name will be saved in the cached value 'dentry->_full_path'.
 *
 * Whenever possible, use dentry_full_path() instead of calling this and
 * accessing _full_path directly.
 *
 * Returns 0 or an error code resulting from a failed string conversion.
 */
int
calculate_dentry_full_path(struct wim_dentry *dentry)
{
        size_t ulen;
        size_t dummy;
        const struct wim_dentry *d;

        if (dentry->_full_path)
                return 0;

        ulen = 0;
        d = dentry;
        do {
                ulen += d->file_name_nbytes / sizeof(utf16lechar);
                ulen++;
                d = d->d_parent;  /* assumes d == d->d_parent for root  */
        } while (!dentry_is_root(d));

        utf16lechar ubuf[ulen];
        utf16lechar *p = &ubuf[ulen];

        d = dentry;
        do {
                p -= d->file_name_nbytes / sizeof(utf16lechar);
                memcpy(p, d->file_name, d->file_name_nbytes);
                *--p = cpu_to_le16(WIM_PATH_SEPARATOR);
                d = d->d_parent;  /* assumes d == d->d_parent for root  */
        } while (!dentry_is_root(d));

        wimlib_assert(p == ubuf);

        return utf16le_to_tstr(ubuf, ulen * sizeof(utf16lechar),
                               &dentry->_full_path, &dummy);
}

/*
 * Return the full path to the @dentry within the WIM image, or NULL if the full
 * path could not be determined due to a string conversion error.
 *
 * The returned memory will be cached in the dentry, so the caller is not
 * responsible for freeing it.
 */
tchar *
dentry_full_path(struct wim_dentry *dentry)
{
        calculate_dentry_full_path(dentry);
        return dentry->_full_path;
}

static int
dentry_calculate_subdir_offset(struct wim_dentry *dentry, void *_subdir_offset_p)
{
        if (dentry_is_directory(dentry)) {
                u64 *subdir_offset_p = _subdir_offset_p;
                struct wim_dentry *child;

                /* Set offset of directory's child dentries  */
                dentry->subdir_offset = *subdir_offset_p;

                /* Account for child dentries  */
                for_dentry_child(child, dentry)
                        *subdir_offset_p += dentry_out_total_length(child);

                /* Account for end-of-directory entry  */
                *subdir_offset_p += 8;
        } else {
                /* Not a directory; set subdir_offset to 0  */
                dentry->subdir_offset = 0;
        }
        return 0;
}

/*
 * Calculate the subdir offsets for a dentry tree, in preparation of writing
 * that dentry tree to a metadata resource.
 *
 * The subdir offset of each dentry is the offset in the uncompressed metadata
 * resource at which its child dentries begin, or 0 if that dentry has no
 * children.
 *
 * The caller must initialize *subdir_offset_p to the first subdir offset that
 * is available to use after the root dentry is written.
 *
 * When this function returns, *subdir_offset_p will have been advanced past the
 * size needed for the dentry tree within the uncompressed metadata resource.
 */
void
calculate_subdir_offsets(struct wim_dentry *root, u64 *subdir_offset_p)
{
        for_dentry_in_tree(root, dentry_calculate_subdir_offset, subdir_offset_p);
}

/* Compare the UTF-16LE long filenames of two dentries case insensitively.  */
static int
dentry_compare_names_case_insensitive(const struct wim_dentry *d1,
                                      const struct wim_dentry *d2)
{
        return cmp_utf16le_strings(d1->file_name,
                                   d1->file_name_nbytes / 2,
                                   d2->file_name,
                                   d2->file_name_nbytes / 2,
                                   true);
}

/* Compare the UTF-16LE long filenames of two dentries case sensitively.  */
static int
dentry_compare_names_case_sensitive(const struct wim_dentry *d1,
                                    const struct wim_dentry *d2)
{
        return cmp_utf16le_strings(d1->file_name,
                                   d1->file_name_nbytes / 2,
                                   d2->file_name,
                                   d2->file_name_nbytes / 2,
                                   false);
}

static int
_avl_dentry_compare_names_ci(const struct avl_tree_node *n1,
                             const struct avl_tree_node *n2)
{
        const struct wim_dentry *d1, *d2;

        d1 = avl_tree_entry(n1, struct wim_dentry, d_index_node_ci);
        d2 = avl_tree_entry(n2, struct wim_dentry, d_index_node_ci);
        return dentry_compare_names_case_insensitive(d1, d2);
}

static int
_avl_dentry_compare_names(const struct avl_tree_node *n1,
                          const struct avl_tree_node *n2)
{
        const struct wim_dentry *d1, *d2;

        d1 = avl_tree_entry(n1, struct wim_dentry, d_index_node);
        d2 = avl_tree_entry(n2, struct wim_dentry, d_index_node);
        return dentry_compare_names_case_sensitive(d1, d2);
}

/* Default case sensitivity behavior for searches with
 * WIMLIB_CASE_PLATFORM_DEFAULT specified.  This can be modified by passing
 * WIMLIB_INIT_FLAG_DEFAULT_CASE_SENSITIVE or
 * WIMLIB_INIT_FLAG_DEFAULT_CASE_INSENSITIVE to wimlib_global_init().  */
bool default_ignore_case =
#ifdef __WIN32__
        true
#else
        false
#endif
;

/* Case-sensitive dentry lookup.  Only @file_name and @file_name_nbytes of
 * @dummy must be valid.  */
static struct wim_dentry *
dir_lookup(const struct wim_inode *dir, const struct wim_dentry *dummy)
{
        struct avl_tree_node *node;

        node = avl_tree_lookup_node(dir->i_children,
                                    &dummy->d_index_node,
                                    _avl_dentry_compare_names);
        if (!node)
                return NULL;
        return avl_tree_entry(node, struct wim_dentry, d_index_node);
}

/* Case-insensitive dentry lookup.  Only @file_name and @file_name_nbytes of
 * @dummy must be valid.  */
static struct wim_dentry *
dir_lookup_ci(const struct wim_inode *dir, const struct wim_dentry *dummy)
{
        struct avl_tree_node *node;

        node = avl_tree_lookup_node(dir->i_children_ci,
                                    &dummy->d_index_node_ci,
                                    _avl_dentry_compare_names_ci);
        if (!node)
                return NULL;
        return avl_tree_entry(node, struct wim_dentry, d_index_node_ci);
}

/* Given a UTF-16LE filename and a directory, look up the dentry for the file.
 * Return it if found, otherwise NULL.  This has configurable case sensitivity,
 * and @name need not be null-terminated.  */
struct wim_dentry *
get_dentry_child_with_utf16le_name(const struct wim_dentry *dentry,
                                   const utf16lechar *name,
                                   size_t name_nbytes,
                                   CASE_SENSITIVITY_TYPE case_ctype)
{
        const struct wim_inode *dir = dentry->d_inode;
        bool ignore_case = will_ignore_case(case_ctype);
        struct wim_dentry dummy;
        struct wim_dentry *child;

        dummy.file_name = (utf16lechar*)name;
        dummy.file_name_nbytes = name_nbytes;

        if (!ignore_case)
                /* Case-sensitive lookup.  */
                return dir_lookup(dir, &dummy);

        /* Case-insensitive lookup.  */

        child = dir_lookup_ci(dir, &dummy);
        if (!child)
                return NULL;

        if (likely(list_empty(&child->d_ci_conflict_list)))
                /* Only one dentry has this case-insensitive name; return it */
                return child;

        /* Multiple dentries have the same case-insensitive name.  Choose the
         * dentry with the same case-sensitive name, if one exists; otherwise
         * print a warning and choose one of the possible dentries arbitrarily.
         */
        struct wim_dentry *alt = child;
        size_t num_alts = 0;

        do {
                num_alts++;
                if (!dentry_compare_names_case_sensitive(&dummy, alt))
                        return alt;
                alt = list_entry(alt->d_ci_conflict_list.next,
                                 struct wim_dentry, d_ci_conflict_list);
        } while (alt != child);

        WARNING("Result of case-insensitive lookup is ambiguous\n"
                "          (returning \"%"TS"\" of %zu "
                "possible files, including \"%"TS"\")",
                dentry_full_path(child),
                num_alts,
                dentry_full_path(list_entry(child->d_ci_conflict_list.next,
                                            struct wim_dentry,
                                            d_ci_conflict_list)));
        return child;
}

/* Given a 'tchar' filename and a directory, look up the dentry for the file.
 * If the filename was successfully converted to UTF-16LE and the dentry was
 * found, return it; otherwise return NULL.  This has configurable case
 * sensitivity.  */
struct wim_dentry *
get_dentry_child_with_name(const struct wim_dentry *dentry, const tchar *name,
                           CASE_SENSITIVITY_TYPE case_type)
{
        int ret;
        const utf16lechar *name_utf16le;
        size_t name_utf16le_nbytes;
        struct wim_dentry *child;

        ret = tstr_get_utf16le_and_len(name, &name_utf16le,
                                       &name_utf16le_nbytes);
        if (ret)
                return NULL;

        child = get_dentry_child_with_utf16le_name(dentry,
                                                   name_utf16le,
                                                   name_utf16le_nbytes,
                                                   case_type);
        tstr_put_utf16le(name_utf16le);
        return child;
}

/* This is the UTF-16LE version of get_dentry(), currently private to this file
 * because no one needs it besides get_dentry().  */
static struct wim_dentry *
get_dentry_utf16le(WIMStruct *wim, const utf16lechar *path,
                   CASE_SENSITIVITY_TYPE case_type)
{
        struct wim_dentry *cur_dentry;
        const utf16lechar *name_start, *name_end;

        /* Start with the root directory of the image.  Note: this will be NULL
         * if an image has been added directly with wimlib_add_empty_image() but
         * no files have been added yet; in that case we fail with ENOENT.  */
        cur_dentry = wim_get_current_root_dentry(wim);

        name_start = path;
        for (;;) {
                if (cur_dentry == NULL) {
                        errno = ENOENT;
                        return NULL;
                }

                if (*name_start && !dentry_is_directory(cur_dentry)) {
                        errno = ENOTDIR;
                        return NULL;
                }

                while (*name_start == cpu_to_le16(WIM_PATH_SEPARATOR))
                        name_start++;

                if (!*name_start)
                        return cur_dentry;

                name_end = name_start;
                do {
                        ++name_end;
                } while (*name_end != cpu_to_le16(WIM_PATH_SEPARATOR) && *name_end);

                cur_dentry = get_dentry_child_with_utf16le_name(cur_dentry,
                                                                name_start,
                                                                (u8*)name_end - (u8*)name_start,
                                                                case_type);
                name_start = name_end;
        }
}

/*
 * WIM path lookup: translate a path in the currently selected WIM image to the
 * corresponding dentry, if it exists.
 *
 * @wim
 *      The WIMStruct for the WIM.  The search takes place in the currently
 *      selected image.
 *
 * @path
 *      The path to look up, given relative to the root of the WIM image.
 *      Characters with value WIM_PATH_SEPARATOR are taken to be path
 *      separators.  Leading path separators are ignored, whereas one or more
 *      trailing path separators cause the path to only match a directory.
 *
 * @case_type
 *      The case-sensitivity behavior of this function, as one of the following
 *      constants:
 *
 *    - WIMLIB_CASE_SENSITIVE:  Perform the search case sensitively.  This means
 *      that names must match exactly.
 *
 *    - WIMLIB_CASE_INSENSITIVE:  Perform the search case insensitively.  This
 *      means that names are considered to match if they are equal when
 *      transformed to upper case.  If a path component matches multiple names
 *      case-insensitively, the name that matches the path component
 *      case-sensitively is chosen, if existent; otherwise one
 *      case-insensitively matching name is chosen arbitrarily.
 *
 *    - WIMLIB_CASE_PLATFORM_DEFAULT:  Perform either case-sensitive or
 *      case-insensitive search, depending on the value of the global variable
 *      default_ignore_case.
 *
 *    In any case, no Unicode normalization is done before comparing strings.
 *
 * Returns a pointer to the dentry that is the result of the lookup, or NULL if
 * no such dentry exists.  If NULL is returned, errno is set to one of the
 * following values:
 *
 *      ENOTDIR if one of the path components used as a directory existed but
 *      was not, in fact, a directory.
 *
 *      ENOENT otherwise.
 *
 * Additional notes:
 *
 *    - This function does not consider a reparse point to be a directory, even
 *      if it has FILE_ATTRIBUTE_DIRECTORY set.
 *
 *    - This function does not dereference symbolic links or junction points
 *      when performing the search.
 *
 *    - Since this function ignores leading slashes, the empty path is valid and
 *      names the root directory of the WIM image.
 *
 *    - An image added with wimlib_add_empty_image() does not have a root
 *      directory yet, and this function will fail with ENOENT for any path on
 *      such an image.
 */
struct wim_dentry *
get_dentry(WIMStruct *wim, const tchar *path, CASE_SENSITIVITY_TYPE case_type)
{
        int ret;
        const utf16lechar *path_utf16le;
        struct wim_dentry *dentry;

        ret = tstr_get_utf16le(path, &path_utf16le);
        if (ret)
                return NULL;
        dentry = get_dentry_utf16le(wim, path_utf16le, case_type);
        tstr_put_utf16le(path_utf16le);
        return dentry;
}

/* Modify @path, which is a null-terminated string @len 'tchars' in length,
 * in-place to produce the path to its parent directory.  */
static void
to_parent_name(tchar *path, size_t len)
{
        ssize_t i = (ssize_t)len - 1;
        while (i >= 0 && path[i] == WIM_PATH_SEPARATOR)
                i--;
        while (i >= 0 && path[i] != WIM_PATH_SEPARATOR)
                i--;
        while (i >= 0 && path[i] == WIM_PATH_SEPARATOR)
                i--;
        path[i + 1] = T('\0');
}

/* Similar to get_dentry(), but returns the dentry named by @path with the last
 * component stripped off.
 *
 * Note: The returned dentry is NOT guaranteed to be a directory.  */
struct wim_dentry *
get_parent_dentry(WIMStruct *wim, const tchar *path,
                  CASE_SENSITIVITY_TYPE case_type)
{
        size_t path_len = tstrlen(path);
        tchar buf[path_len + 1];

        tmemcpy(buf, path, path_len + 1);
        to_parent_name(buf, path_len);
        return get_dentry(wim, buf, case_type);
}

/*
 * Create an unlinked dentry.
 *
 * @name specifies the long name to give the new dentry.  If NULL or empty, the
 * new dentry will be given no long name.
 *
 * The new dentry will have no short name and no associated inode.
 *
 * On success, returns 0 and a pointer to the new, allocated dentry is stored in
 * *dentry_ret.  On failure, returns WIMLIB_ERR_NOMEM or an error code resulting
 * from a failed string conversion.
 */
static int
new_dentry(const tchar *name, struct wim_dentry **dentry_ret)
{
        struct wim_dentry *dentry;
        int ret;

        dentry = CALLOC(1, sizeof(struct wim_dentry));
        if (!dentry)
                return WIMLIB_ERR_NOMEM;

        if (name && *name) {
                ret = dentry_set_name(dentry, name);
                if (ret) {
                        FREE(dentry);
                        return ret;
                }
        }
        dentry->d_parent = dentry;
        *dentry_ret = dentry;
        return 0;
}

/* Like new_dentry(), but also allocate an inode and associate it with the
 * dentry.  If set_timestamps=true, the timestamps for the inode will be set to
 * the current time; otherwise, they will be left 0.  */
int
new_dentry_with_new_inode(const tchar *name, bool set_timestamps,
                          struct wim_dentry **dentry_ret)
{
        struct wim_dentry *dentry;
        struct wim_inode *inode;
        int ret;

        ret = new_dentry(name, &dentry);
        if (ret)
                return ret;

        inode = new_inode(dentry, set_timestamps);
        if (!inode) {
                free_dentry(dentry);
                return WIMLIB_ERR_NOMEM;
        }

        *dentry_ret = dentry;
        return 0;
}

/* Like new_dentry(), but also associate the new dentry with the specified inode
 * and acquire a reference to each of the inode's blobs.  */
int
new_dentry_with_existing_inode(const tchar *name, struct wim_inode *inode,
                               struct wim_dentry **dentry_ret)
{
        int ret = new_dentry(name, dentry_ret);
        if (ret)
                return ret;
        d_associate(*dentry_ret, inode);
        inode_ref_blobs(inode);
        return 0;
}

/* Create an unnamed dentry with a new inode for a directory with the default
 * metadata.  */
int
new_filler_directory(struct wim_dentry **dentry_ret)
{
        int ret;
        struct wim_dentry *dentry;

        ret = new_dentry_with_new_inode(NULL, true, &dentry);
        if (ret)
                return ret;
        /* Leave the inode number as 0; this is allowed for non
         * hard-linked files. */
        dentry->d_inode->i_attributes = FILE_ATTRIBUTE_DIRECTORY;
        *dentry_ret = dentry;
        return 0;
}

static int
dentry_clear_inode_visited(struct wim_dentry *dentry, void *_ignore)
{
        dentry->d_inode->i_visited = 0;
        return 0;
}

void
dentry_tree_clear_inode_visited(struct wim_dentry *root)
{
        for_dentry_in_tree(root, dentry_clear_inode_visited, NULL);
}

/*
 * Free a WIM dentry.
 *
 * In addition to freeing the dentry itself, this disassociates the dentry from
 * its inode.  If the inode is no longer in use, it will be freed as well.
 */
void
free_dentry(struct wim_dentry *dentry)
{
        if (dentry) {
                d_disassociate(dentry);
                FREE(dentry->file_name);
                FREE(dentry->short_name);
                FREE(dentry->_full_path);
                FREE(dentry);
        }
}

static int
do_free_dentry(struct wim_dentry *dentry, void *_ignore)
{
        free_dentry(dentry);
        return 0;
}

static int
do_free_dentry_and_unref_blobs(struct wim_dentry *dentry, void *blob_table)
{
        inode_unref_blobs(dentry->d_inode, blob_table);
        free_dentry(dentry);
        return 0;
}

/*
 * Free all dentries in a tree.
 *
 * @root:
 *      The root of the dentry tree to free.  If NULL, this function has no
 *      effect.
 *
 * @blob_table:
 *      A pointer to the blob table for the WIM, or NULL if not specified.  If
 *      specified, this function will decrement the reference counts of the
 *      blobs referenced by the dentries.
 *
 * This function also releases references to the corresponding inodes.
 *
 * This function does *not* unlink @root from its parent directory, if it has
 * one.  If @root has a parent, the caller must unlink @root before calling this
 * function.
 */
void
free_dentry_tree(struct wim_dentry *root, struct blob_table *blob_table)
{
        int (*f)(struct wim_dentry *, void *);

        if (blob_table)
                f = do_free_dentry_and_unref_blobs;
        else
                f = do_free_dentry;

        for_dentry_in_tree_depth(root, f, blob_table);
}

/* Insert the @child dentry into the case sensitive index of the @dir directory.
 * Return NULL if successfully inserted, otherwise a pointer to the
 * already-inserted duplicate.  */
static struct wim_dentry *
dir_index_child(struct wim_inode *dir, struct wim_dentry *child)
{
        struct avl_tree_node *duplicate;

        duplicate = avl_tree_insert(&dir->i_children,
                                    &child->d_index_node,
                                    _avl_dentry_compare_names);
        if (!duplicate)
                return NULL;
        return avl_tree_entry(duplicate, struct wim_dentry, d_index_node);
}

/* Insert the @child dentry into the case insensitive index of the @dir
 * directory.  Return NULL if successfully inserted, otherwise a pointer to the
 * already-inserted duplicate.  */
static struct wim_dentry *
dir_index_child_ci(struct wim_inode *dir, struct wim_dentry *child)
{
        struct avl_tree_node *duplicate;

        duplicate = avl_tree_insert(&dir->i_children_ci,
                                    &child->d_index_node_ci,
                                    _avl_dentry_compare_names_ci);
        if (!duplicate)
                return NULL;
        return avl_tree_entry(duplicate, struct wim_dentry, d_index_node_ci);
}

/* Remove the specified dentry from its directory's case-sensitive index.  */
static void
dir_unindex_child(struct wim_inode *dir, struct wim_dentry *child)
{
        avl_tree_remove(&dir->i_children, &child->d_index_node);
}

/* Remove the specified dentry from its directory's case-insensitive index.  */
static void
dir_unindex_child_ci(struct wim_inode *dir, struct wim_dentry *child)
{
        avl_tree_remove(&dir->i_children_ci, &child->d_index_node_ci);
}

/* Return true iff the specified dentry is in its parent directory's
 * case-insensitive index.  */
static bool
dentry_in_ci_index(const struct wim_dentry *dentry)
{
        return !avl_tree_node_is_unlinked(&dentry->d_index_node_ci);
}

/*
 * Link a dentry into the tree.
 *
 * @parent:
 *      The dentry that will be the parent of @child.  It must name a directory.
 *
 * @child:
 *      The dentry to link.  It must be currently unlinked.
 *
 * Returns NULL if successful.  If @parent already contains a dentry with the
 * same case-sensitive name as @child, returns a pointer to this duplicate
 * dentry.
 */
struct wim_dentry *
dentry_add_child(struct wim_dentry *parent, struct wim_dentry *child)
{
        struct wim_dentry *duplicate;
        struct wim_inode *dir;

        wimlib_assert(parent != child);

        dir = parent->d_inode;

        wimlib_assert(inode_is_directory(dir));

        duplicate = dir_index_child(dir, child);
        if (duplicate)
                return duplicate;

        duplicate = dir_index_child_ci(dir, child);
        if (duplicate) {
                list_add(&child->d_ci_conflict_list, &duplicate->d_ci_conflict_list);
                avl_tree_node_set_unlinked(&child->d_index_node_ci);
        } else {
                INIT_LIST_HEAD(&child->d_ci_conflict_list);
        }
        child->d_parent = parent;
        return NULL;
}

/* Unlink a dentry from the tree.  */
void
unlink_dentry(struct wim_dentry *dentry)
{
        struct wim_inode *dir;

        /* Do nothing if the dentry is root or it's already unlinked.  Not
         * actually necessary based on the current callers, but we do the check
         * here to be safe.  */
        if (unlikely(dentry->d_parent == dentry))
                return;

        dir = dentry->d_parent->d_inode;

        dir_unindex_child(dir, dentry);

        if (dentry_in_ci_index(dentry)) {

                dir_unindex_child_ci(dir, dentry);

                if (!list_empty(&dentry->d_ci_conflict_list)) {
                        /* Make a different case-insensitively-the-same dentry
                         * be the "representative" in the search index.  */
                        struct list_head *next;
                        struct wim_dentry *other;
                        struct wim_dentry *existing;

                        next = dentry->d_ci_conflict_list.next;
                        other = list_entry(next, struct wim_dentry, d_ci_conflict_list);
                        existing = dir_index_child_ci(dir, other);
                        wimlib_assert(existing == NULL);
                }
        }
        list_del(&dentry->d_ci_conflict_list);

        /* Not actually necessary, but to be safe don't retain the now-obsolete
         * parent pointer.  */
        dentry->d_parent = dentry;
}

static int
read_extra_data(const u8 *p, const u8 *end, struct wim_inode *inode)
{
        while (((uintptr_t)p & 7) && p < end)
                p++;

        if (unlikely(p < end)) {
                inode->i_extra = memdup(p, end - p);
                if (!inode->i_extra)
                        return WIMLIB_ERR_NOMEM;
                inode->i_extra_size = end - p;
        }
        return 0;
}

/*
 * Set the type of each stream for an encrypted file.
 *
 * All data streams of the encrypted file should have been packed into a single
 * stream in the format provided by ReadEncryptedFileRaw() on Windows.  We
 * assign this stream type STREAM_TYPE_EFSRPC_RAW_DATA.
 *
 * Encrypted files can't have a reparse point stream.  In the on-disk NTFS
 * format they can, but as far as I know the reparse point stream of an
 * encrypted file can't be stored in the WIM format in a way that's compatible
 * with WIMGAPI, nor is there even any way for it to be read or written on
 * Windows when the process does not have access to the file encryption key.
 */
static void
assign_stream_types_encrypted(struct wim_inode *inode)
{
        for (unsigned i = 0; i < inode->i_num_streams; i++) {
                struct wim_inode_stream *strm = &inode->i_streams[i];
                if (!stream_is_named(strm) && !is_zero_hash(strm->_stream_hash))
                {
                        strm->stream_type = STREAM_TYPE_EFSRPC_RAW_DATA;
                        return;
                }
        }
}

/*
 * Set the type of each stream for an unencrypted file.
 *
 * There will be an unnamed data stream, a reparse point stream, or both an
 * unnamed data stream and a reparse point stream.  In addition, there may be
 * named data streams.
 */
static void
assign_stream_types_unencrypted(struct wim_inode *inode)
{
        bool found_reparse_point_stream = false;
        bool found_unnamed_data_stream = false;
        struct wim_inode_stream *unnamed_stream_with_zero_hash = NULL;

        for (unsigned i = 0; i < inode->i_num_streams; i++) {
                struct wim_inode_stream *strm = &inode->i_streams[i];

                if (stream_is_named(strm)) {
                        /* Named data stream  */
                        strm->stream_type = STREAM_TYPE_DATA;
                } else if (!is_zero_hash(strm->_stream_hash)) {
                        if ((inode->i_attributes & FILE_ATTRIBUTE_REPARSE_POINT) &&
                            !found_reparse_point_stream) {
                                found_reparse_point_stream = true;
                                strm->stream_type = STREAM_TYPE_REPARSE_POINT;
                        } else if (!found_unnamed_data_stream) {
                                found_unnamed_data_stream = true;
                                strm->stream_type = STREAM_TYPE_DATA;
                        }
                } else {
                        /* If no stream name is specified and the hash is zero,
                         * then remember this stream for later so that we can
                         * assign it to the unnamed data stream if we don't find
                         * a better candidate.  */
                        unnamed_stream_with_zero_hash = strm;
                }
        }

        if (!found_unnamed_data_stream && unnamed_stream_with_zero_hash != NULL)
                unnamed_stream_with_zero_hash->stream_type = STREAM_TYPE_DATA;
}

/*
 * Read and interpret the collection of streams for the specified inode.
 */
static int
setup_inode_streams(const u8 *p, const u8 *end, struct wim_inode *inode,
                    unsigned num_extra_streams, const u8 *default_hash,
                    u64 *offset_p)
{
        const u8 *orig_p = p;

        inode->i_num_streams = 1 + num_extra_streams;

        if (unlikely(inode->i_num_streams > ARRAY_LEN(inode->i_embedded_streams))) {
                inode->i_streams = CALLOC(inode->i_num_streams,
                                          sizeof(inode->i_streams[0]));
                if (!inode->i_streams)
                        return WIMLIB_ERR_NOMEM;
        }

        /* Use the default hash field for the first stream  */
        inode->i_streams[0].stream_name = (utf16lechar *)NO_STREAM_NAME;
        copy_hash(inode->i_streams[0]._stream_hash, default_hash);
        inode->i_streams[0].stream_type = STREAM_TYPE_UNKNOWN;
        inode->i_streams[0].stream_id = 0;

        /* Read the extra stream entries  */
        for (unsigned i = 1; i < inode->i_num_streams; i++) {
                struct wim_inode_stream *strm;
                const struct wim_extra_stream_entry_on_disk *disk_strm;
                u64 length;
                u16 name_nbytes;

                strm = &inode->i_streams[i];

                strm->stream_id = i;

                /* Do we have at least the size of the fixed-length data we know
                 * need?  */
                if ((end - p) < sizeof(struct wim_extra_stream_entry_on_disk))
                        return WIMLIB_ERR_INVALID_METADATA_RESOURCE;

                disk_strm = (const struct wim_extra_stream_entry_on_disk *)p;

                /* Read the length field  */
                length = le64_to_cpu(disk_strm->length);

                /* 8-byte align the length  */
                length = (length + 7) & ~7;

                /* Make sure the length field is neither so small it doesn't
                 * include all the fixed-length data nor so large it overflows
                 * the metadata resource buffer. */
                if (length < sizeof(struct wim_extra_stream_entry_on_disk) ||
                    length > (end - p))
                        return WIMLIB_ERR_INVALID_METADATA_RESOURCE;

                /* Read the rest of the fixed-length data. */

                copy_hash(strm->_stream_hash, disk_strm->hash);
                name_nbytes = le16_to_cpu(disk_strm->name_nbytes);

                /* If stream_name_nbytes != 0, the stream is named.  */
                if (name_nbytes != 0) {
                        /* The name is encoded in UTF16-LE, which uses 2-byte
                         * coding units, so the length of the name had better be
                         * an even number of bytes.  */
                        if (name_nbytes & 1)
                                return WIMLIB_ERR_INVALID_METADATA_RESOURCE;

                        /* Add the length of the stream name to get the length
                         * we actually need to read.  Make sure this isn't more
                         * than the specified length of the entry.  */
                        if (sizeof(struct wim_extra_stream_entry_on_disk) +
                            name_nbytes > length)
                                return WIMLIB_ERR_INVALID_METADATA_RESOURCE;

                        strm->stream_name = utf16le_dupz(disk_strm->name,
                                                         name_nbytes);
                        if (!strm->stream_name)
                                return WIMLIB_ERR_NOMEM;
                } else {
                        strm->stream_name = (utf16lechar *)NO_STREAM_NAME;
                }

                strm->stream_type = STREAM_TYPE_UNKNOWN;

                p += length;
        }

        inode->i_next_stream_id = inode->i_num_streams;

        /* Now, assign a type to each stream.  Unfortunately this requires
         * various hacks because stream types aren't explicitly provided in the
         * WIM on-disk format.  */

        if (unlikely(inode->i_attributes & FILE_ATTRIBUTE_ENCRYPTED))
                assign_stream_types_encrypted(inode);
        else
                assign_stream_types_unencrypted(inode);

        *offset_p += p - orig_p;
        return 0;
}

/* Read a dentry, including all extra stream entries that follow it, from an
 * uncompressed metadata resource buffer.  */
static int
read_dentry(const u8 * restrict buf, size_t buf_len,
            u64 *offset_p, struct wim_dentry **dentry_ret)
{
        u64 offset = *offset_p;
        u64 length;
        const u8 *p;
        const struct wim_dentry_on_disk *disk_dentry;
        struct wim_dentry *dentry;
        struct wim_inode *inode;
        u16 short_name_nbytes;
        u16 file_name_nbytes;
        u64 calculated_size;
        int ret;

        BUILD_BUG_ON(sizeof(struct wim_dentry_on_disk) != WIM_DENTRY_DISK_SIZE);

        /* Before reading the whole dentry, we need to read just the length.
         * This is because a dentry of length 8 (that is, just the length field)
         * terminates the list of sibling directory entries. */

        /* Check for buffer overrun.  */
        if (unlikely(offset + sizeof(u64) > buf_len ||
                     offset + sizeof(u64) < offset))
                return WIMLIB_ERR_INVALID_METADATA_RESOURCE;

        /* Get pointer to the dentry data.  */
        p = &buf[offset];
        disk_dentry = (const struct wim_dentry_on_disk*)p;

        /* Get dentry length.  */
        length = (le64_to_cpu(disk_dentry->length) + 7) & ~7;

        /* Check for end-of-directory.  */
        if (length <= 8) {
                *dentry_ret = NULL;
                return 0;
        }

        /* Validate dentry length.  */
        if (unlikely(length < sizeof(struct wim_dentry_on_disk)))
                return WIMLIB_ERR_INVALID_METADATA_RESOURCE;

        /* Check for buffer overrun.  */
        if (unlikely(offset + length > buf_len ||
                     offset + length < offset))
                return WIMLIB_ERR_INVALID_METADATA_RESOURCE;

        /* Allocate new dentry structure, along with a preliminary inode.  */
        ret = new_dentry_with_new_inode(NULL, false, &dentry);
        if (ret)
                return ret;

        inode = dentry->d_inode;

        /* Read more fields: some into the dentry, and some into the inode.  */
        inode->i_attributes = le32_to_cpu(disk_dentry->attributes);
        inode->i_security_id = le32_to_cpu(disk_dentry->security_id);
        dentry->subdir_offset = le64_to_cpu(disk_dentry->subdir_offset);
        inode->i_creation_time = le64_to_cpu(disk_dentry->creation_time);
        inode->i_last_access_time = le64_to_cpu(disk_dentry->last_access_time);
        inode->i_last_write_time = le64_to_cpu(disk_dentry->last_write_time);

        /* I don't know what's going on here.  It seems like M$ screwed up the
         * reparse points, then put the fields in the same place and didn't
         * document it.  So we have some fields we read for reparse points, and
         * some fields in the same place for non-reparse-points.  */
        if (inode->i_attributes & FILE_ATTRIBUTE_REPARSE_POINT) {
                inode->i_rp_unknown_1 = le32_to_cpu(disk_dentry->reparse.rp_unknown_1);
                inode->i_reparse_tag = le32_to_cpu(disk_dentry->reparse.reparse_tag);
                inode->i_rp_unknown_2 = le16_to_cpu(disk_dentry->reparse.rp_unknown_2);
                inode->i_not_rpfixed = le16_to_cpu(disk_dentry->reparse.not_rpfixed);
                /* Leave inode->i_ino at 0.  Note: this means that WIM cannot
                 * represent multiple hard links to a reparse point file.  */
        } else {
                inode->i_rp_unknown_1 = le32_to_cpu(disk_dentry->nonreparse.rp_unknown_1);
                inode->i_ino = le64_to_cpu(disk_dentry->nonreparse.hard_link_group_id);
        }

        /* Now onto reading the names.  There are two of them: the (long) file
         * name, and the short name.  */

        short_name_nbytes = le16_to_cpu(disk_dentry->short_name_nbytes);
        file_name_nbytes = le16_to_cpu(disk_dentry->file_name_nbytes);

        if (unlikely((short_name_nbytes & 1) | (file_name_nbytes & 1))) {
                ret = WIMLIB_ERR_INVALID_METADATA_RESOURCE;
                goto err_free_dentry;
        }

        /* We now know the length of the file name and short name.  Make sure
         * the length of the dentry is large enough to actually hold them.  */
        calculated_size = dentry_min_len_with_names(file_name_nbytes,
                                                    short_name_nbytes);

        if (unlikely(length < calculated_size)) {
                ret = WIMLIB_ERR_INVALID_METADATA_RESOURCE;
                goto err_free_dentry;
        }

        /* Advance p to point past the base dentry, to the first name.  */
        p += sizeof(struct wim_dentry_on_disk);

        /* Read the filename if present.  Note: if the filename is empty, there
         * is no null terminator following it.  */
        if (file_name_nbytes) {
                dentry->file_name = utf16le_dupz(p, file_name_nbytes);
                if (dentry->file_name == NULL) {
                        ret = WIMLIB_ERR_NOMEM;
                        goto err_free_dentry;
                }
                dentry->file_name_nbytes = file_name_nbytes;
                p += (u32)file_name_nbytes + 2;
        }

        /* Read the short filename if present.  Note: if there is no short
         * filename, there is no null terminator following it. */
        if (short_name_nbytes) {
                dentry->short_name = utf16le_dupz(p, short_name_nbytes);
                if (dentry->short_name == NULL) {
                        ret = WIMLIB_ERR_NOMEM;
                        goto err_free_dentry;
                }
                dentry->short_name_nbytes = short_name_nbytes;
                p += (u32)short_name_nbytes + 2;
        }

        /* Read extra data at end of dentry (but before extra stream entries).
         * This may contain tagged metadata items.  */
        ret = read_extra_data(p, &buf[offset + length], inode);
        if (ret)
                goto err_free_dentry;

        offset += length;

        /* Set up the inode's collection of streams.  */
        ret = setup_inode_streams(&buf[offset],
                                  &buf[buf_len],
                                  inode,
                                  le16_to_cpu(disk_dentry->num_extra_streams),
                                  disk_dentry->default_hash,
                                  &offset);
        if (ret)
                goto err_free_dentry;

        *offset_p = offset;  /* Sets offset of next dentry in directory  */
        *dentry_ret = dentry;
        return 0;

err_free_dentry:
        free_dentry(dentry);
        return ret;
}

/* Is the dentry named "." or ".." ?  */
static bool
dentry_is_dot_or_dotdot(const struct wim_dentry *dentry)
{
        if (dentry->file_name_nbytes <= 4) {
                if (dentry->file_name_nbytes == 4) {
                        if (dentry->file_name[0] == cpu_to_le16('.') &&
                            dentry->file_name[1] == cpu_to_le16('.'))
                                return true;
                } else if (dentry->file_name_nbytes == 2) {
                        if (dentry->file_name[0] == cpu_to_le16('.'))
                                return true;
                }
        }
        return false;
}

static int
read_dentry_tree_recursive(const u8 * restrict buf, size_t buf_len,
                           struct wim_dentry * restrict dir)
{
        u64 cur_offset = dir->subdir_offset;

        /* Check for cyclic directory structure, which would cause infinite
         * recursion if not handled.  */
        for (struct wim_dentry *d = dir->d_parent;
             !dentry_is_root(d); d = d->d_parent)
        {
                if (unlikely(d->subdir_offset == cur_offset)) {
                        ERROR("Cyclic directory structure detected: children "
                              "of \"%"TS"\" coincide with children of \"%"TS"\"",
                              dentry_full_path(dir), dentry_full_path(d));
                        return WIMLIB_ERR_INVALID_METADATA_RESOURCE;
                }
        }

        for (;;) {
                struct wim_dentry *child;
                struct wim_dentry *duplicate;
                int ret;

                /* Read next child of @dir.  */
                ret = read_dentry(buf, buf_len, &cur_offset, &child);
                if (ret)
                        return ret;

                /* Check for end of directory.  */
                if (child == NULL)
                        return 0;

                /* All dentries except the root should be named.  */
                if (unlikely(!dentry_has_long_name(child))) {
                        WARNING("Ignoring unnamed dentry in "
                                "directory \"%"TS"\"", dentry_full_path(dir));
                        free_dentry(child);
                        continue;
                }

                /* Don't allow files named "." or "..".  */
                if (unlikely(dentry_is_dot_or_dotdot(child))) {
                        WARNING("Ignoring file named \".\" or \"..\"; "
                                "potentially malicious archive!!!");
                        free_dentry(child);
                        continue;
                }

                /* Link the child into the directory.  */
                duplicate = dentry_add_child(dir, child);
                if (unlikely(duplicate)) {
                        /* We already found a dentry with this same
                         * case-sensitive long name.  Only keep the first one.
                         */
                        WARNING("Ignoring duplicate file \"%"TS"\" "
                                "(the WIM image already contains a file "
                                "at that path with the exact same name)",
                                dentry_full_path(duplicate));
                        free_dentry(child);
                        continue;
                }

                /* If this child is a directory that itself has children, call
                 * this procedure recursively.  */
                if (child->subdir_offset != 0) {
                        if (likely(dentry_is_directory(child))) {
                                ret = read_dentry_tree_recursive(buf,
                                                                 buf_len,
                                                                 child);
                                if (ret)
                                        return ret;
                        } else {
                                WARNING("Ignoring children of "
                                        "non-directory file \"%"TS"\"",
                                        dentry_full_path(child));
                        }
                }
        }
}

/*
 * Read a tree of dentries from a WIM metadata resource.
 *
 * @buf:
 *      Buffer containing an uncompressed WIM metadata resource.
 *
 * @buf_len:
 *      Length of the uncompressed metadata resource, in bytes.
 *
 * @root_offset
 *      Offset in the metadata resource of the root of the dentry tree.
 *
 * @root_ret:
 *      On success, either NULL or a pointer to the root dentry is written to
 *      this location.  The former case only occurs in the unexpected case that
 *      the tree began with an end-of-directory entry.
 *
 * Return values:
 *      WIMLIB_ERR_SUCCESS (0)
 *      WIMLIB_ERR_INVALID_METADATA_RESOURCE
 *      WIMLIB_ERR_NOMEM
 */
int
read_dentry_tree(const u8 *buf, size_t buf_len,
                 u64 root_offset, struct wim_dentry **root_ret)
{
        int ret;
        struct wim_dentry *root;

        DEBUG("Reading dentry tree (root_offset=%"PRIu64")", root_offset);

        ret = read_dentry(buf, buf_len, &root_offset, &root);
        if (ret)
                return ret;

        if (likely(root != NULL)) {
                if (unlikely(dentry_has_long_name(root) ||
                             dentry_has_short_name(root)))
                {
                        WARNING("The root directory has a nonempty name; "
                                "removing it.");
                        dentry_set_name(root, NULL);
                }

                if (unlikely(!dentry_is_directory(root))) {
                        ERROR("The root of the WIM image is not a directory!");
                        ret = WIMLIB_ERR_INVALID_METADATA_RESOURCE;
                        goto err_free_dentry_tree;
                }

                if (likely(root->subdir_offset != 0)) {
                        ret = read_dentry_tree_recursive(buf, buf_len, root);
                        if (ret)
                                goto err_free_dentry_tree;
                }
        } else {
                WARNING("The metadata resource has no directory entries; "
                        "treating as an empty image.");
        }
        *root_ret = root;
        return 0;

err_free_dentry_tree:
        free_dentry_tree(root, NULL);
        return ret;
}

static u8 *
write_extra_stream_entry(u8 * restrict p, const utf16lechar * restrict name,
                         const u8 * restrict hash)
{
        struct wim_extra_stream_entry_on_disk *disk_strm =
                        (struct wim_extra_stream_entry_on_disk *)p;
        u8 *orig_p = p;
        size_t name_nbytes;

        if (name == NO_STREAM_NAME)
                name_nbytes = 0;
        else
                name_nbytes = utf16le_len_bytes(name);

        disk_strm->reserved = 0;
        copy_hash(disk_strm->hash, hash);
        disk_strm->name_nbytes = cpu_to_le16(name_nbytes);
        p += sizeof(struct wim_extra_stream_entry_on_disk);
        if (name_nbytes != 0)
                p = mempcpy(p, name, name_nbytes + 2);
        /* Align to 8-byte boundary */
        while ((uintptr_t)p & 7)
                *p++ = 0;
        disk_strm->length = cpu_to_le64(p - orig_p);
        return p;
}

/*
 * Write a WIM dentry to an output buffer.
 *
 * This includes any extra stream entries that may follow the dentry itself.
 *
 * @dentry:
 *      The dentry to write.
 *
 * @p:
 *      The memory location to which to write the data.
 *
 * Returns a pointer to the byte following the last written.
 */
static u8 *
write_dentry(const struct wim_dentry * restrict dentry, u8 * restrict p)
{
        const struct wim_inode *inode;
        struct wim_dentry_on_disk *disk_dentry;
        const u8 *orig_p;

        wimlib_assert(((uintptr_t)p & 7) == 0); /* 8 byte aligned */
        orig_p = p;

        inode = dentry->d_inode;
        disk_dentry = (struct wim_dentry_on_disk*)p;

        disk_dentry->attributes = cpu_to_le32(inode->i_attributes);
        disk_dentry->security_id = cpu_to_le32(inode->i_security_id);
        disk_dentry->subdir_offset = cpu_to_le64(dentry->subdir_offset);

        disk_dentry->unused_1 = cpu_to_le64(0);
        disk_dentry->unused_2 = cpu_to_le64(0);

        disk_dentry->creation_time = cpu_to_le64(inode->i_creation_time);
        disk_dentry->last_access_time = cpu_to_le64(inode->i_last_access_time);
        disk_dentry->last_write_time = cpu_to_le64(inode->i_last_write_time);
        if (inode->i_attributes & FILE_ATTRIBUTE_REPARSE_POINT) {
                disk_dentry->reparse.rp_unknown_1 = cpu_to_le32(inode->i_rp_unknown_1);
                disk_dentry->reparse.reparse_tag = cpu_to_le32(inode->i_reparse_tag);
                disk_dentry->reparse.rp_unknown_2 = cpu_to_le16(inode->i_rp_unknown_2);
                disk_dentry->reparse.not_rpfixed = cpu_to_le16(inode->i_not_rpfixed);
        } else {
                disk_dentry->nonreparse.rp_unknown_1 = cpu_to_le32(inode->i_rp_unknown_1);
                disk_dentry->nonreparse.hard_link_group_id =
                        cpu_to_le64((inode->i_nlink == 1) ? 0 : inode->i_ino);
        }

        disk_dentry->short_name_nbytes = cpu_to_le16(dentry->short_name_nbytes);
        disk_dentry->file_name_nbytes = cpu_to_le16(dentry->file_name_nbytes);
        p += sizeof(struct wim_dentry_on_disk);

        wimlib_assert(dentry_is_root(dentry) != dentry_has_long_name(dentry));

        if (dentry_has_long_name(dentry))
                p = mempcpy(p, dentry->file_name, (u32)dentry->file_name_nbytes + 2);

        if (dentry_has_short_name(dentry))
                p = mempcpy(p, dentry->short_name, (u32)dentry->short_name_nbytes + 2);

        /* Align to 8-byte boundary */
        while ((uintptr_t)p & 7)
                *p++ = 0;

        if (inode->i_extra_size) {
                /* Extra tagged items --- not usually present.  */
                p = mempcpy(p, inode->i_extra, inode->i_extra_size);

                /* Align to 8-byte boundary */
                while ((uintptr_t)p & 7)
                        *p++ = 0;
        }

        disk_dentry->length = cpu_to_le64(p - orig_p);

        /* Streams  */

        if (unlikely(inode->i_attributes & FILE_ATTRIBUTE_ENCRYPTED)) {
                const struct wim_inode_stream *efs_strm;
                const u8 *efs_hash;

                efs_strm = inode_get_stream(inode, STREAM_TYPE_EFSRPC_RAW_DATA,
                                            NO_STREAM_NAME);
                efs_hash = efs_strm ? stream_hash(efs_strm) : zero_hash;
                copy_hash(disk_dentry->default_hash, efs_hash);
                disk_dentry->num_extra_streams = cpu_to_le16(0);
        } else {
                /*
                 * Extra stream entries:
                 *
                 * - Use one extra stream entry for each named data stream
                 * - Use one extra stream entry for the unnamed data stream when there is either:
                 *      - a reparse point stream
                 *      - at least one named data stream (for Windows PE bug workaround)
                 * - Use one extra stream entry for the reparse point stream if there is one
                 */
                bool have_named_data_stream = false;
                bool have_reparse_point_stream = false;
                u16 num_extra_streams = 0;
                const u8 *unnamed_data_stream_hash = zero_hash;
                const u8 *reparse_point_hash;
                for (unsigned i = 0; i < inode->i_num_streams; i++) {
                        const struct wim_inode_stream *strm = &inode->i_streams[i];
                        if (strm->stream_type == STREAM_TYPE_DATA) {
                                if (stream_is_named(strm))
                                        have_named_data_stream = true;
                                else
                                        unnamed_data_stream_hash = stream_hash(strm);
                        } else if (strm->stream_type == STREAM_TYPE_REPARSE_POINT) {
                                have_reparse_point_stream = true;
                                reparse_point_hash = stream_hash(strm);
                        }
                }

                if (have_reparse_point_stream || have_named_data_stream) {

                        copy_hash(disk_dentry->default_hash, zero_hash);

                        if (have_reparse_point_stream) {
                                p = write_extra_stream_entry(p, NO_STREAM_NAME,
                                                             reparse_point_hash);
                                num_extra_streams++;
                        }

                        p = write_extra_stream_entry(p, NO_STREAM_NAME,
                                                     unnamed_data_stream_hash);
                        num_extra_streams++;
                } else {
                        copy_hash(disk_dentry->default_hash, unnamed_data_stream_hash);
                }

                for (unsigned i = 0; i < inode->i_num_streams; i++) {
                        const struct wim_inode_stream *strm = &inode->i_streams[i];
                        if (stream_is_named_data_stream(strm)) {
                                p = write_extra_stream_entry(p, strm->stream_name,
                                                             stream_hash(strm));
                                num_extra_streams++;
                        }
                }
                disk_dentry->num_extra_streams = cpu_to_le16(num_extra_streams);
        }

        return p;
}

static int
write_dir_dentries(struct wim_dentry *dir, void *_pp)
{
        if (dir->subdir_offset != 0) {
                u8 **pp = _pp;
                u8 *p = *pp;
                struct wim_dentry *child;

                /* write child dentries */
                for_dentry_child(child, dir)
                        p = write_dentry(child, p);

                /* write end of directory entry */
                *(u64*)p = 0;
                p += 8;
                *pp = p;
        }
        return 0;
}

/*
 * Write a directory tree to the metadata resource.
 *
 * @root:
 *      The root of a dentry tree on which calculate_subdir_offsets() has been
 *      called.  This cannot be NULL; if the dentry tree is empty, the caller is
 *      expected to first generate a dummy root directory.
 *
 * @p:
 *      Pointer to a buffer with enough space for the dentry tree.  This size
 *      must have been obtained by calculate_subdir_offsets().
 *
 * Returns a pointer to the byte following the last written.
 */
u8 *
write_dentry_tree(struct wim_dentry *root, u8 *p)
{
        DEBUG("Writing dentry tree.");

        wimlib_assert(root != NULL);

        /* write root dentry and end-of-directory entry following it */
        p = write_dentry(root, p);
        *(u64*)p = 0;
        p += 8;

        /* write the rest of the dentry tree */
        for_dentry_in_tree(root, write_dir_dentries, &p);

        return p;
}