Merge experimental LZX compressor

[wimlib] / src / write.c

/*
 * write.c
 *
 * Support for writing WIM files; write a WIM file, overwrite a WIM file, write
 * compressed file resources, etc.
 */

/*
 * Copyright (C) 2012, 2013 Eric Biggers
 *
 * This file is part of wimlib, a library for working with WIM files.
 *
 * wimlib is free software; you can redistribute it and/or modify it under the
 * terms of the GNU General Public License as published by the Free
 * Software Foundation; either version 3 of the License, or (at your option)
 * any later version.
 *
 * wimlib 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 General Public License for more
 * details.
 *
 * You should have received a copy of the GNU General Public License
 * along with wimlib; if not, see http://www.gnu.org/licenses/.
 */

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

#if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
/* On BSD, this should be included before "wimlib/list.h" so that "wimlib/list.h" can
 * overwrite the LIST_HEAD macro. */
#  include <sys/file.h>
#endif

#include "wimlib/endianness.h"
#include "wimlib/error.h"
#include "wimlib/file_io.h"
#include "wimlib/header.h"
#include "wimlib/integrity.h"
#include "wimlib/lookup_table.h"
#include "wimlib/metadata.h"
#include "wimlib/resource.h"
#include "wimlib/write.h"
#include "wimlib/xml.h"

#ifdef __WIN32__
#  include "wimlib/win32.h" /* win32_get_number_of_processors() */
#endif

#ifdef ENABLE_MULTITHREADED_COMPRESSION
#  include <pthread.h>
#endif

#include <errno.h>
#include <fcntl.h>
#include <limits.h>
#include <stdlib.h>
#include <unistd.h>

#ifdef HAVE_ALLOCA_H
#  include <alloca.h>
#endif


#ifndef __WIN32__
#  include <sys/uio.h> /* for `struct iovec' */
#endif

static int
alloc_lzx_context(int write_resource_flags, struct wimlib_lzx_context **ctx_pp)
{
        struct wimlib_lzx_params params;
        params.size_of_this = sizeof(params);
        if (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_COMPRESS_SLOW)
                params.algorithm = WIMLIB_LZX_ALGORITHM_SLOW;
        else
                params.algorithm = WIMLIB_LZX_ALGORITHM_FAST;
        params.use_defaults = 1;
        return wimlib_lzx_alloc_context(&params, ctx_pp);
}

static unsigned
compress_chunk(const void * uncompressed_data,
               unsigned uncompressed_len,
               void *compressed_data,
               int out_ctype,
               struct wimlib_lzx_context *comp_ctx)
{
        switch (out_ctype) {
        case WIMLIB_COMPRESSION_TYPE_XPRESS:
                return wimlib_xpress_compress(uncompressed_data,
                                              uncompressed_len,
                                              compressed_data);
        case WIMLIB_COMPRESSION_TYPE_LZX:
                return wimlib_lzx_compress2(uncompressed_data,
                                            uncompressed_len,
                                            compressed_data,
                                            comp_ctx);
        default:
                wimlib_assert(0);
                return 0;
        }
}

/* Chunk table that's located at the beginning of each compressed resource in
 * the WIM.  (This is not the on-disk format; the on-disk format just has an
 * array of offsets.) */
struct chunk_table {
        u64 original_resource_size;
        u64 num_chunks;
        u64 table_disk_size;
        unsigned bytes_per_chunk_entry;
        void *cur_offset_p;
        union {
                u32 cur_offset_u32;
                u64 cur_offset_u64;
        };
        /* Beginning of chunk offsets, in either 32-bit or 64-bit little endian
         * integers, including the first offset of 0, which will not be written.
         * */
        u8 offsets[] _aligned_attribute(8);
};

/* Allocate and initializes a chunk table, then reserve space for it in the
 * output file unless writing a pipable resource.  */
static int
begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
                             struct filedes *out_fd,
                             struct chunk_table **chunk_tab_ret,
                             int resource_flags)
{
        u64 size;
        u64 num_chunks;
        unsigned bytes_per_chunk_entry;
        size_t alloc_size;
        struct chunk_table *chunk_tab;
        int ret;

        size = wim_resource_size(lte);
        num_chunks = wim_resource_chunks(lte);
        bytes_per_chunk_entry = (size > (1ULL << 32)) ? 8 : 4;
        alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
        chunk_tab = CALLOC(1, alloc_size);

        if (!chunk_tab) {
                ERROR("Failed to allocate chunk table for %"PRIu64" byte "
                      "resource", size);
                return WIMLIB_ERR_NOMEM;
        }
        chunk_tab->num_chunks = num_chunks;
        chunk_tab->original_resource_size = size;
        chunk_tab->bytes_per_chunk_entry = bytes_per_chunk_entry;
        chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
                                     (num_chunks - 1);
        chunk_tab->cur_offset_p = chunk_tab->offsets;

        /* We don't know the correct offsets yet; so just write zeroes to
         * reserve space for the table, so we can go back to it later after
         * we've written the compressed chunks following it.
         *
         * Special case: if writing a pipable WIM, compressed resources are in a
         * modified format (see comment above write_pipable_wim()) and do not
         * have a chunk table at the beginning, so don't reserve any space for
         * one.  */
        if (!(resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE)) {
                ret = full_write(out_fd, chunk_tab->offsets,
                                 chunk_tab->table_disk_size);
                if (ret) {
                        ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
                                         "file resource");
                        FREE(chunk_tab);
                        return ret;
                }
        }
        *chunk_tab_ret = chunk_tab;
        return 0;
}

/* Add the offset for the next chunk to the chunk table being constructed for a
 * compressed stream. */
static void
chunk_tab_record_chunk(struct chunk_table *chunk_tab, unsigned out_chunk_size)
{
        if (chunk_tab->bytes_per_chunk_entry == 4) {
                *(le32*)chunk_tab->cur_offset_p = cpu_to_le32(chunk_tab->cur_offset_u32);
                chunk_tab->cur_offset_p = (le32*)chunk_tab->cur_offset_p + 1;
                chunk_tab->cur_offset_u32 += out_chunk_size;
        } else {
                *(le64*)chunk_tab->cur_offset_p = cpu_to_le64(chunk_tab->cur_offset_u64);
                chunk_tab->cur_offset_p = (le64*)chunk_tab->cur_offset_p + 1;
                chunk_tab->cur_offset_u64 += out_chunk_size;
        }
}

/* Finishes a WIM chunk table and writes it to the output file at the correct
 * offset.  */
static int
finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
                              struct filedes *out_fd,
                              off_t res_start_offset,
                              int write_resource_flags)
{
        int ret;

        if (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
                ret = full_write(out_fd,
                                 chunk_tab->offsets +
                                         chunk_tab->bytes_per_chunk_entry,
                                 chunk_tab->table_disk_size);
        } else {
                ret  = full_pwrite(out_fd,
                                   chunk_tab->offsets +
                                           chunk_tab->bytes_per_chunk_entry,
                                   chunk_tab->table_disk_size,
                                   res_start_offset);
        }
        if (ret) {
                ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
                                 "file resource");
        }
        return ret;
}

/* Write the header for a stream in a pipable WIM.
 */
static int
write_pwm_stream_header(const struct wim_lookup_table_entry *lte,
                        struct filedes *out_fd,
                        int additional_reshdr_flags)
{
        struct pwm_stream_hdr stream_hdr;
        u32 reshdr_flags;
        int ret;

        stream_hdr.magic = PWM_STREAM_MAGIC;
        stream_hdr.uncompressed_size = cpu_to_le64(lte->resource_entry.original_size);
        if (additional_reshdr_flags & PWM_RESHDR_FLAG_UNHASHED) {
                zero_out_hash(stream_hdr.hash);
        } else {
                wimlib_assert(!lte->unhashed);
                copy_hash(stream_hdr.hash, lte->hash);
        }

        reshdr_flags = lte->resource_entry.flags & ~WIM_RESHDR_FLAG_COMPRESSED;
        reshdr_flags |= additional_reshdr_flags;
        stream_hdr.flags = cpu_to_le32(reshdr_flags);
        ret = full_write(out_fd, &stream_hdr, sizeof(stream_hdr));
        if (ret)
                ERROR_WITH_ERRNO("Error writing stream header");
        return ret;
}

static int
seek_and_truncate(struct filedes *out_fd, off_t offset)
{
        if (filedes_seek(out_fd, offset) == -1 ||
            ftruncate(out_fd->fd, offset))
        {
                ERROR_WITH_ERRNO("Failed to truncate output WIM file");
                return WIMLIB_ERR_WRITE;
        }
        return 0;
}

static int
finalize_and_check_sha1(SHA_CTX *sha_ctx, struct wim_lookup_table_entry *lte)
{
        u8 md[SHA1_HASH_SIZE];

        sha1_final(md, sha_ctx);
        if (lte->unhashed) {
                copy_hash(lte->hash, md);
        } else if (!hashes_equal(md, lte->hash)) {
                ERROR("WIM resource has incorrect hash!");
                if (lte_filename_valid(lte)) {
                        ERROR("We were reading it from \"%"TS"\"; maybe "
                              "it changed while we were reading it.",
                              lte->file_on_disk);
                }
                return WIMLIB_ERR_INVALID_RESOURCE_HASH;
        }
        return 0;
}

struct write_resource_ctx {
        int out_ctype;
        struct wimlib_lzx_context *comp_ctx;
        struct chunk_table *chunk_tab;
        struct filedes *out_fd;
        SHA_CTX sha_ctx;
        bool doing_sha;
        int resource_flags;
};

static int
write_resource_cb(const void *chunk, size_t chunk_size, void *_ctx)
{
        struct write_resource_ctx *ctx = _ctx;
        const void *out_chunk;
        unsigned out_chunk_size;
        int ret;

        if (ctx->doing_sha)
                sha1_update(&ctx->sha_ctx, chunk, chunk_size);

        out_chunk = chunk;
        out_chunk_size = chunk_size;
        if (ctx->out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
                void *compressed_chunk;
                unsigned compressed_size;

                /* Compress the chunk.  */
                compressed_chunk = alloca(chunk_size);

                compressed_size = compress_chunk(chunk, chunk_size,
                                                 compressed_chunk,
                                                 ctx->out_ctype,
                                                 ctx->comp_ctx);
                /* Use compressed data if compression to less than input size
                 * was successful.  */
                if (compressed_size) {
                        out_chunk = compressed_chunk;
                        out_chunk_size = compressed_size;
                }
        }

        if (ctx->chunk_tab) {
                /* Update chunk table accounting.  */
                chunk_tab_record_chunk(ctx->chunk_tab, out_chunk_size);

                /* If writing compressed chunks to a pipable WIM, before the
                 * chunk data write a chunk header that provides the compressed
                 * chunk size.  */
                if (ctx->resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
                        struct pwm_chunk_hdr chunk_hdr = {
                                .compressed_size = cpu_to_le32(out_chunk_size),
                        };
                        ret = full_write(ctx->out_fd, &chunk_hdr,
                                         sizeof(chunk_hdr));
                        if (ret)
                                goto error;
                }
        }

        /* Write the chunk data.  */
        ret = full_write(ctx->out_fd, out_chunk, out_chunk_size);
        if (ret)
                goto error;
        return 0;

error:
        ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
        return ret;
}

/*
 * write_wim_resource()-
 *
 * Write a resource to an output WIM.
 *
 * @lte:
 *      Lookup table entry for the resource, which could be in another WIM, in
 *      an external file, or in another location.
 *
 * @out_fd:
 *      File descriptor opened to the output WIM.
 *
 * @out_ctype:
 *      One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate which
 *      compression algorithm to use.
 *
 * @out_res_entry:
 *      On success, this is filled in with the offset, flags, compressed size,
 *      and uncompressed size of the resource in the output WIM.
 *
 * @resource_flags:
 *      * WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed even
 *        if it could otherwise be copied directly from the input;
 *      * WIMLIB_WRITE_RESOURCE_FLAG_COMPRESS_SLOW to compress the data as much
 *        as possible;
 *      * WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE if writing a resource for a pipable WIM
 *        (and the output file descriptor may be a pipe).
 *
 * Additional notes:  The SHA1 message digest of the uncompressed data is
 * calculated (except when doing a raw copy --- see below).  If the @unhashed
 * flag is set on the lookup table entry, this message digest is simply copied
 * to it; otherwise, the message digest is compared with the existing one, and
 * the function will fail if they do not match.
 */
int
write_wim_resource(struct wim_lookup_table_entry *lte,
                   struct filedes *out_fd, int out_ctype,
                   struct resource_entry *out_res_entry,
                   int resource_flags,
                   struct wimlib_lzx_context **comp_ctx)
{
        struct write_resource_ctx write_ctx;
        off_t res_start_offset;
        u64 read_size;
        int ret;

        /* Mask out any irrelevant flags, since this function also uses this
         * variable to store WIMLIB_READ_RESOURCE flags.  */
        resource_flags &= WIMLIB_WRITE_RESOURCE_MASK;

        /* Get current position in output WIM.  */
        res_start_offset = out_fd->offset;

        /* If we are not forcing the data to be recompressed, and the input
         * resource is located in a WIM with the same compression type as that
         * desired other than no compression, we can simply copy the compressed
         * data without recompressing it.  This also means we must skip
         * calculating the SHA1, as we never will see the uncompressed data.  */
        if (lte->resource_location == RESOURCE_IN_WIM &&
            out_ctype == wim_resource_compression_type(lte) &&
            out_ctype != WIMLIB_COMPRESSION_TYPE_NONE &&
            !(resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS))
        {
                /* Normally we can request a RAW_FULL read, but if we're reading
                 * from a pipable resource and writing a non-pipable resource or
                 * vice versa, then a RAW_CHUNKS read needs to be requested so
                 * that the written resource can be appropriately formatted.
                 * However, in neither case is any actual decompression needed.
                 */
                if (lte->is_pipable == !!(resource_flags &
                                          WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE))
                        resource_flags |= WIMLIB_READ_RESOURCE_FLAG_RAW_FULL;
                else
                        resource_flags |= WIMLIB_READ_RESOURCE_FLAG_RAW_CHUNKS;
                write_ctx.doing_sha = false;
                read_size = lte->resource_entry.size;
        } else {
                write_ctx.doing_sha = true;
                sha1_init(&write_ctx.sha_ctx);
                read_size = lte->resource_entry.original_size;
        }


        /* If the output resource is to be compressed, initialize the chunk
         * table and set the function to use for chunk compression.  Exceptions:
         * no compression function is needed if doing a raw copy; also, no chunk
         * table is needed if doing a *full* (not per-chunk) raw copy.  */
        write_ctx.out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
        write_ctx.chunk_tab = NULL;
        if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
                if (!(resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW)) {
                        write_ctx.out_ctype = out_ctype;
                        if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX) {
                                ret = alloc_lzx_context(resource_flags, comp_ctx);
                                if (ret)
                                        goto out;
                        }
                        write_ctx.comp_ctx = *comp_ctx;
                }
                if (!(resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW_FULL)) {
                        ret = begin_wim_resource_chunk_tab(lte, out_fd,
                                                           &write_ctx.chunk_tab,
                                                           resource_flags);
                        if (ret)
                                goto out;
                }
        }

        /* If writing a pipable resource, write the stream header and update
         * @res_start_offset to be the end of the stream header.  */
        if (resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
                int reshdr_flags = 0;
                if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE)
                        reshdr_flags |= WIM_RESHDR_FLAG_COMPRESSED;
                ret = write_pwm_stream_header(lte, out_fd, reshdr_flags);
                if (ret)
                        goto out_free_chunk_tab;
                res_start_offset = out_fd->offset;
        }

        /* Write the entire resource by reading the entire resource and feeding
         * the data through the write_resource_cb function. */
        write_ctx.out_fd = out_fd;
        write_ctx.resource_flags = resource_flags;
try_write_again:
        ret = read_resource_prefix(lte, read_size,
                                   write_resource_cb, &write_ctx, resource_flags);
        if (ret)
                goto out_free_chunk_tab;

        /* Verify SHA1 message digest of the resource, or set the hash for the
         * first time. */
        if (write_ctx.doing_sha) {
                ret = finalize_and_check_sha1(&write_ctx.sha_ctx, lte);
                if (ret)
                        goto out_free_chunk_tab;
        }

        /* Write chunk table if needed.  */
        if (write_ctx.chunk_tab) {
                ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab,
                                                    out_fd,
                                                    res_start_offset,
                                                    resource_flags);
                if (ret)
                        goto out_free_chunk_tab;
        }

        /* Fill in out_res_entry with information about the newly written
         * resource.  */
        out_res_entry->size          = out_fd->offset - res_start_offset;
        out_res_entry->flags         = lte->resource_entry.flags;
        if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE)
                out_res_entry->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
        else
                out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
        out_res_entry->offset        = res_start_offset;
        out_res_entry->original_size = wim_resource_size(lte);

        /* Check for resources compressed to greater than their original size
         * and write them uncompressed instead.  (But never do this if writing
         * to a pipe, and don't bother if we did a raw copy.)  */
        if (out_res_entry->size > out_res_entry->original_size &&
            !(resource_flags & (WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE |
                                WIMLIB_READ_RESOURCE_FLAG_RAW)))
        {
                DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
                      "writing uncompressed instead",
                      out_res_entry->original_size, out_res_entry->size);
                ret = seek_and_truncate(out_fd, res_start_offset);
                if (ret)
                        goto out_free_chunk_tab;
                out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
                FREE(write_ctx.chunk_tab);
                write_ctx.out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
                write_ctx.chunk_tab = NULL;
                write_ctx.doing_sha = false;
                goto try_write_again;
        }
        if (resource_flags & (WIMLIB_READ_RESOURCE_FLAG_RAW)) {
                DEBUG("Copied raw compressed data "
                      "(%"PRIu64" => %"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
                      out_res_entry->original_size, out_res_entry->size,
                      out_res_entry->offset, out_res_entry->flags);
        } else if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
                DEBUG("Wrote compressed resource "
                      "(%"PRIu64" => %"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
                      out_res_entry->original_size, out_res_entry->size,
                      out_res_entry->offset, out_res_entry->flags);
        } else {
                DEBUG("Wrote uncompressed resource "
                      "(%"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
                      out_res_entry->original_size,
                      out_res_entry->offset, out_res_entry->flags);
        }
        ret = 0;
out_free_chunk_tab:
        FREE(write_ctx.chunk_tab);
out:
        return ret;
}

/* Like write_wim_resource(), but the resource is specified by a buffer of
 * uncompressed data rather a lookup table entry; also writes the SHA1 hash of
 * the buffer to @hash_ret.  */
int
write_wim_resource_from_buffer(const void *buf, size_t buf_size,
                               int reshdr_flags, struct filedes *out_fd,
                               int out_ctype,
                               struct resource_entry *out_res_entry,
                               u8 *hash_ret, int write_resource_flags,
                               struct wimlib_lzx_context **comp_ctx)
{
        /* Set up a temporary lookup table entry to provide to
         * write_wim_resource(). */
        struct wim_lookup_table_entry lte;
        int ret;

        lte.resource_location            = RESOURCE_IN_ATTACHED_BUFFER;
        lte.attached_buffer              = (void*)buf;
        lte.resource_entry.original_size = buf_size;
        lte.resource_entry.flags         = reshdr_flags;

        if (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
                sha1_buffer(buf, buf_size, lte.hash);
                lte.unhashed = 0;
        } else {
                lte.unhashed = 1;
        }

        ret = write_wim_resource(&lte, out_fd, out_ctype, out_res_entry,
                                 write_resource_flags, comp_ctx);
        if (ret)
                return ret;
        if (hash_ret)
                copy_hash(hash_ret, lte.hash);
        return 0;
}


#ifdef ENABLE_MULTITHREADED_COMPRESSION

/* Blocking shared queue (solves the producer-consumer problem) */
struct shared_queue {
        unsigned size;
        unsigned front;
        unsigned back;
        unsigned filled_slots;
        void **array;
        pthread_mutex_t lock;
        pthread_cond_t msg_avail_cond;
        pthread_cond_t space_avail_cond;
};

static int
shared_queue_init(struct shared_queue *q, unsigned size)
{
        wimlib_assert(size != 0);
        q->array = CALLOC(sizeof(q->array[0]), size);
        if (!q->array)
                goto err;
        q->filled_slots = 0;
        q->front = 0;
        q->back = size - 1;
        q->size = size;
        if (pthread_mutex_init(&q->lock, NULL)) {
                ERROR_WITH_ERRNO("Failed to initialize mutex");
                goto err;
        }
        if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
                ERROR_WITH_ERRNO("Failed to initialize condition variable");
                goto err_destroy_lock;
        }
        if (pthread_cond_init(&q->space_avail_cond, NULL)) {
                ERROR_WITH_ERRNO("Failed to initialize condition variable");
                goto err_destroy_msg_avail_cond;
        }
        return 0;
err_destroy_msg_avail_cond:
        pthread_cond_destroy(&q->msg_avail_cond);
err_destroy_lock:
        pthread_mutex_destroy(&q->lock);
err:
        return WIMLIB_ERR_NOMEM;
}

static void
shared_queue_destroy(struct shared_queue *q)
{
        FREE(q->array);
        pthread_mutex_destroy(&q->lock);
        pthread_cond_destroy(&q->msg_avail_cond);
        pthread_cond_destroy(&q->space_avail_cond);
}

static void
shared_queue_put(struct shared_queue *q, void *obj)
{
        pthread_mutex_lock(&q->lock);
        while (q->filled_slots == q->size)
                pthread_cond_wait(&q->space_avail_cond, &q->lock);

        q->back = (q->back + 1) % q->size;
        q->array[q->back] = obj;
        q->filled_slots++;

        pthread_cond_broadcast(&q->msg_avail_cond);
        pthread_mutex_unlock(&q->lock);
}

static void *
shared_queue_get(struct shared_queue *q)
{
        void *obj;

        pthread_mutex_lock(&q->lock);
        while (q->filled_slots == 0)
                pthread_cond_wait(&q->msg_avail_cond, &q->lock);

        obj = q->array[q->front];
        q->array[q->front] = NULL;
        q->front = (q->front + 1) % q->size;
        q->filled_slots--;

        pthread_cond_broadcast(&q->space_avail_cond);
        pthread_mutex_unlock(&q->lock);
        return obj;
}

struct compressor_thread_params {
        struct shared_queue *res_to_compress_queue;
        struct shared_queue *compressed_res_queue;
        int out_ctype;
        struct wimlib_lzx_context *comp_ctx;
};

#define MAX_CHUNKS_PER_MSG 2

struct message {
        struct wim_lookup_table_entry *lte;
        u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
        u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
        unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
        struct iovec out_chunks[MAX_CHUNKS_PER_MSG];
        unsigned num_chunks;
        struct list_head list;
        bool complete;
        u64 begin_chunk;
};

static void
compress_chunks(struct message *msg, int out_ctype,
                struct wimlib_lzx_context *comp_ctx)
{
        for (unsigned i = 0; i < msg->num_chunks; i++) {
                unsigned len;

                len = compress_chunk(msg->uncompressed_chunks[i],
                                     msg->uncompressed_chunk_sizes[i],
                                     msg->compressed_chunks[i],
                                     out_ctype,
                                     comp_ctx);

                void *out_chunk;
                unsigned out_len;
                if (len) {
                        /* To be written compressed */
                        out_chunk = msg->compressed_chunks[i];
                        out_len = len;
                } else {
                        /* To be written uncompressed */
                        out_chunk = msg->uncompressed_chunks[i];
                        out_len = msg->uncompressed_chunk_sizes[i];
                }
                msg->out_chunks[i].iov_base = out_chunk;
                msg->out_chunks[i].iov_len = out_len;
        }
}

/* Compressor thread routine.  This is a lot simpler than the main thread
 * routine: just repeatedly get a group of chunks from the
 * res_to_compress_queue, compress them, and put them in the
 * compressed_res_queue.  A NULL pointer indicates that the thread should stop.
 * */
static void *
compressor_thread_proc(void *arg)
{
        struct compressor_thread_params *params = arg;
        struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
        struct shared_queue *compressed_res_queue = params->compressed_res_queue;
        struct message *msg;

        DEBUG("Compressor thread ready");
        while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
                compress_chunks(msg, params->out_ctype, params->comp_ctx);
                shared_queue_put(compressed_res_queue, msg);
        }
        DEBUG("Compressor thread terminating");
        return NULL;
}
#endif /* ENABLE_MULTITHREADED_COMPRESSION */

struct write_streams_progress_data {
        wimlib_progress_func_t progress_func;
        union wimlib_progress_info progress;
        uint64_t next_progress;
        WIMStruct *prev_wim_part;
};

static void
do_write_streams_progress(struct write_streams_progress_data *progress_data,
                          struct wim_lookup_table_entry *lte,
                          bool stream_discarded)
{
        union wimlib_progress_info *progress = &progress_data->progress;
        bool new_wim_part;

        if (stream_discarded) {
                progress->write_streams.total_bytes -= wim_resource_size(lte);
                if (progress_data->next_progress != ~(uint64_t)0 &&
                    progress_data->next_progress > progress->write_streams.total_bytes)
                {
                        progress_data->next_progress = progress->write_streams.total_bytes;
                }
        } else {
                progress->write_streams.completed_bytes += wim_resource_size(lte);
        }
        new_wim_part = false;
        if (lte->resource_location == RESOURCE_IN_WIM &&
            lte->wim != progress_data->prev_wim_part)
        {
                if (progress_data->prev_wim_part) {
                        new_wim_part = true;
                        progress->write_streams.completed_parts++;
                }
                progress_data->prev_wim_part = lte->wim;
        }
        progress->write_streams.completed_streams++;
        if (progress_data->progress_func
            && (progress->write_streams.completed_bytes >= progress_data->next_progress
                || new_wim_part))
        {
                progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
                                             progress);
                if (progress_data->next_progress == progress->write_streams.total_bytes) {
                        progress_data->next_progress = ~(uint64_t)0;
                } else {
                        progress_data->next_progress =
                                min(progress->write_streams.total_bytes,
                                    progress->write_streams.completed_bytes +
                                        progress->write_streams.total_bytes / 100);
                }
        }
}

struct serial_write_stream_ctx {
        struct filedes *out_fd;
        int out_ctype;
        struct wimlib_lzx_context **comp_ctx;
        int write_resource_flags;
};

static int
serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
{
        struct serial_write_stream_ctx *ctx = _ctx;
        return write_wim_resource(lte, ctx->out_fd,
                                  ctx->out_ctype, &lte->output_resource_entry,
                                  ctx->write_resource_flags,
                                  ctx->comp_ctx);
}


/* Write a list of streams, taking into account that some streams may be
 * duplicates that are checksummed and discarded on the fly, and also delegating
 * the actual writing of a stream to a function @write_stream_cb, which is
 * passed the context @write_stream_ctx. */
static int
do_write_stream_list(struct list_head *stream_list,
                     struct wim_lookup_table *lookup_table,
                     int (*write_stream_cb)(struct wim_lookup_table_entry *, void *),
                     void *write_stream_ctx,
                     struct write_streams_progress_data *progress_data)
{
        int ret = 0;
        struct wim_lookup_table_entry *lte;
        bool stream_discarded;

        /* For each stream in @stream_list ... */
        while (!list_empty(stream_list)) {
                stream_discarded = false;
                lte = container_of(stream_list->next,
                                   struct wim_lookup_table_entry,
                                   write_streams_list);
                list_del(&lte->write_streams_list);
                if (lte->unhashed && !lte->unique_size) {
                        /* Unhashed stream that shares a size with some other
                         * stream in the WIM we are writing.  The stream must be
                         * checksummed to know if we need to write it or not. */
                        struct wim_lookup_table_entry *tmp;
                        u32 orig_out_refcnt = lte->out_refcnt;

                        ret = hash_unhashed_stream(lte, lookup_table, &tmp);
                        if (ret)
                                break;
                        if (tmp != lte) {
                                /* We found a duplicate stream.  'lte' was
                                 * freed, so replace it with the duplicate.  */
                                lte = tmp;

                                /* 'out_refcnt' was transferred to the
                                 * duplicate, and we can detect if the duplicate
                                 * stream was already referenced for writing by
                                 * checking if its 'out_refcnt' is higher than
                                 * that of the original stream.  In such cases,
                                 * the current stream can be discarded.  We can
                                 * also discard the current stream if it was
                                 * previously marked as filtered (e.g. already
                                 * present in the WIM being written).  */
                                if (lte->out_refcnt > orig_out_refcnt ||
                                    lte->filtered) {
                                        DEBUG("Discarding duplicate stream of "
                                              "length %"PRIu64,
                                              wim_resource_size(lte));
                                        lte->no_progress = 0;
                                        stream_discarded = true;
                                        goto skip_to_progress;
                                }
                        }
                }

                /* Here, @lte is either a hashed stream or an unhashed stream
                 * with a unique size.  In either case we know that the stream
                 * has to be written.  In either case the SHA1 message digest
                 * will be calculated over the stream while writing it; however,
                 * in the former case this is done merely to check the data,
                 * while in the latter case this is done because we do not have
                 * the SHA1 message digest yet.  */
                wimlib_assert(lte->out_refcnt != 0);
                lte->deferred = 0;
                lte->no_progress = 0;
                ret = (*write_stream_cb)(lte, write_stream_ctx);
                if (ret)
                        break;
                /* In parallel mode, some streams are deferred for later,
                 * serialized processing; ignore them here. */
                if (lte->deferred)
                        continue;
                if (lte->unhashed) {
                        list_del(&lte->unhashed_list);
                        lookup_table_insert(lookup_table, lte);
                        lte->unhashed = 0;
                }
        skip_to_progress:
                if (!lte->no_progress) {
                        do_write_streams_progress(progress_data,
                                                  lte, stream_discarded);
                }
        }
        return ret;
}

static int
do_write_stream_list_serial(struct list_head *stream_list,
                            struct wim_lookup_table *lookup_table,
                            struct filedes *out_fd,
                            int out_ctype,
                            struct wimlib_lzx_context **comp_ctx,
                            int write_resource_flags,
                            struct write_streams_progress_data *progress_data)
{
        struct serial_write_stream_ctx ctx = {
                .out_fd = out_fd,
                .out_ctype = out_ctype,
                .write_resource_flags = write_resource_flags,
                .comp_ctx = comp_ctx,
        };
        return do_write_stream_list(stream_list,
                                    lookup_table,
                                    serial_write_stream,
                                    &ctx,
                                    progress_data);
}

static inline int
write_flags_to_resource_flags(int write_flags)
{
        int resource_flags = 0;

        if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
                resource_flags |= WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS;
        if (write_flags & WIMLIB_WRITE_FLAG_COMPRESS_SLOW)
                resource_flags |= WIMLIB_WRITE_RESOURCE_FLAG_COMPRESS_SLOW;
        if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
                resource_flags |= WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE;
        return resource_flags;
}

static int
write_stream_list_serial(struct list_head *stream_list,
                         struct wim_lookup_table *lookup_table,
                         struct filedes *out_fd,
                         int out_ctype,
                         struct wimlib_lzx_context **comp_ctx,
                         int write_resource_flags,
                         struct write_streams_progress_data *progress_data)
{
        union wimlib_progress_info *progress = &progress_data->progress;
        DEBUG("Writing stream list of size %"PRIu64" (serial version)",
              progress->write_streams.total_streams);
        progress->write_streams.num_threads = 1;
        if (progress_data->progress_func) {
                progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
                                             progress);
        }
        return do_write_stream_list_serial(stream_list,
                                           lookup_table,
                                           out_fd,
                                           out_ctype,
                                           comp_ctx,
                                           write_resource_flags,
                                           progress_data);
}

#ifdef ENABLE_MULTITHREADED_COMPRESSION
static int
write_wim_chunks(struct message *msg, struct filedes *out_fd,
                 struct chunk_table *chunk_tab,
                 int write_resource_flags)
{
        struct iovec *vecs;
        struct pwm_chunk_hdr *chunk_hdrs;
        unsigned nvecs;
        int ret;

        for (unsigned i = 0; i < msg->num_chunks; i++)
                chunk_tab_record_chunk(chunk_tab, msg->out_chunks[i].iov_len);

        if (!(write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE)) {
                nvecs = msg->num_chunks;
                vecs = msg->out_chunks;
        } else {
                /* Special case:  If writing a compressed resource to a pipable
                 * WIM, prefix each compressed chunk with a header that gives
                 * its compressed size.  */
                nvecs = msg->num_chunks * 2;
                vecs = alloca(nvecs * sizeof(vecs[0]));
                chunk_hdrs = alloca(msg->num_chunks * sizeof(chunk_hdrs[0]));

                for (unsigned i = 0; i < msg->num_chunks; i++) {
                        chunk_hdrs[i].compressed_size = cpu_to_le32(msg->out_chunks[i].iov_len);
                        vecs[i * 2].iov_base = &chunk_hdrs[i];
                        vecs[i * 2].iov_len = sizeof(chunk_hdrs[i]);
                        vecs[i * 2 + 1].iov_base = msg->out_chunks[i].iov_base;
                        vecs[i * 2 + 1].iov_len = msg->out_chunks[i].iov_len;
                }
        }
        ret = full_writev(out_fd, vecs, nvecs);
        if (ret)
                ERROR_WITH_ERRNO("Failed to write WIM chunks");
        return ret;
}

struct main_writer_thread_ctx {
        struct list_head *stream_list;
        struct wim_lookup_table *lookup_table;
        struct filedes *out_fd;
        off_t res_start_offset;
        int out_ctype;
        struct wimlib_lzx_context **comp_ctx;
        int write_resource_flags;
        struct shared_queue *res_to_compress_queue;
        struct shared_queue *compressed_res_queue;
        size_t num_messages;
        struct write_streams_progress_data *progress_data;

        struct list_head available_msgs;
        struct list_head outstanding_streams;
        struct list_head serial_streams;
        size_t num_outstanding_messages;

        SHA_CTX next_sha_ctx;
        u64 next_chunk;
        u64 next_num_chunks;
        struct wim_lookup_table_entry *next_lte;

        struct message *msgs;
        struct message *next_msg;
        struct chunk_table *cur_chunk_tab;
};

static int
init_message(struct message *msg)
{
        for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
                msg->compressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
                msg->uncompressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
                if (msg->compressed_chunks[i] == NULL ||
                    msg->uncompressed_chunks[i] == NULL)
                        return WIMLIB_ERR_NOMEM;
        }
        return 0;
}

static void
destroy_message(struct message *msg)
{
        for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
                FREE(msg->compressed_chunks[i]);
                FREE(msg->uncompressed_chunks[i]);
        }
}

static void
free_messages(struct message *msgs, size_t num_messages)
{
        if (msgs) {
                for (size_t i = 0; i < num_messages; i++)
                        destroy_message(&msgs[i]);
                FREE(msgs);
        }
}

static struct message *
allocate_messages(size_t num_messages)
{
        struct message *msgs;

        msgs = CALLOC(num_messages, sizeof(struct message));
        if (!msgs)
                return NULL;
        for (size_t i = 0; i < num_messages; i++) {
                if (init_message(&msgs[i])) {
                        free_messages(msgs, num_messages);
                        return NULL;
                }
        }
        return msgs;
}

static void
main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
{
        while (ctx->num_outstanding_messages--)
                shared_queue_get(ctx->compressed_res_queue);
        free_messages(ctx->msgs, ctx->num_messages);
        FREE(ctx->cur_chunk_tab);
}

static int
main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
{
        /* Pre-allocate all the buffers that will be needed to do the chunk
         * compression. */
        ctx->msgs = allocate_messages(ctx->num_messages);
        if (!ctx->msgs)
                return WIMLIB_ERR_NOMEM;

        /* Initially, all the messages are available to use. */
        INIT_LIST_HEAD(&ctx->available_msgs);
        for (size_t i = 0; i < ctx->num_messages; i++)
                list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);

        /* outstanding_streams is the list of streams that currently have had
         * chunks sent off for compression.
         *
         * The first stream in outstanding_streams is the stream that is
         * currently being written.
         *
         * The last stream in outstanding_streams is the stream that is
         * currently being read and having chunks fed to the compressor threads.
         * */
        INIT_LIST_HEAD(&ctx->outstanding_streams);
        ctx->num_outstanding_messages = 0;

        ctx->next_msg = NULL;

        /* Resources that don't need any chunks compressed are added to this
         * list and written directly by the main thread. */
        INIT_LIST_HEAD(&ctx->serial_streams);

        ctx->cur_chunk_tab = NULL;

        return 0;
}

static int
receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
{
        struct message *msg;
        struct wim_lookup_table_entry *cur_lte;
        int ret;

        wimlib_assert(!list_empty(&ctx->outstanding_streams));
        wimlib_assert(ctx->num_outstanding_messages != 0);

        cur_lte = container_of(ctx->outstanding_streams.next,
                               struct wim_lookup_table_entry,
                               being_compressed_list);

        /* Get the next message from the queue and process it.
         * The message will contain 1 or more data chunks that have been
         * compressed. */
        msg = shared_queue_get(ctx->compressed_res_queue);
        msg->complete = true;
        --ctx->num_outstanding_messages;

        /* Is this the next chunk in the current resource?  If it's not
         * (i.e., an earlier chunk in a same or different resource
         * hasn't been compressed yet), do nothing, and keep this
         * message around until all earlier chunks are received.
         *
         * Otherwise, write all the chunks we can. */
        while (cur_lte != NULL &&
               !list_empty(&cur_lte->msg_list)
               && (msg = container_of(cur_lte->msg_list.next,
                                      struct message,
                                      list))->complete)
        {
                list_move(&msg->list, &ctx->available_msgs);
                if (msg->begin_chunk == 0) {
                        /* First set of chunks.  */

                        /* Write pipable WIM stream header if needed.  */
                        if (ctx->write_resource_flags &
                            WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE)
                        {
                                ret = write_pwm_stream_header(cur_lte, ctx->out_fd,
                                                              WIM_RESHDR_FLAG_COMPRESSED);
                                if (ret)
                                        return ret;
                        }

                        /* Save current offset.  */
                        ctx->res_start_offset = ctx->out_fd->offset;

                        /* Begin building the chunk table, and leave space for
                         * it if needed.  */
                        ret = begin_wim_resource_chunk_tab(cur_lte,
                                                           ctx->out_fd,
                                                           &ctx->cur_chunk_tab,
                                                           ctx->write_resource_flags);
                        if (ret)
                                return ret;

                }

                /* Write the compressed chunks from the message. */
                ret = write_wim_chunks(msg, ctx->out_fd, ctx->cur_chunk_tab,
                                       ctx->write_resource_flags);
                if (ret)
                        return ret;

                /* Was this the last chunk of the stream?  If so, finish
                 * it. */
                if (list_empty(&cur_lte->msg_list) &&
                    msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
                {
                        u64 res_csize;

                        ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
                                                            ctx->out_fd,
                                                            ctx->res_start_offset,
                                                            ctx->write_resource_flags);
                        if (ret)
                                return ret;

                        list_del(&cur_lte->being_compressed_list);

                        res_csize = ctx->out_fd->offset - ctx->res_start_offset;

                        FREE(ctx->cur_chunk_tab);
                        ctx->cur_chunk_tab = NULL;

                        /* Check for resources compressed to greater than or
                         * equal to their original size and write them
                         * uncompressed instead.  (But never do this if writing
                         * to a pipe.)  */
                        if (res_csize >= wim_resource_size(cur_lte) &&
                            !(ctx->write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE))
                        {
                                DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
                                      "writing uncompressed instead",
                                      wim_resource_size(cur_lte), res_csize);
                                ret = seek_and_truncate(ctx->out_fd, ctx->res_start_offset);
                                if (ret)
                                        return ret;
                                ret = write_wim_resource(cur_lte,
                                                         ctx->out_fd,
                                                         WIMLIB_COMPRESSION_TYPE_NONE,
                                                         &cur_lte->output_resource_entry,
                                                         ctx->write_resource_flags,
                                                         ctx->comp_ctx);
                                if (ret)
                                        return ret;
                        } else {
                                cur_lte->output_resource_entry.size =
                                        res_csize;

                                cur_lte->output_resource_entry.original_size =
                                        cur_lte->resource_entry.original_size;

                                cur_lte->output_resource_entry.offset =
                                        ctx->res_start_offset;

                                cur_lte->output_resource_entry.flags =
                                        cur_lte->resource_entry.flags |
                                                WIM_RESHDR_FLAG_COMPRESSED;

                                DEBUG("Wrote compressed resource "
                                      "(%"PRIu64" => %"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
                                      cur_lte->output_resource_entry.original_size,
                                      cur_lte->output_resource_entry.size,
                                      cur_lte->output_resource_entry.offset,
                                      cur_lte->output_resource_entry.flags);
                        }

                        do_write_streams_progress(ctx->progress_data,
                                                  cur_lte, false);

                        /* Since we just finished writing a stream, write any
                         * streams that have been added to the serial_streams
                         * list for direct writing by the main thread (e.g.
                         * resources that don't need to be compressed because
                         * the desired compression type is the same as the
                         * previous compression type). */
                        if (!list_empty(&ctx->serial_streams)) {
                                ret = do_write_stream_list_serial(&ctx->serial_streams,
                                                                  ctx->lookup_table,
                                                                  ctx->out_fd,
                                                                  ctx->out_ctype,
                                                                  ctx->comp_ctx,
                                                                  ctx->write_resource_flags,
                                                                  ctx->progress_data);
                                if (ret)
                                        return ret;
                        }

                        /* Advance to the next stream to write. */
                        if (list_empty(&ctx->outstanding_streams)) {
                                cur_lte = NULL;
                        } else {
                                cur_lte = container_of(ctx->outstanding_streams.next,
                                                       struct wim_lookup_table_entry,
                                                       being_compressed_list);
                        }
                }
        }
        return 0;
}

/* Called when the main thread has read a new chunk of data. */
static int
main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
{
        struct main_writer_thread_ctx *ctx = _ctx;
        int ret;
        struct message *next_msg;
        u64 next_chunk_in_msg;

        /* Update SHA1 message digest for the stream currently being read by the
         * main thread. */
        sha1_update(&ctx->next_sha_ctx, chunk, chunk_size);

        /* We send chunks of data to the compressor chunks in batches which we
         * refer to as "messages".  @next_msg is the message that is currently
         * being prepared to send off.  If it is NULL, that indicates that we
         * need to start a new message. */
        next_msg = ctx->next_msg;
        if (!next_msg) {
                /* We need to start a new message.  First check to see if there
                 * is a message available in the list of available messages.  If
                 * so, we can just take one.  If not, all the messages (there is
                 * a fixed number of them, proportional to the number of
                 * threads) have been sent off to the compressor threads, so we
                 * receive messages from the compressor threads containing
                 * compressed chunks of data.
                 *
                 * We may need to receive multiple messages before one is
                 * actually available to use because messages received that are
                 * *not* for the very next set of chunks to compress must be
                 * buffered until it's time to write those chunks. */
                while (list_empty(&ctx->available_msgs)) {
                        ret = receive_compressed_chunks(ctx);
                        if (ret)
                                return ret;
                }

                next_msg = container_of(ctx->available_msgs.next,
                                        struct message, list);
                list_del(&next_msg->list);
                next_msg->complete = false;
                next_msg->begin_chunk = ctx->next_chunk;
                next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
                                           ctx->next_num_chunks - ctx->next_chunk);
                ctx->next_msg = next_msg;
        }

        /* Fill in the next chunk to compress */
        next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;

        next_msg->uncompressed_chunk_sizes[next_chunk_in_msg] = chunk_size;
        memcpy(next_msg->uncompressed_chunks[next_chunk_in_msg],
               chunk, chunk_size);
        ctx->next_chunk++;
        if (++next_chunk_in_msg == next_msg->num_chunks) {
                /* Send off an array of chunks to compress */
                list_add_tail(&next_msg->list, &ctx->next_lte->msg_list);
                shared_queue_put(ctx->res_to_compress_queue, next_msg);
                ++ctx->num_outstanding_messages;
                ctx->next_msg = NULL;
        }
        return 0;
}

static int
main_writer_thread_finish(void *_ctx)
{
        struct main_writer_thread_ctx *ctx = _ctx;
        int ret;
        while (ctx->num_outstanding_messages != 0) {
                ret = receive_compressed_chunks(ctx);
                if (ret)
                        return ret;
        }
        wimlib_assert(list_empty(&ctx->outstanding_streams));
        return do_write_stream_list_serial(&ctx->serial_streams,
                                           ctx->lookup_table,
                                           ctx->out_fd,
                                           ctx->out_ctype,
                                           ctx->comp_ctx,
                                           ctx->write_resource_flags,
                                           ctx->progress_data);
}

static int
submit_stream_for_compression(struct wim_lookup_table_entry *lte,
                              struct main_writer_thread_ctx *ctx)
{
        int ret;

        /* Read the entire stream @lte, feeding its data chunks to the
         * compressor threads.  Also SHA1-sum the stream; this is required in
         * the case that @lte is unhashed, and a nice additional verification
         * when @lte is already hashed. */
        sha1_init(&ctx->next_sha_ctx);
        ctx->next_chunk = 0;
        ctx->next_num_chunks = wim_resource_chunks(lte);
        ctx->next_lte = lte;
        INIT_LIST_HEAD(&lte->msg_list);
        list_add_tail(&lte->being_compressed_list, &ctx->outstanding_streams);
        ret = read_resource_prefix(lte, wim_resource_size(lte),
                                   main_writer_thread_cb, ctx, 0);
        if (ret)
                return ret;
        wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
        return finalize_and_check_sha1(&ctx->next_sha_ctx, lte);
}

static int
main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
{
        struct main_writer_thread_ctx *ctx = _ctx;
        int ret;

        if (wim_resource_size(lte) < 1000 ||
            ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
            (lte->resource_location == RESOURCE_IN_WIM &&
             !(ctx->write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS) &&
             lte->wim->compression_type == ctx->out_ctype))
        {
                /* Stream is too small or isn't being compressed.  Process it by
                 * the main thread when we have a chance.  We can't necessarily
                 * process it right here, as the main thread could be in the
                 * middle of writing a different stream. */
                list_add_tail(&lte->write_streams_list, &ctx->serial_streams);
                lte->deferred = 1;
                ret = 0;
        } else {
                ret = submit_stream_for_compression(lte, ctx);
        }
        lte->no_progress = 1;
        return ret;
}

static long
get_default_num_threads(void)
{
#ifdef __WIN32__
        return win32_get_number_of_processors();
#else
        return sysconf(_SC_NPROCESSORS_ONLN);
#endif
}

/* Equivalent to write_stream_list_serial(), except this takes a @num_threads
 * parameter and will perform compression using that many threads.  Falls
 * back to write_stream_list_serial() on certain errors, such as a failure to
 * create the number of threads requested.
 *
 * High level description of the algorithm for writing compressed streams in
 * parallel:  We perform compression on chunks of size WIM_CHUNK_SIZE bytes
 * rather than on full files.  The currently executing thread becomes the main
 * thread and is entirely in charge of reading the data to compress (which may
 * be in any location understood by the resource code--- such as in an external
 * file being captured, or in another WIM file from which an image is being
 * exported) and actually writing the compressed data to the output file.
 * Additional threads are "compressor threads" and all execute the
 * compressor_thread_proc, where they repeatedly retrieve buffers of data from
 * the main thread, compress them, and hand them back to the main thread.
 *
 * Certain streams, such as streams that do not need to be compressed (e.g.
 * input compression type same as output compression type) or streams of very
 * small size are placed in a list (main_writer_thread_ctx.serial_list) and
 * handled entirely by the main thread at an appropriate time.
 *
 * At any given point in time, multiple streams may be having chunks compressed
 * concurrently.  The stream that the main thread is currently *reading* may be
 * later in the list that the stream that the main thread is currently
 * *writing*.
 */
static int
write_stream_list_parallel(struct list_head *stream_list,
                           struct wim_lookup_table *lookup_table,
                           struct filedes *out_fd,
                           int out_ctype,
                           struct wimlib_lzx_context **comp_ctx,
                           int write_resource_flags,
                           struct write_streams_progress_data *progress_data,
                           unsigned num_threads)
{
        int ret;
        struct shared_queue res_to_compress_queue;
        struct shared_queue compressed_res_queue;
        pthread_t *compressor_threads = NULL;
        union wimlib_progress_info *progress = &progress_data->progress;

        if (num_threads == 0) {
                long nthreads = get_default_num_threads();
                if (nthreads < 1 || nthreads > UINT_MAX) {
                        WARNING("Could not determine number of processors! Assuming 1");
                        goto out_serial;
                } else if (nthreads == 1) {
                        goto out_serial_quiet;
                } else {
                        num_threads = nthreads;
                }
        }

        DEBUG("Writing stream list of size %"PRIu64" "
              "(parallel version, num_threads=%u)",
              progress->write_streams.total_streams, num_threads);

        progress->write_streams.num_threads = num_threads;

        static const size_t MESSAGES_PER_THREAD = 2;
        size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);

        DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);

        ret = shared_queue_init(&res_to_compress_queue, queue_size);
        if (ret)
                goto out_serial;

        ret = shared_queue_init(&compressed_res_queue, queue_size);
        if (ret)
                goto out_destroy_res_to_compress_queue;

        struct compressor_thread_params *params;

        params = CALLOC(num_threads, sizeof(params[0]));
        if (params == NULL) {
                ret = WIMLIB_ERR_NOMEM;
                goto out_destroy_compressed_res_queue;
        }

        for (unsigned i = 0; i < num_threads; i++) {
                params[i].res_to_compress_queue = &res_to_compress_queue;
                params[i].compressed_res_queue = &compressed_res_queue;
                params[i].out_ctype = out_ctype;
                if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX) {
                        ret = alloc_lzx_context(write_resource_flags,
                                                &params[i].comp_ctx);
                        if (ret)
                                goto out_free_params;
                }
        }

        compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
        if (!compressor_threads) {
                ret = WIMLIB_ERR_NOMEM;
                goto out_free_params;
        }

        for (unsigned i = 0; i < num_threads; i++) {
                DEBUG("pthread_create thread %u of %u", i + 1, num_threads);
                ret = pthread_create(&compressor_threads[i], NULL,
                                     compressor_thread_proc, &params[i]);
                if (ret != 0) {
                        ret = -1;
                        ERROR_WITH_ERRNO("Failed to create compressor "
                                         "thread %u of %u",
                                         i + 1, num_threads);
                        num_threads = i;
                        goto out_join;
                }
        }

        if (progress_data->progress_func) {
                progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
                                             progress);
        }

        struct main_writer_thread_ctx ctx;
        ctx.stream_list           = stream_list;
        ctx.lookup_table          = lookup_table;
        ctx.out_fd                = out_fd;
        ctx.out_ctype             = out_ctype;
        ctx.comp_ctx              = comp_ctx;
        ctx.res_to_compress_queue = &res_to_compress_queue;
        ctx.compressed_res_queue  = &compressed_res_queue;
        ctx.num_messages          = queue_size;
        ctx.write_resource_flags  = write_resource_flags;
        ctx.progress_data         = progress_data;
        ret = main_writer_thread_init_ctx(&ctx);
        if (ret)
                goto out_join;
        ret = do_write_stream_list(stream_list, lookup_table,
                                   main_thread_process_next_stream,
                                   &ctx, progress_data);
        if (ret)
                goto out_destroy_ctx;

        /* The main thread has finished reading all streams that are going to be
         * compressed in parallel, and it now needs to wait for all remaining
         * chunks to be compressed so that the remaining streams can actually be
         * written to the output file.  Furthermore, any remaining streams that
         * had processing deferred to the main thread need to be handled.  These
         * tasks are done by the main_writer_thread_finish() function. */
        ret = main_writer_thread_finish(&ctx);
out_destroy_ctx:
        main_writer_thread_destroy_ctx(&ctx);
out_join:
        for (unsigned i = 0; i < num_threads; i++)
                shared_queue_put(&res_to_compress_queue, NULL);

        for (unsigned i = 0; i < num_threads; i++) {
                if (pthread_join(compressor_threads[i], NULL)) {
                        WARNING_WITH_ERRNO("Failed to join compressor "
                                           "thread %u of %u",
                                           i + 1, num_threads);
                }
        }
        FREE(compressor_threads);
out_free_params:
        for (unsigned i = 0; i < num_threads; i++)
                wimlib_lzx_free_context(params[i].comp_ctx);
        FREE(params);
out_destroy_compressed_res_queue:
        shared_queue_destroy(&compressed_res_queue);
out_destroy_res_to_compress_queue:
        shared_queue_destroy(&res_to_compress_queue);
        if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
                return ret;
out_serial:
        WARNING("Falling back to single-threaded compression");
out_serial_quiet:
        return write_stream_list_serial(stream_list,
                                        lookup_table,
                                        out_fd,
                                        out_ctype,
                                        comp_ctx,
                                        write_resource_flags,
                                        progress_data);

}
#endif

/*
 * Write a list of streams to a WIM (@out_fd) using the compression type
 * @out_ctype and up to @num_threads compressor threads.
 */
static int
write_stream_list(struct list_head *stream_list,
                  struct wim_lookup_table *lookup_table,
                  struct filedes *out_fd, int out_ctype,
                  struct wimlib_lzx_context **comp_ctx,
                  int write_flags,
                  unsigned num_threads, wimlib_progress_func_t progress_func)
{
        struct wim_lookup_table_entry *lte;
        size_t num_streams = 0;
        u64 total_bytes = 0;
        u64 total_compression_bytes = 0;
        struct write_streams_progress_data progress_data;
        int ret;
        int write_resource_flags;
        unsigned total_parts = 0;
        WIMStruct *prev_wim_part = NULL;

        if (list_empty(stream_list)) {
                DEBUG("No streams to write.");
                return 0;
        }

        write_resource_flags = write_flags_to_resource_flags(write_flags);

        DEBUG("Writing stream list (offset = %"PRIu64", write_resource_flags=0x%08x)",
              out_fd->offset, write_resource_flags);

        sort_stream_list_by_sequential_order(stream_list,
                                             offsetof(struct wim_lookup_table_entry,
                                                      write_streams_list));

        /* Calculate the total size of the streams to be written.  Note: this
         * will be the uncompressed size, as we may not know the compressed size
         * yet, and also this will assume that every unhashed stream will be
         * written (which will not necessarily be the case). */
        list_for_each_entry(lte, stream_list, write_streams_list) {
                num_streams++;
                total_bytes += wim_resource_size(lte);
                if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
                       && (wim_resource_compression_type(lte) != out_ctype ||
                           (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS)))
                {
                        total_compression_bytes += wim_resource_size(lte);
                }
                if (lte->resource_location == RESOURCE_IN_WIM) {
                        if (prev_wim_part != lte->wim) {
                                prev_wim_part = lte->wim;
                                total_parts++;
                        }
                }
        }

        memset(&progress_data, 0, sizeof(progress_data));
        progress_data.progress_func = progress_func;

        progress_data.progress.write_streams.total_bytes       = total_bytes;
        progress_data.progress.write_streams.total_streams     = num_streams;
        progress_data.progress.write_streams.completed_bytes   = 0;
        progress_data.progress.write_streams.completed_streams = 0;
        progress_data.progress.write_streams.num_threads       = num_threads;
        progress_data.progress.write_streams.compression_type  = out_ctype;
        progress_data.progress.write_streams.total_parts       = total_parts;
        progress_data.progress.write_streams.completed_parts   = 0;

        progress_data.next_progress = 0;
        progress_data.prev_wim_part = NULL;

#ifdef ENABLE_MULTITHREADED_COMPRESSION
        if (total_compression_bytes >= 2000000 && num_threads != 1)
                ret = write_stream_list_parallel(stream_list,
                                                 lookup_table,
                                                 out_fd,
                                                 out_ctype,
                                                 comp_ctx,
                                                 write_resource_flags,
                                                 &progress_data,
                                                 num_threads);
        else
#endif
                ret = write_stream_list_serial(stream_list,
                                               lookup_table,
                                               out_fd,
                                               out_ctype,
                                               comp_ctx,
                                               write_resource_flags,
                                               &progress_data);
        if (ret == 0)
                DEBUG("Successfully wrote stream list.");
        else
                DEBUG("Failed to write stream list.");
        return ret;
}

struct stream_size_table {
        struct hlist_head *array;
        size_t num_entries;
        size_t capacity;
};

static int
init_stream_size_table(struct stream_size_table *tab, size_t capacity)
{
        tab->array = CALLOC(capacity, sizeof(tab->array[0]));
        if (!tab->array)
                return WIMLIB_ERR_NOMEM;
        tab->num_entries = 0;
        tab->capacity = capacity;
        return 0;
}

static void
destroy_stream_size_table(struct stream_size_table *tab)
{
        FREE(tab->array);
}

static int
stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
{
        struct stream_size_table *tab = _tab;
        size_t pos;
        struct wim_lookup_table_entry *same_size_lte;
        struct hlist_node *tmp;

        pos = hash_u64(wim_resource_size(lte)) % tab->capacity;
        lte->unique_size = 1;
        hlist_for_each_entry(same_size_lte, tmp, &tab->array[pos], hash_list_2) {
                if (wim_resource_size(same_size_lte) == wim_resource_size(lte)) {
                        lte->unique_size = 0;
                        same_size_lte->unique_size = 0;
                        break;
                }
        }

        hlist_add_head(&lte->hash_list_2, &tab->array[pos]);
        tab->num_entries++;
        return 0;
}

struct find_streams_ctx {
        WIMStruct *wim;
        int write_flags;
        struct list_head stream_list;
        struct stream_size_table stream_size_tab;
};

static void
lte_reference_for_logical_write(struct wim_lookup_table_entry *lte,
                                struct find_streams_ctx *ctx,
                                unsigned nref)
{
        if (lte->out_refcnt == 0) {
                stream_size_table_insert(lte, &ctx->stream_size_tab);
                list_add_tail(&lte->write_streams_list, &ctx->stream_list);
        }
        lte->out_refcnt += nref;
}

static int
do_lte_full_reference_for_logical_write(struct wim_lookup_table_entry *lte,
                                        void *_ctx)
{
        struct find_streams_ctx *ctx = _ctx;
        lte->out_refcnt = 0;
        lte_reference_for_logical_write(lte, ctx,
                                        (lte->refcnt ? lte->refcnt : 1));
        return 0;
}

static int
inode_find_streams_to_write(struct wim_inode *inode,
                            struct wim_lookup_table *table,
                            struct find_streams_ctx *ctx)
{
        struct wim_lookup_table_entry *lte;
        unsigned i;

        for (i = 0; i <= inode->i_num_ads; i++) {
                lte = inode_stream_lte(inode, i, table);
                if (lte)
                        lte_reference_for_logical_write(lte, ctx, inode->i_nlink);
                else if (!is_zero_hash(inode_stream_hash(inode, i)))
                        return WIMLIB_ERR_RESOURCE_NOT_FOUND;
        }
        return 0;
}

static int
image_find_streams_to_write(WIMStruct *wim)
{
        struct find_streams_ctx *ctx;
        struct wim_image_metadata *imd;
        struct wim_inode *inode;
        struct wim_lookup_table_entry *lte;
        int ret;

        ctx = wim->private;
        imd = wim_get_current_image_metadata(wim);

        image_for_each_unhashed_stream(lte, imd)
                lte->out_refcnt = 0;

        /* Go through this image's inodes to find any streams that have not been
         * found yet. */
        image_for_each_inode(inode, imd) {
                ret = inode_find_streams_to_write(inode, wim->lookup_table, ctx);
                if (ret)
                        return ret;
        }
        return 0;
}

/*
 * Build a list of streams (via `struct wim_lookup_table_entry's) included in
 * the "logical write" of the WIM, meaning all streams that are referenced at
 * least once by dentries in the the image(s) being written.  'out_refcnt' on
 * each stream being included in the logical write is set to the number of
 * references from dentries in the image(s).  Furthermore, 'unique_size' on each
 * stream being included in the logical write is set to indicate whether that
 * stream has a unique size relative to the streams being included in the
 * logical write.  Still furthermore, 'part_number' on each stream being
 * included in the logical write is set to the part number given in the
 * in-memory header of @p wim.
 *
 * This is considered a "logical write" because it does not take into account
 * filtering out streams already present in the WIM (in the case of an in place
 * overwrite) or present in other WIMs (in case of creating delta WIM).
 */
static int
prepare_logical_stream_list(WIMStruct *wim, int image, bool streams_ok,
                            struct find_streams_ctx *ctx)
{
        int ret;
        struct wim_lookup_table_entry *lte;

        if (streams_ok && (image == WIMLIB_ALL_IMAGES ||
                           (image == 1 && wim->hdr.image_count == 1)))
        {
                /* Fast case:  Assume that all streams are being written and
                 * that the reference counts are correct.  */
                struct wim_lookup_table_entry *lte;
                struct wim_image_metadata *imd;
                unsigned i;

                for_lookup_table_entry(wim->lookup_table,
                                       do_lte_full_reference_for_logical_write, ctx);
                for (i = 0; i < wim->hdr.image_count; i++) {
                        imd = wim->image_metadata[i];
                        image_for_each_unhashed_stream(lte, imd)
                                do_lte_full_reference_for_logical_write(lte, ctx);
                }
        } else {
                /* Slow case:  Walk through the images being written and
                 * determine the streams referenced.  */
                for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
                wim->private = ctx;
                ret = for_image(wim, image, image_find_streams_to_write);
                if (ret)
                        return ret;
        }

        list_for_each_entry(lte, &ctx->stream_list, write_streams_list)
                lte->part_number = wim->hdr.part_number;
        return 0;
}

static int
process_filtered_stream(struct wim_lookup_table_entry *lte, void *_ctx)
{
        struct find_streams_ctx *ctx = _ctx;
        u16 filtered = 0;

        /* Calculate and set lte->filtered.  */
        if (lte->resource_location == RESOURCE_IN_WIM) {
                if (lte->wim == ctx->wim &&
                    (ctx->write_flags & WIMLIB_WRITE_FLAG_OVERWRITE))
                        filtered |= FILTERED_SAME_WIM;
                if (lte->wim != ctx->wim &&
                    (ctx->write_flags & WIMLIB_WRITE_FLAG_SKIP_EXTERNAL_WIMS))
                        filtered |= FILTERED_EXTERNAL_WIM;
        }
        lte->filtered = filtered;

        /* Filtered streams get inserted into the stream size table too, unless
         * they already were.  This is because streams that are checksummed
         * on-the-fly during the write should not be written if they are
         * duplicates of filtered stream.  */
        if (lte->filtered && lte->out_refcnt == 0)
                stream_size_table_insert(lte, &ctx->stream_size_tab);
        return 0;
}

static int
mark_stream_not_filtered(struct wim_lookup_table_entry *lte, void *_ignore)
{
        lte->filtered = 0;
        return 0;
}

/* Given the list of streams to include in a logical write of a WIM, handle
 * filtering out streams already present in the WIM or already present in
 * external WIMs, depending on the write flags provided.  */
static void
handle_stream_filtering(struct find_streams_ctx *ctx)
{
        struct wim_lookup_table_entry *lte, *tmp;

        if (!(ctx->write_flags & (WIMLIB_WRITE_FLAG_OVERWRITE |
                                  WIMLIB_WRITE_FLAG_SKIP_EXTERNAL_WIMS)))
        {
                for_lookup_table_entry(ctx->wim->lookup_table,
                                       mark_stream_not_filtered, ctx);
                return;
        }

        for_lookup_table_entry(ctx->wim->lookup_table,
                               process_filtered_stream, ctx);

        /* Streams in logical write list that were filtered can be removed.  */
        list_for_each_entry_safe(lte, tmp, &ctx->stream_list,
                                 write_streams_list)
                if (lte->filtered)
                        list_del(&lte->write_streams_list);
}

/* Prepares list of streams to write for the specified WIM image(s).  This wraps
 * around prepare_logical_stream_list() to handle filtering out streams already
 * present in the WIM or already present in external WIMs, depending on the
 * write flags provided.
 *
 * Note: some additional data is stored in each `struct wim_lookup_table_entry':
 *
 * - 'out_refcnt' is set to the number of references found for the logical write.
 *    This will be nonzero on all streams in the list returned by this function,
 *    but will also be nonzero on streams not in the list that were included in
 *    the logical write list, but filtered out from the returned list.
 * - 'filtered' is set to nonzero if the stream was filtered.  Filtered streams
 *   are not included in the list of streams returned by this function.
 * - 'unique_size' is set if the stream has a unique size among all streams in
 *   the logical write plus any filtered streams in the entire WIM that could
 *   potentially turn out to have the same checksum as a yet-to-be-checksummed
 *   stream being written.
 */
static int
prepare_stream_list(WIMStruct *wim, int image, int write_flags,
                    struct list_head *stream_list)
{
        int ret;
        bool streams_ok;
        struct find_streams_ctx ctx;

        INIT_LIST_HEAD(&ctx.stream_list);
        ret = init_stream_size_table(&ctx.stream_size_tab,
                                     wim->lookup_table->capacity);
        if (ret)
                return ret;
        ctx.write_flags = write_flags;
        ctx.wim = wim;

        streams_ok = ((write_flags & WIMLIB_WRITE_FLAG_STREAMS_OK) != 0);

        ret = prepare_logical_stream_list(wim, image, streams_ok, &ctx);
        if (ret)
                goto out_destroy_table;

        handle_stream_filtering(&ctx);
        list_transfer(&ctx.stream_list, stream_list);
        ret = 0;
out_destroy_table:
        destroy_stream_size_table(&ctx.stream_size_tab);
        return ret;
}

static int
write_wim_streams(WIMStruct *wim, int image, int write_flags,
                  unsigned num_threads,
                  wimlib_progress_func_t progress_func,
                  struct list_head *stream_list_override)
{
        int ret;
        struct list_head _stream_list;
        struct list_head *stream_list;
        struct wim_lookup_table_entry *lte;

        if (stream_list_override == NULL) {
                /* Normal case: prepare stream list from image(s) being written.
                 */
                stream_list = &_stream_list;
                ret = prepare_stream_list(wim, image, write_flags, stream_list);
                if (ret)
                        return ret;
        } else {
                /* Currently only as a result of wimlib_split() being called:
                 * use stream list already explicitly provided.  Use existing
                 * reference counts.  */
                stream_list = stream_list_override;
                list_for_each_entry(lte, stream_list, write_streams_list) {
                        lte->out_refcnt = (lte->refcnt ? lte->refcnt : 1);
                        lte->part_number = wim->hdr.part_number;
                }
        }

        return write_stream_list(stream_list,
                                 wim->lookup_table,
                                 &wim->out_fd,
                                 wim->compression_type,
                                 &wim->lzx_context,
                                 write_flags,
                                 num_threads,
                                 progress_func);
}

static int
write_wim_metadata_resources(WIMStruct *wim, int image, int write_flags,
                             wimlib_progress_func_t progress_func)
{
        int ret;
        int start_image;
        int end_image;
        int write_resource_flags;

        if (write_flags & WIMLIB_WRITE_FLAG_NO_METADATA) {
                DEBUG("Not writing any metadata resources.");
                return 0;
        }

        write_resource_flags = write_flags_to_resource_flags(write_flags);

        DEBUG("Writing metadata resources (offset=%"PRIu64")",
              wim->out_fd.offset);

        if (progress_func)
                progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);

        if (image == WIMLIB_ALL_IMAGES) {
                start_image = 1;
                end_image = wim->hdr.image_count;
        } else {
                start_image = image;
                end_image = image;
        }

        for (int i = start_image; i <= end_image; i++) {
                struct wim_image_metadata *imd;

                imd = wim->image_metadata[i - 1];
                /* Build a new metadata resource only if image was modified from
                 * the original (or was newly added).  Otherwise just copy the
                 * existing one.  */
                if (imd->modified) {
                        DEBUG("Image %u was modified; building and writing new "
                              "metadata resource", i);
                        ret = write_metadata_resource(wim, i,
                                                      write_resource_flags);
                } else if (write_flags & WIMLIB_WRITE_FLAG_OVERWRITE) {
                        DEBUG("Image %u was not modified; re-using existing "
                              "metadata resource.", i);
                        copy_resource_entry(&imd->metadata_lte->output_resource_entry,
                                            &imd->metadata_lte->resource_entry);
                        ret = 0;
                } else {
                        DEBUG("Image %u was not modified; copying existing "
                              "metadata resource.", i);
                        ret = write_wim_resource(imd->metadata_lte,
                                                 &wim->out_fd,
                                                 wim->compression_type,
                                                 &imd->metadata_lte->output_resource_entry,
                                                 write_resource_flags,
                                                 &wim->lzx_context);
                }
                if (ret)
                        return ret;
        }
        if (progress_func)
                progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
        return 0;
}

static int
open_wim_writable(WIMStruct *wim, const tchar *path, int open_flags)
{
        int raw_fd;
        DEBUG("Opening \"%"TS"\" for writing.", path);

        raw_fd = topen(path, open_flags | O_BINARY, 0644);
        if (raw_fd < 0) {
                ERROR_WITH_ERRNO("Failed to open \"%"TS"\" for writing", path);
                return WIMLIB_ERR_OPEN;
        }
        filedes_init(&wim->out_fd, raw_fd);
        return 0;
}

static int
close_wim_writable(WIMStruct *wim, int write_flags)
{
        int ret = 0;

        if (!(write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR))
                if (filedes_valid(&wim->out_fd))
                        if (filedes_close(&wim->out_fd))
                                ret = WIMLIB_ERR_WRITE;
        filedes_invalidate(&wim->out_fd);
        return ret;
}

/*
 * finish_write():
 *
 * Finish writing a WIM file: write the lookup table, xml data, and integrity
 * table, then overwrite the WIM header.  By default, closes the WIM file
 * descriptor (@wim->out_fd) if successful.
 *
 * write_flags is a bitwise OR of the following:
 *
 *      (public) WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
 *              Include an integrity table.
 *
 *      (public) WIMLIB_WRITE_FLAG_FSYNC:
 *              fsync() the output file before closing it.
 *
 *      (public) WIMLIB_WRITE_FLAG_PIPABLE:
 *              Writing a pipable WIM, possibly to a pipe; include pipable WIM
 *              stream headers before the lookup table and XML data, and also
 *              write the WIM header at the end instead of seeking to the
 *              beginning.  Can't be combined with
 *              WIMLIB_WRITE_FLAG_CHECK_INTEGRITY.
 *
 *      (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
 *              Don't write the lookup table.
 *
 *      (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
 *              When (if) writing the integrity table, re-use entries from the
 *              existing integrity table, if possible.
 *
 *      (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
 *              After writing the XML data but before writing the integrity
 *              table, write a temporary WIM header and flush the stream so that
 *              the WIM is less likely to become corrupted upon abrupt program
 *              termination.
 *      (private) WIMLIB_WRITE_FLAG_HEADER_AT_END:
 *              Instead of overwriting the WIM header at the beginning of the
 *              file, simply append it to the end of the file.  (Used when
 *              writing to pipe.)
 *      (private) WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR:
 *              Do not close the file descriptor @wim->out_fd on either success
 *              on failure.
 *      (private) WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES:
 *              Use the existing <TOTALBYTES> stored in the in-memory XML
 *              information, rather than setting it to the offset of the XML
 *              data being written.
 */
static int
finish_write(WIMStruct *wim, int image, int write_flags,
             wimlib_progress_func_t progress_func,
             struct list_head *stream_list_override)
{
        int ret;
        off_t hdr_offset;
        int write_resource_flags;
        off_t old_lookup_table_end;
        off_t new_lookup_table_end;
        u64 xml_totalbytes;

        write_resource_flags = write_flags_to_resource_flags(write_flags);

        /* In the WIM header, there is room for the resource entry for a
         * metadata resource labeled as the "boot metadata".  This entry should
         * be zeroed out if there is no bootable image (boot_idx 0).  Otherwise,
         * it should be a copy of the resource entry for the image that is
         * marked as bootable.  This is not well documented...  */
        if (wim->hdr.boot_idx == 0) {
                zero_resource_entry(&wim->hdr.boot_metadata_res_entry);
        } else {
                copy_resource_entry(&wim->hdr.boot_metadata_res_entry,
                            &wim->image_metadata[wim->hdr.boot_idx- 1
                                        ]->metadata_lte->output_resource_entry);
        }

        /* Write lookup table.  (Save old position first.)  */
        old_lookup_table_end = wim->hdr.lookup_table_res_entry.offset +
                               wim->hdr.lookup_table_res_entry.size;
        if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
                ret = write_wim_lookup_table(wim, image, write_flags,
                                             &wim->hdr.lookup_table_res_entry,
                                             stream_list_override);
                if (ret)
                        return ret;
        }

        /* Write XML data.  */
        xml_totalbytes = wim->out_fd.offset;
        if (write_flags & WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES)
                xml_totalbytes = WIM_TOTALBYTES_USE_EXISTING;
        ret = write_wim_xml_data(wim, image, xml_totalbytes,
                                 &wim->hdr.xml_res_entry,
                                 write_resource_flags);
        if (ret)
                return ret;

        /* Write integrity table (optional).  */
        if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
                if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
                        struct wim_header checkpoint_hdr;
                        memcpy(&checkpoint_hdr, &wim->hdr, sizeof(struct wim_header));
                        zero_resource_entry(&checkpoint_hdr.integrity);
                        checkpoint_hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
                        ret = write_wim_header_at_offset(&checkpoint_hdr,
                                                         &wim->out_fd, 0);
                        if (ret)
                                return ret;
                }

                if (!(write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE))
                        old_lookup_table_end = 0;

                new_lookup_table_end = wim->hdr.lookup_table_res_entry.offset +
                                       wim->hdr.lookup_table_res_entry.size;

                ret = write_integrity_table(wim,
                                            new_lookup_table_end,
                                            old_lookup_table_end,
                                            progress_func);
                if (ret)
                        return ret;
        } else {
                /* No integrity table.  */
                zero_resource_entry(&wim->hdr.integrity);
        }

        /* Now that all information in the WIM header has been determined, the
         * preliminary header written earlier can be overwritten, the header of
         * the existing WIM file can be overwritten, or the final header can be
         * written to the end of the pipable WIM.  */
        wim->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
        hdr_offset = 0;
        if (write_flags & WIMLIB_WRITE_FLAG_HEADER_AT_END)
                hdr_offset = wim->out_fd.offset;
        ret = write_wim_header_at_offset(&wim->hdr, &wim->out_fd, hdr_offset);
        if (ret)
                return ret;

        /* Possibly sync file data to disk before closing.  On POSIX systems, it
         * is necessary to do this before using rename() to overwrite an
         * existing file with a new file.  Otherwise, data loss would occur if
         * the system is abruptly terminated when the metadata for the rename
         * operation has been written to disk, but the new file data has not.
         */
        if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
                if (fsync(wim->out_fd.fd)) {
                        ERROR_WITH_ERRNO("Error syncing data to WIM file");
                        return WIMLIB_ERR_WRITE;
                }
        }

        if (close_wim_writable(wim, write_flags)) {
                ERROR_WITH_ERRNO("Failed to close the output WIM file");
                return WIMLIB_ERR_WRITE;
        }

        return 0;
}

#if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
int
lock_wim(WIMStruct *wim, int fd)
{
        int ret = 0;
        if (fd != -1 && !wim->wim_locked) {
                ret = flock(fd, LOCK_EX | LOCK_NB);
                if (ret != 0) {
                        if (errno == EWOULDBLOCK) {
                                ERROR("`%"TS"' is already being modified or has been "
                                      "mounted read-write\n"
                                      "        by another process!", wim->filename);
                                ret = WIMLIB_ERR_ALREADY_LOCKED;
                        } else {
                                WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
                                                   wim->filename);
                                ret = 0;
                        }
                } else {
                        wim->wim_locked = 1;
                }
        }
        return ret;
}
#endif

/*
 * write_pipable_wim():
 *
 * Perform the intermediate stages of creating a "pipable" WIM (i.e. a WIM
 * capable of being applied from a pipe).
 *
 * Pipable WIMs are a wimlib-specific modification of the WIM format such that
 * images can be applied from them sequentially when the file data is sent over
 * a pipe.  In addition, a pipable WIM can be written sequentially to a pipe.
 * The modifications made to the WIM format for pipable WIMs are:
 *
 * - Magic characters in header are "WLPWM\0\0\0" (wimlib pipable WIM) instead
 *   of "MSWIM\0\0\0".  This lets wimlib know that the WIM is pipable and also
 *   stops other software from trying to read the file as a normal WIM.
 *
 * - The header at the beginning of the file does not contain all the normal
 *   information; in particular it will have all 0's for the lookup table and
 *   XML data resource entries.  This is because this information cannot be
 *   determined until the lookup table and XML data have been written.
 *   Consequently, wimlib will write the full header at the very end of the
 *   file.  The header at the end, however, is only used when reading the WIM
 *   from a seekable file (not a pipe).
 *
 * - An extra copy of the XML data is placed directly after the header.  This
 *   allows image names and sizes to be determined at an appropriate time when
 *   reading the WIM from a pipe.  This copy of the XML data is ignored if the
 *   WIM is read from a seekable file (not a pipe).
 *
 * - The format of resources, or streams, has been modified to allow them to be
 *   used before the "lookup table" has been read.  Each stream is prefixed with
 *   a `struct pwm_stream_hdr' that is basically an abbreviated form of `struct
 *   wim_lookup_table_entry_disk' that only contains the SHA1 message digest,
 *   uncompressed stream size, and flags that indicate whether the stream is
 *   compressed.  The data of uncompressed streams then follows literally, while
 *   the data of compressed streams follows in a modified format.  Compressed
 *   streams do not begin with a chunk table, since the chunk table cannot be
 *   written until all chunks have been compressed.  Instead, each compressed
 *   chunk is prefixed by a `struct pwm_chunk_hdr' that gives its size.
 *   Furthermore, the chunk table is written at the end of the resource instead
 *   of the start.  Note: chunk offsets are given in the chunk table as if the
 *   `struct pwm_chunk_hdr's were not present; also, the chunk table is only
 *   used if the WIM is being read from a seekable file (not a pipe).
 *
 * - Metadata resources always come before other file resources (streams).
 *   (This does not by itself constitute an incompatibility with normal WIMs,
 *   since this is valid in normal WIMs.)
 *
 * - At least up to the end of the file resources, all components must be packed
 *   as tightly as possible; there cannot be any "holes" in the WIM.  (This does
 *   not by itself consititute an incompatibility with normal WIMs, since this
 *   is valid in normal WIMs.)
 *
 * Note: the lookup table, XML data, and header at the end are not used when
 * applying from a pipe.  They exist to support functionality such as image
 * application and export when the WIM is *not* read from a pipe.
 *
 *   Layout of pipable WIM:
 *
 * ---------+----------+--------------------+----------------+--------------+-----------+--------+
 * | Header | XML data | Metadata resources | File resources | Lookup table | XML data  | Header |
 * ---------+----------+--------------------+----------------+--------------+-----------+--------+
 *
 *   Layout of normal WIM:
 *
 * +--------+-----------------------------+-------------------------+
 * | Header | File and metadata resources | Lookup table | XML data |
 * +--------+-----------------------------+-------------------------+
 *
 * An optional integrity table can follow the final XML data in both normal and
 * pipable WIMs.  However, due to implementation details, wimlib currently can
 * only include an integrity table in a pipable WIM when writing it to a
 * seekable file (not a pipe).
 *
 * Do note that since pipable WIMs are not supported by Microsoft's software,
 * wimlib does not create them unless explicitly requested (with
 * WIMLIB_WRITE_FLAG_PIPABLE) and as stated above they use different magic
 * characters to identify the file.
 */
static int
write_pipable_wim(WIMStruct *wim, int image, int write_flags,
                  unsigned num_threads, wimlib_progress_func_t progress_func,
                  struct list_head *stream_list_override)
{
        int ret;
        struct resource_entry xml_res_entry;

        WARNING("Creating a pipable WIM, which will "
                "be incompatible\n"
                "          with Microsoft's software (wimgapi/imagex/Dism).");

        /* At this point, the header at the beginning of the file has already
         * been written.  */

        /* For efficiency, when wimlib adds an image to the WIM with
         * wimlib_add_image(), the SHA1 message digests of files is not
         * calculated; instead, they are calculated while the files are being
         * written.  However, this does not work when writing a pipable WIM,
         * since when writing a stream to a pipable WIM, its SHA1 message digest
         * needs to be known before the stream data is written.  Therefore,
         * before getting much farther, we need to pre-calculate the SHA1
         * message digests of all streams that will be written.  */
        ret = wim_checksum_unhashed_streams(wim);
        if (ret)
                return ret;

        /* Write extra copy of the XML data.  */
        ret = write_wim_xml_data(wim, image, WIM_TOTALBYTES_OMIT,
                                 &xml_res_entry,
                                 WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE);
        if (ret)
                return ret;

        /* Write metadata resources for the image(s) being included in the
         * output WIM.  */
        ret = write_wim_metadata_resources(wim, image, write_flags,
                                           progress_func);
        if (ret)
                return ret;

        /* Write streams needed for the image(s) being included in the output
         * WIM, or streams needed for the split WIM part.  */
        return write_wim_streams(wim, image, write_flags, num_threads,
                                 progress_func, stream_list_override);

        /* The lookup table, XML data, and header at end are handled by
         * finish_write().  */
}

/* Write a standalone WIM or split WIM (SWM) part to a new file or to a file
 * descriptor.  */
int
write_wim_part(WIMStruct *wim,
               const void *path_or_fd,
               int image,
               int write_flags,
               unsigned num_threads,
               wimlib_progress_func_t progress_func,
               unsigned part_number,
               unsigned total_parts,
               struct list_head *stream_list_override,
               const u8 *guid)
{
        int ret;
        struct wim_header hdr_save;
        struct list_head lt_stream_list_override;

        if (total_parts == 1)
                DEBUG("Writing standalone WIM.");
        else
                DEBUG("Writing split WIM part %u/%u", part_number, total_parts);
        if (image == WIMLIB_ALL_IMAGES)
                DEBUG("Including all images.");
        else
                DEBUG("Including image %d only.", image);
        if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
                DEBUG("File descriptor: %d", *(const int*)path_or_fd);
        else
                DEBUG("Path: \"%"TS"\"", (const tchar*)path_or_fd);
        DEBUG("Write flags: 0x%08x", write_flags);
        if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
                DEBUG("\tCHECK_INTEGRITY");
        if (write_flags & WIMLIB_WRITE_FLAG_REBUILD)
                DEBUG("\tREBUILD");
        if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
                DEBUG("\tRECOMPRESS");
        if (write_flags & WIMLIB_WRITE_FLAG_FSYNC)
                DEBUG("\tFSYNC");
        if (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE)
                DEBUG("\tFSYNC");
        if (write_flags & WIMLIB_WRITE_FLAG_IGNORE_READONLY_FLAG)
                DEBUG("\tIGNORE_READONLY_FLAG");
        if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
                DEBUG("\tPIPABLE");
        if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
                DEBUG("\tFILE_DESCRIPTOR");
        if (write_flags & WIMLIB_WRITE_FLAG_NO_METADATA)
                DEBUG("\tNO_METADATA");
        if (write_flags & WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES)
                DEBUG("\tUSE_EXISTING_TOTALBYTES");
        if (num_threads == 0)
                DEBUG("Number of threads: autodetect");
        else
                DEBUG("Number of threads: %u", num_threads);
        DEBUG("Progress function: %s", (progress_func ? "yes" : "no"));
        DEBUG("Stream list:       %s", (stream_list_override ? "specified" : "autodetect"));
        DEBUG("GUID:              %s", ((guid || wim->guid_set_explicitly) ?
                                        "specified" : "generate new"));

        /* Internally, this is always called with a valid part number and total
         * parts.  */
        wimlib_assert(total_parts >= 1);
        wimlib_assert(part_number >= 1 && part_number <= total_parts);

        /* A valid image (or all images) must be specified.  */
        if (image != WIMLIB_ALL_IMAGES &&
             (image < 1 || image > wim->hdr.image_count))
                return WIMLIB_ERR_INVALID_IMAGE;

        /* If we need to write metadata resources, make sure the ::WIMStruct has
         * the needed information attached (e.g. is not a resource-only WIM,
         * such as a non-first part of a split WIM).  */
        if (!wim_has_metadata(wim) &&
            !(write_flags & WIMLIB_WRITE_FLAG_NO_METADATA))
                return WIMLIB_ERR_METADATA_NOT_FOUND;

        /* Check for contradictory flags.  */
        if ((write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
                            WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY))
                                == (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
                                    WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY))
                return WIMLIB_ERR_INVALID_PARAM;

        if ((write_flags & (WIMLIB_WRITE_FLAG_PIPABLE |
                            WIMLIB_WRITE_FLAG_NOT_PIPABLE))
                                == (WIMLIB_WRITE_FLAG_PIPABLE |
                                    WIMLIB_WRITE_FLAG_NOT_PIPABLE))
                return WIMLIB_ERR_INVALID_PARAM;

        /* Save previous header, then start initializing the new one.  */
        memcpy(&hdr_save, &wim->hdr, sizeof(struct wim_header));

        /* Set default integrity and pipable flags.  */
        if (!(write_flags & (WIMLIB_WRITE_FLAG_PIPABLE |
                             WIMLIB_WRITE_FLAG_NOT_PIPABLE)))
                if (wim_is_pipable(wim))
                        write_flags |= WIMLIB_WRITE_FLAG_PIPABLE;

        if (!(write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
                             WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY)))
                if (wim_has_integrity_table(wim))
                        write_flags |= WIMLIB_WRITE_FLAG_CHECK_INTEGRITY;

        /* Set appropriate magic number.  */
        if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
                wim->hdr.magic = PWM_MAGIC;
        else
                wim->hdr.magic = WIM_MAGIC;

        /* Clear header flags that will be set automatically.  */
        wim->hdr.flags &= ~(WIM_HDR_FLAG_METADATA_ONLY          |
                            WIM_HDR_FLAG_RESOURCE_ONLY          |
                            WIM_HDR_FLAG_SPANNED                |
                            WIM_HDR_FLAG_WRITE_IN_PROGRESS);

        /* Set SPANNED header flag if writing part of a split WIM.  */
        if (total_parts != 1)
                wim->hdr.flags |= WIM_HDR_FLAG_SPANNED;

        /* Set part number and total parts of split WIM.  This will be 1 and 1
         * if the WIM is standalone.  */
        wim->hdr.part_number = part_number;
        wim->hdr.total_parts = total_parts;

        /* Use GUID if specified; otherwise generate a new one.  */
        if (guid)
                memcpy(wim->hdr.guid, guid, WIMLIB_GUID_LEN);
        else if (!wim->guid_set_explicitly)
                randomize_byte_array(wim->hdr.guid, WIMLIB_GUID_LEN);

        /* Clear references to resources that have not been written yet.  */
        zero_resource_entry(&wim->hdr.lookup_table_res_entry);
        zero_resource_entry(&wim->hdr.xml_res_entry);
        zero_resource_entry(&wim->hdr.boot_metadata_res_entry);
        zero_resource_entry(&wim->hdr.integrity);

        /* Set image count and boot index correctly for single image writes.  */
        if (image != WIMLIB_ALL_IMAGES) {
                wim->hdr.image_count = 1;
                if (wim->hdr.boot_idx == image)
                        wim->hdr.boot_idx = 1;
                else
                        wim->hdr.boot_idx = 0;
        }

        /* Split WIMs can't be bootable.  */
        if (total_parts != 1)
                wim->hdr.boot_idx = 0;

        /* Initialize output file descriptor.  */
        if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR) {
                /* File descriptor was explicitly provided.  Return error if
                 * file descriptor is not seekable, unless writing a pipable WIM
                 * was requested.  */
                wim->out_fd.fd = *(const int*)path_or_fd;
                wim->out_fd.offset = 0;
                if (!filedes_is_seekable(&wim->out_fd)) {
                        ret = WIMLIB_ERR_INVALID_PARAM;
                        if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE))
                                goto out_restore_hdr;
                        if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
                                ERROR("Can't include integrity check when "
                                      "writing pipable WIM to pipe!");
                                goto out_restore_hdr;
                        }
                }

        } else {
                /* Filename of WIM to write was provided; open file descriptor
                 * to it.  */
                ret = open_wim_writable(wim, (const tchar*)path_or_fd,
                                        O_TRUNC | O_CREAT | O_RDWR);
                if (ret)
                        goto out_restore_hdr;
        }

        /* Write initial header.  This is merely a "dummy" header since it
         * doesn't have all the information yet, so it will be overwritten later
         * (unless writing a pipable WIM).  */
        if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE))
                wim->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
        ret = write_wim_header(&wim->hdr, &wim->out_fd);
        wim->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
        if (ret)
                goto out_restore_hdr;

        if (stream_list_override) {
                struct wim_lookup_table_entry *lte;
                INIT_LIST_HEAD(&lt_stream_list_override);
                list_for_each_entry(lte, stream_list_override,
                                    write_streams_list)
                {
                        list_add_tail(&lte->lookup_table_list,
                                      &lt_stream_list_override);
                }
        }

        /* Write metadata resources and streams.  */
        if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE)) {
                /* Default case: create a normal (non-pipable) WIM.  */
                ret = write_wim_streams(wim, image, write_flags, num_threads,
                                        progress_func, stream_list_override);
                if (ret)
                        goto out_restore_hdr;

                ret = write_wim_metadata_resources(wim, image, write_flags,
                                                   progress_func);
                if (ret)
                        goto out_restore_hdr;
        } else {
                /* Non-default case: create pipable WIM.  */
                ret = write_pipable_wim(wim, image, write_flags, num_threads,
                                        progress_func, stream_list_override);
                if (ret)
                        goto out_restore_hdr;
                write_flags |= WIMLIB_WRITE_FLAG_HEADER_AT_END;
        }

        if (stream_list_override)
                stream_list_override = &lt_stream_list_override;

        /* Write lookup table, XML data, and (optional) integrity table.  */
        ret = finish_write(wim, image, write_flags, progress_func,
                           stream_list_override);
out_restore_hdr:
        memcpy(&wim->hdr, &hdr_save, sizeof(struct wim_header));
        (void)close_wim_writable(wim, write_flags);
        return ret;
}

/* Write a standalone WIM to a file or file descriptor.  */
static int
write_standalone_wim(WIMStruct *wim, const void *path_or_fd,
                     int image, int write_flags, unsigned num_threads,
                     wimlib_progress_func_t progress_func)
{
        return write_wim_part(wim, path_or_fd, image, write_flags,
                              num_threads, progress_func, 1, 1, NULL, NULL);
}

/* API function documented in wimlib.h  */
WIMLIBAPI int
wimlib_write(WIMStruct *wim, const tchar *path,
             int image, int write_flags, unsigned num_threads,
             wimlib_progress_func_t progress_func)
{
        if (!path)
                return WIMLIB_ERR_INVALID_PARAM;

        write_flags &= WIMLIB_WRITE_MASK_PUBLIC;

        return write_standalone_wim(wim, path, image, write_flags,
                                    num_threads, progress_func);
}

/* API function documented in wimlib.h  */
WIMLIBAPI int
wimlib_write_to_fd(WIMStruct *wim, int fd,
                   int image, int write_flags, unsigned num_threads,
                   wimlib_progress_func_t progress_func)
{
        if (fd < 0)
                return WIMLIB_ERR_INVALID_PARAM;

        write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
        write_flags |= WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR;

        return write_standalone_wim(wim, &fd, image, write_flags,
                                    num_threads, progress_func);
}

static bool
any_images_modified(WIMStruct *wim)
{
        for (int i = 0; i < wim->hdr.image_count; i++)
                if (wim->image_metadata[i]->modified)
                        return true;
        return false;
}

static int
check_resource_offset(struct wim_lookup_table_entry *lte, void *_wim)
{
        const WIMStruct *wim = _wim;
        off_t end_offset = *(const off_t*)wim->private;

        if (lte->resource_location == RESOURCE_IN_WIM && lte->wim == wim &&
            lte->resource_entry.offset + lte->resource_entry.size > end_offset)
                return WIMLIB_ERR_RESOURCE_ORDER;
        return 0;
}

/* Make sure no file or metadata resources are located after the XML data (or
 * integrity table if present)--- otherwise we can't safely overwrite the WIM in
 * place and we return WIMLIB_ERR_RESOURCE_ORDER.  */
static int
check_resource_offsets(WIMStruct *wim, off_t end_offset)
{
        int ret;
        unsigned i;

        wim->private = &end_offset;
        ret = for_lookup_table_entry(wim->lookup_table, check_resource_offset, wim);
        if (ret)
                return ret;

        for (i = 0; i < wim->hdr.image_count; i++) {
                ret = check_resource_offset(wim->image_metadata[i]->metadata_lte, wim);
                if (ret)
                        return ret;
        }
        return 0;
}

/*
 * Overwrite a WIM, possibly appending streams to it.
 *
 * A WIM looks like (or is supposed to look like) the following:
 *
 *                   Header (212 bytes)
 *                   Streams and metadata resources (variable size)
 *                   Lookup table (variable size)
 *                   XML data (variable size)
 *                   Integrity table (optional) (variable size)
 *
 * If we are not adding any streams or metadata resources, the lookup table is
 * unchanged--- so we only need to overwrite the XML data, integrity table, and
 * header.  This operation is potentially unsafe if the program is abruptly
 * terminated while the XML data or integrity table are being overwritten, but
 * before the new header has been written.  To partially alleviate this problem,
 * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
 * finish_write() to cause a temporary WIM header to be written after the XML
 * data has been written.  This may prevent the WIM from becoming corrupted if
 * the program is terminated while the integrity table is being calculated (but
 * no guarantees, due to write re-ordering...).
 *
 * If we are adding new streams or images (metadata resources), the lookup table
 * needs to be changed, and those streams need to be written.  In this case, we
 * try to perform a safe update of the WIM file by writing the streams *after*
 * the end of the previous WIM, then writing the new lookup table, XML data, and
 * (optionally) integrity table following the new streams.  This will produce a
 * layout like the following:
 *
 *                   Header (212 bytes)
 *                   (OLD) Streams and metadata resources (variable size)
 *                   (OLD) Lookup table (variable size)
 *                   (OLD) XML data (variable size)
 *                   (OLD) Integrity table (optional) (variable size)
 *                   (NEW) Streams and metadata resources (variable size)
 *                   (NEW) Lookup table (variable size)
 *                   (NEW) XML data (variable size)
 *                   (NEW) Integrity table (optional) (variable size)
 *
 * At all points, the WIM is valid as nothing points to the new data yet.  Then,
 * the header is overwritten to point to the new lookup table, XML data, and
 * integrity table, to produce the following layout:
 *
 *                   Header (212 bytes)
 *                   Streams and metadata resources (variable size)
 *                   Nothing (variable size)
 *                   More Streams and metadata resources (variable size)
 *                   Lookup table (variable size)
 *                   XML data (variable size)
 *                   Integrity table (optional) (variable size)
 *
 * This method allows an image to be appended to a large WIM very quickly, and
 * is is crash-safe except in the case of write re-ordering, but the
 * disadvantage is that a small hole is left in the WIM where the old lookup
 * table, xml data, and integrity table were.  (These usually only take up a
 * small amount of space compared to the streams, however.)
 */
static int
overwrite_wim_inplace(WIMStruct *wim, int write_flags,
                      unsigned num_threads,
                      wimlib_progress_func_t progress_func)
{
        int ret;
        struct list_head stream_list;
        off_t old_wim_end;
        u64 old_lookup_table_end, old_xml_begin, old_xml_end;
        struct wim_header hdr_save;

        DEBUG("Overwriting `%"TS"' in-place", wim->filename);

        /* Set default integrity flag.  */
        if (!(write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
                             WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY)))
                if (wim_has_integrity_table(wim))
                        write_flags |= WIMLIB_WRITE_FLAG_CHECK_INTEGRITY;

        /* Set additional flags for overwrite.  */
        write_flags |= WIMLIB_WRITE_FLAG_OVERWRITE |
                       WIMLIB_WRITE_FLAG_STREAMS_OK;

        /* Make sure that the integrity table (if present) is after the XML
         * data, and that there are no stream resources, metadata resources, or
         * lookup tables after the XML data.  Otherwise, these data would be
         * overwritten. */
        old_xml_begin = wim->hdr.xml_res_entry.offset;
        old_xml_end = old_xml_begin + wim->hdr.xml_res_entry.size;
        old_lookup_table_end = wim->hdr.lookup_table_res_entry.offset +
                               wim->hdr.lookup_table_res_entry.size;
        if (wim->hdr.integrity.offset != 0 && wim->hdr.integrity.offset < old_xml_end) {
                WARNING("Didn't expect the integrity table to be before the XML data");
                return WIMLIB_ERR_RESOURCE_ORDER;
        }

        if (old_lookup_table_end > old_xml_begin) {
                WARNING("Didn't expect the lookup table to be after the XML data");
                return WIMLIB_ERR_RESOURCE_ORDER;
        }

        /* Set @old_wim_end, which indicates the point beyond which we don't
         * allow any file and metadata resources to appear without returning
         * WIMLIB_ERR_RESOURCE_ORDER (due to the fact that we would otherwise
         * overwrite these resources). */
        if (!wim->deletion_occurred && !any_images_modified(wim)) {
                /* If no images have been modified and no images have been
                 * deleted, a new lookup table does not need to be written.  We
                 * shall write the new XML data and optional integrity table
                 * immediately after the lookup table.  Note that this may
                 * overwrite an existing integrity table. */
                DEBUG("Skipping writing lookup table "
                      "(no images modified or deleted)");
                old_wim_end = old_lookup_table_end;
                write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
                               WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
        } else if (wim->hdr.integrity.offset) {
                /* Old WIM has an integrity table; begin writing new streams
                 * after it. */
                old_wim_end = wim->hdr.integrity.offset + wim->hdr.integrity.size;
        } else {
                /* No existing integrity table; begin writing new streams after
                 * the old XML data. */
                old_wim_end = old_xml_end;
        }

        ret = check_resource_offsets(wim, old_wim_end);
        if (ret)
                return ret;

        ret = prepare_stream_list(wim, WIMLIB_ALL_IMAGES, write_flags,
                                  &stream_list);
        if (ret)
                return ret;

        ret = open_wim_writable(wim, wim->filename, O_RDWR);
        if (ret)
                return ret;

        ret = lock_wim(wim, wim->out_fd.fd);
        if (ret)
                goto out_close_wim;

        /* Save original header so it can be restored in case of error  */
        memcpy(&hdr_save, &wim->hdr, sizeof(struct wim_header));

        /* Set WIM_HDR_FLAG_WRITE_IN_PROGRESS flag in header. */
        wim->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
        ret = write_wim_header_flags(wim->hdr.flags, &wim->out_fd);
        if (ret) {
                ERROR_WITH_ERRNO("Error updating WIM header flags");
                goto out_restore_memory_hdr;
        }

        if (filedes_seek(&wim->out_fd, old_wim_end) == -1) {
                ERROR_WITH_ERRNO("Can't seek to end of WIM");
                ret = WIMLIB_ERR_WRITE;
                goto out_restore_physical_hdr;
        }

        ret = write_stream_list(&stream_list,
                                wim->lookup_table,
                                &wim->out_fd,
                                wim->compression_type,
                                &wim->lzx_context,
                                write_flags,
                                num_threads,
                                progress_func);
        if (ret)
                goto out_truncate;

        ret = write_wim_metadata_resources(wim, WIMLIB_ALL_IMAGES,
                                           write_flags, progress_func);
        if (ret)
                goto out_truncate;

        write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
        ret = finish_write(wim, WIMLIB_ALL_IMAGES, write_flags,
                           progress_func, NULL);
        if (ret)
                goto out_truncate;

        goto out_unlock_wim;

out_truncate:
        if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
                WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
                        wim->filename, old_wim_end);
                /* Return value of ftruncate() is ignored because this is
                 * already an error path.  */
                (void)ftruncate(wim->out_fd.fd, old_wim_end);
        }
out_restore_physical_hdr:
        (void)write_wim_header_flags(hdr_save.flags, &wim->out_fd);
out_restore_memory_hdr:
        memcpy(&wim->hdr, &hdr_save, sizeof(struct wim_header));
out_close_wim:
        (void)close_wim_writable(wim, write_flags);
out_unlock_wim:
        wim->wim_locked = 0;
        return ret;
}

static int
overwrite_wim_via_tmpfile(WIMStruct *wim, int write_flags,
                          unsigned num_threads,
                          wimlib_progress_func_t progress_func)
{
        size_t wim_name_len;
        int ret;

        DEBUG("Overwriting `%"TS"' via a temporary file", wim->filename);

        /* Write the WIM to a temporary file in the same directory as the
         * original WIM. */
        wim_name_len = tstrlen(wim->filename);
        tchar tmpfile[wim_name_len + 10];
        tmemcpy(tmpfile, wim->filename, wim_name_len);
        randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
        tmpfile[wim_name_len + 9] = T('\0');

        ret = wimlib_write(wim, tmpfile, WIMLIB_ALL_IMAGES,
                           write_flags | WIMLIB_WRITE_FLAG_FSYNC,
                           num_threads, progress_func);
        if (ret) {
                tunlink(tmpfile);
                return ret;
        }

        close_wim(wim);

        /* Rename the new WIM file to the original WIM file.  Note: on Windows
         * this actually calls win32_rename_replacement(), not _wrename(), so
         * that removing the existing destination file can be handled.  */
        DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, wim->filename);
        ret = trename(tmpfile, wim->filename);
        if (ret) {
                ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
                                 tmpfile, wim->filename);
        #ifdef __WIN32__
                if (ret < 0)
        #endif
                {
                        tunlink(tmpfile);
                }
                return WIMLIB_ERR_RENAME;
        }

        if (progress_func) {
                union wimlib_progress_info progress;
                progress.rename.from = tmpfile;
                progress.rename.to = wim->filename;
                progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
        }
        return 0;
}

/* API function documented in wimlib.h  */
WIMLIBAPI int
wimlib_overwrite(WIMStruct *wim, int write_flags,
                 unsigned num_threads,
                 wimlib_progress_func_t progress_func)
{
        int ret;
        u32 orig_hdr_flags;

        write_flags &= WIMLIB_WRITE_MASK_PUBLIC;

        if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
                return WIMLIB_ERR_INVALID_PARAM;

        if (!wim->filename)
                return WIMLIB_ERR_NO_FILENAME;

        orig_hdr_flags = wim->hdr.flags;
        if (write_flags & WIMLIB_WRITE_FLAG_IGNORE_READONLY_FLAG)
                wim->hdr.flags &= ~WIM_HDR_FLAG_READONLY;
        ret = can_modify_wim(wim);
        wim->hdr.flags = orig_hdr_flags;
        if (ret)
                return ret;

        if ((!wim->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
            && !(write_flags & (WIMLIB_WRITE_FLAG_REBUILD |
                                WIMLIB_WRITE_FLAG_PIPABLE))
            && !(wim_is_pipable(wim)))
        {
                ret = overwrite_wim_inplace(wim, write_flags, num_threads,
                                            progress_func);
                if (ret != WIMLIB_ERR_RESOURCE_ORDER)
                        return ret;
                WARNING("Falling back to re-building entire WIM");
        }
        return overwrite_wim_via_tmpfile(wim, write_flags, num_threads,
                                         progress_func);
}