]> wimlib.net Git - wimlib/blobdiff - src/write.c
write_pipable_wim(): Fix/improve comment
[wimlib] / src / write.c
index b1d8708f63a8c391f7857ad63e621a9603c02769..1a429039b38dd80fb3141705ef8d1a5aa8c7ac84 100644 (file)
 #  include <pthread.h>
 #endif
 
-#include <unistd.h>
-#include <fcntl.h>
 #include <errno.h>
-
-#ifdef WITH_NTFS_3G
-#  include <time.h>
-#  include <ntfs-3g/attrib.h>
-#  include <ntfs-3g/inode.h>
-#  include <ntfs-3g/dir.h>
-#endif
+#include <fcntl.h>
+#include <limits.h>
+#include <stdlib.h>
+#include <unistd.h>
 
 #ifdef HAVE_ALLOCA_H
 #  include <alloca.h>
-#else
-#  include <stdlib.h>
 #endif
 
-#include <limits.h>
 
 #ifndef __WIN32__
 #  include <sys/uio.h> /* for `struct iovec' */
@@ -80,7 +72,6 @@
  * the WIM.  (This is not the on-disk format; the on-disk format just has an
  * array of offsets.) */
 struct chunk_table {
-       off_t file_offset;
        u64 original_resource_size;
        u64 num_chunks;
        u64 table_disk_size;
@@ -96,30 +87,32 @@ struct chunk_table {
        u8 offsets[] _aligned_attribute(8);
 };
 
-/*
- * Allocates and initializes a chunk table, and reserves space for it in the
- * output file.
- */
+/* 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,
-                            int out_fd,
-                            off_t file_offset,
-                            struct chunk_table **chunk_tab_ret)
+                            struct filedes *out_fd,
+                            struct chunk_table **chunk_tab_ret,
+                            int resource_flags)
 {
-       u64 size = wim_resource_size(lte);
-       u64 num_chunks = wim_resource_chunks(lte);
-       unsigned bytes_per_chunk_entry = (size > (1ULL << 32)) ? 8 : 4;
-       size_t alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
-       struct chunk_table *chunk_tab = CALLOC(1, alloc_size);
+       u64 size;
+       u64 num_chunks;
+       unsigned bytes_per_chunk_entry;
+       size_t alloc_size;
+       struct chunk_table *chunk_tab;
+       int ret;
 
-       DEBUG("Beginning chunk table for stream with size %"PRIu64, size);
+       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->file_offset = file_offset;
        chunk_tab->num_chunks = num_chunks;
        chunk_tab->original_resource_size = size;
        chunk_tab->bytes_per_chunk_entry = bytes_per_chunk_entry;
@@ -127,16 +120,23 @@ begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
                                     (num_chunks - 1);
        chunk_tab->cur_offset_p = chunk_tab->offsets;
 
-       /* We don't know the correct offsets yet; this just writes zeroes to
+       /* 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. */
-       if (full_write(out_fd, chunk_tab->offsets,
-                      chunk_tab->table_disk_size) != chunk_tab->table_disk_size)
-       {
-               ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
-                                "file resource");
-               FREE(chunk_tab);
-               return WIMLIB_ERR_WRITE;
+        * 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;
@@ -193,100 +193,81 @@ get_compress_func(int out_ctype)
                return wimlib_xpress_compress;
 }
 
-/*
- * Writes a chunk of a WIM resource to an output file.
- *
- * @chunk:       Uncompressed data of the chunk.
- * @chunk_size:          Size of the chunk (<= WIM_CHUNK_SIZE)
- * @out_fd:      File descriptor to write the chunk to.
- * @compress:     Compression function to use (NULL if writing uncompressed
- *                     data).
- * @chunk_tab:   Pointer to chunk table being created.  It is updated with the
- *                     offset of the chunk we write.
- *
- * Returns 0 on success; nonzero on failure.
- */
+/* Finishes a WIM chunk table and writes it to the output file at the correct
+ * offset.  */
 static int
-write_wim_resource_chunk(const void * restrict chunk,
-                        unsigned chunk_size,
-                        int out_fd,
-                        compress_func_t compress,
-                        struct chunk_table * restrict chunk_tab)
+finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
+                             struct filedes *out_fd,
+                             off_t res_start_offset,
+                             int write_resource_flags)
 {
-       const void *out_chunk;
-       unsigned out_chunk_size;
-       if (compress) {
-               void *compressed_chunk = alloca(chunk_size);
+       int ret;
 
-               out_chunk_size = (*compress)(chunk, chunk_size, compressed_chunk);
-               if (out_chunk_size) {
-                       /* Write compressed */
-                       out_chunk = compressed_chunk;
-               } else {
-                       /* Write uncompressed */
-                       out_chunk = chunk;
-                       out_chunk_size = chunk_size;
-               }
-               chunk_tab_record_chunk(chunk_tab, out_chunk_size);
+       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 {
-               /* Write uncompressed */
-               out_chunk = chunk;
-               out_chunk_size = chunk_size;
+               ret  = full_pwrite(out_fd,
+                                  chunk_tab->offsets +
+                                          chunk_tab->bytes_per_chunk_entry,
+                                  chunk_tab->table_disk_size,
+                                  res_start_offset);
        }
-       if (full_write(out_fd, out_chunk, out_chunk_size) != out_chunk_size) {
-               ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
-               return WIMLIB_ERR_WRITE;
+       if (ret) {
+               ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
+                                "file resource");
        }
-       return 0;
+       return ret;
 }
 
-/*
- * Finishes a WIM chunk table and writes it to the output file at the correct
- * offset.
- *
- * The final size of the full compressed resource is returned in the
- * @compressed_size_p.
+/* Write the header for a stream in a pipable WIM.
  */
 static int
-finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
-                             int out_fd, u64 *compressed_size_p)
+write_pwm_stream_header(const struct wim_lookup_table_entry *lte,
+                       struct filedes *out_fd,
+                       int additional_reshdr_flags)
 {
-       size_t bytes_written;
+       struct pwm_stream_hdr stream_hdr;
+       u32 reshdr_flags;
+       int ret;
 
-       bytes_written = full_pwrite(out_fd,
-                                   chunk_tab->offsets + chunk_tab->bytes_per_chunk_entry,
-                                   chunk_tab->table_disk_size,
-                                   chunk_tab->file_offset);
-       if (bytes_written != chunk_tab->table_disk_size) {
-               ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
-                                "file resource");
-               return WIMLIB_ERR_WRITE;
+       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);
        }
-       if (chunk_tab->bytes_per_chunk_entry == 4)
-               *compressed_size_p = chunk_tab->cur_offset_u32 + chunk_tab->table_disk_size;
-       else
-               *compressed_size_p = chunk_tab->cur_offset_u64 + chunk_tab->table_disk_size;
-       return 0;
+
+       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(int out_fd, off_t offset)
+seek_and_truncate(struct filedes *out_fd, off_t offset)
 {
-       if (lseek(out_fd, offset, SEEK_SET) == -1 ||
-           ftruncate(out_fd, 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;
-       } else {
-               return 0;
        }
+       return 0;
 }
 
 static int
-finalize_and_check_sha1(SHA_CTX * restrict sha_ctx,
-                       struct wim_lookup_table_entry * restrict lte)
+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);
@@ -302,45 +283,99 @@ finalize_and_check_sha1(SHA_CTX * restrict sha_ctx,
        return 0;
 }
 
-
 struct write_resource_ctx {
        compress_func_t compress;
        struct chunk_table *chunk_tab;
-       int out_fd;
+       struct filedes *out_fd;
        SHA_CTX sha_ctx;
        bool doing_sha;
+       int resource_flags;
 };
 
 static int
-write_resource_cb(const void *restrict chunk, size_t chunk_size,
-                 void *restrict _ctx)
+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);
-       return write_wim_resource_chunk(chunk, chunk_size,
-                                       ctx->out_fd, ctx->compress,
-                                       ctx->chunk_tab);
+
+       out_chunk = chunk;
+       out_chunk_size = chunk_size;
+       if (ctx->compress) {
+               void *compressed_chunk;
+               unsigned compressed_size;
+
+               /* Compress the chunk.  */
+               compressed_chunk = alloca(chunk_size);
+               compressed_size = (*ctx->compress)(chunk, chunk_size,
+                                                  compressed_chunk);
+
+               /* 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.
+ * @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_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_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.
+ * @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.
  *
- * @flags:  WIMLIB_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed
- *          even if it could otherwise be copied directly from the input.
+ * @write_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_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
@@ -350,36 +385,43 @@ write_resource_cb(const void *restrict chunk, size_t chunk_size,
  */
 int
 write_wim_resource(struct wim_lookup_table_entry *lte,
-                  int out_fd, int out_ctype,
+                  struct filedes *out_fd, int out_ctype,
                   struct resource_entry *out_res_entry,
-                  int flags)
+                  int resource_flags)
 {
        struct write_resource_ctx write_ctx;
+       off_t res_start_offset;
        u64 read_size;
-       u64 new_size;
-       off_t offset;
        int ret;
 
-       flags &= ~WIMLIB_RESOURCE_FLAG_RECOMPRESS;
+       /* 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 */
-       offset = filedes_offset(out_fd);
-       if (offset == -1) {
-               ERROR_WITH_ERRNO("Can't get position in output WIM");
-               return WIMLIB_ERR_WRITE;
-       }
+       /* 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 (!(flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
-           lte->resource_location == RESOURCE_IN_WIM &&
+        * 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 &&
-           wimlib_get_compression_type(lte->wim) == out_ctype)
+           !(resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS))
        {
-               flags |= WIMLIB_RESOURCE_FLAG_RAW;
+               /* 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 {
@@ -388,27 +430,43 @@ write_wim_resource(struct wim_lookup_table_entry *lte,
                read_size = lte->resource_entry.original_size;
        }
 
-       /* Initialize the chunk table and set the compression function if
-        * compressing the resource. */
-       if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
-           (flags & WIMLIB_RESOURCE_FLAG_RAW)) {
-               write_ctx.compress = NULL;
-               write_ctx.chunk_tab = NULL;
-       } else {
-               write_ctx.compress = get_compress_func(out_ctype);
-               ret = begin_wim_resource_chunk_tab(lte, out_fd,
-                                                  offset,
-                                                  &write_ctx.chunk_tab);
+       /* 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.compress = NULL;
+       write_ctx.chunk_tab = NULL;
+       if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
+               if (!(resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW))
+                       write_ctx.compress = get_compress_func(out_ctype);
+               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)
-                       return 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, flags);
+                                  write_resource_cb, &write_ctx, resource_flags);
        if (ret)
                goto out_free_chunk_tab;
 
@@ -420,49 +478,107 @@ try_write_again:
                        goto out_free_chunk_tab;
        }
 
-       out_res_entry->flags = lte->resource_entry.flags;
-       out_res_entry->original_size = wim_resource_size(lte);
-       out_res_entry->offset = offset;
-       if (flags & WIMLIB_RESOURCE_FLAG_RAW) {
-               /* Doing a raw write:  The new compressed size is the same as
-                * the compressed size in the other WIM. */
-               new_size = lte->resource_entry.size;
-       } else if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE) {
-               /* Using WIMLIB_COMPRESSION_TYPE_NONE:  The new compressed size
-                * is the original size. */
-               new_size = lte->resource_entry.original_size;
-               out_res_entry->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
-       } else {
-               /* Using a different compression type:  Call
-                * finish_wim_resource_chunk_tab() and it will provide the new
-                * compressed size. */
-               ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab, out_fd,
-                                                   &new_size);
+       /* 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;
-               if (new_size >= wim_resource_size(lte)) {
-                       /* Oops!  We compressed the resource to larger than the original
-                        * size.  Write the resource uncompressed instead. */
-                       DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
-                             "writing uncompressed instead",
-                             wim_resource_size(lte), new_size);
-                       ret = seek_and_truncate(out_fd, offset);
-                       if (ret)
-                               goto out_free_chunk_tab;
-                       write_ctx.compress = NULL;
-                       write_ctx.doing_sha = false;
-                       out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
-                       goto try_write_again;
-               }
+       }
+
+       /* 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.compress = NULL;
+               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);
        }
-       out_res_entry->size = new_size;
        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)
+{
+       /* 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);
+       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) */
@@ -566,7 +682,6 @@ struct message {
        u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
        unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
        struct iovec out_chunks[MAX_CHUNKS_PER_MSG];
-       size_t total_out_bytes;
        unsigned num_chunks;
        struct list_head list;
        bool complete;
@@ -576,7 +691,6 @@ struct message {
 static void
 compress_chunks(struct message *msg, compress_func_t compress)
 {
-       msg->total_out_bytes = 0;
        for (unsigned i = 0; i < msg->num_chunks; i++) {
                unsigned len = compress(msg->uncompressed_chunks[i],
                                        msg->uncompressed_chunk_sizes[i],
@@ -594,7 +708,6 @@ compress_chunks(struct message *msg, compress_func_t compress)
                }
                msg->out_chunks[i].iov_base = out_chunk;
                msg->out_chunks[i].iov_len = out_len;
-               msg->total_out_bytes += out_len;
        }
 }
 
@@ -622,32 +735,52 @@ compressor_thread_proc(void *arg)
 }
 #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(union wimlib_progress_info *progress,
-                         wimlib_progress_func_t progress_func,
-                         uint64_t size_added,
+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 -= size_added;
-               if (progress->write_streams._private != ~(uint64_t)0 &&
-                   progress->write_streams._private > progress->write_streams.total_bytes)
+               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->write_streams._private = progress->write_streams.total_bytes;
+                       progress_data->next_progress = progress->write_streams.total_bytes;
                }
        } else {
-               progress->write_streams.completed_bytes += size_added;
+               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_func &&
-           progress->write_streams.completed_bytes >= progress->write_streams._private)
+       if (progress_data->progress_func
+           && (progress->write_streams.completed_bytes >= progress_data->next_progress
+               || new_wim_part))
        {
-               progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
-                             progress);
-               if (progress->write_streams._private == progress->write_streams.total_bytes) {
-                       progress->write_streams._private = ~(uint64_t)0;
+               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->write_streams._private =
+                       progress_data->next_progress =
                                min(progress->write_streams.total_bytes,
                                    progress->write_streams.completed_bytes +
                                        progress->write_streams.total_bytes / 100);
@@ -656,7 +789,7 @@ do_write_streams_progress(union wimlib_progress_info *progress,
 }
 
 struct serial_write_stream_ctx {
-       int out_fd;
+       struct filedes *out_fd;
        int out_ctype;
        int write_resource_flags;
 };
@@ -670,6 +803,7 @@ serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
                                  ctx->write_resource_flags);
 }
 
+
 /* 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
@@ -679,8 +813,7 @@ 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,
-                    wimlib_progress_func_t progress_func,
-                    union wimlib_progress_info *progress)
+                    struct write_streams_progress_data *progress_data)
 {
        int ret = 0;
        struct wim_lookup_table_entry *lte;
@@ -698,19 +831,29 @@ do_write_stream_list(struct list_head *stream_list,
                         * 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_refcnt = lte->out_refcnt;
+                       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;
-                               /* We found a duplicate stream. */
-                               if (orig_refcnt != tmp->out_refcnt) {
-                                       /* We have already written, or are going
-                                        * to write, the duplicate stream.  So
-                                        * just skip to the next stream. */
-                                       DEBUG("Discarding duplicate stream of length %"PRIu64,
+
+                               /* '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;
@@ -743,10 +886,8 @@ do_write_stream_list(struct list_head *stream_list,
                }
        skip_to_progress:
                if (!lte->no_progress) {
-                       do_write_streams_progress(progress,
-                                                 progress_func,
-                                                 wim_resource_size(lte),
-                                                 stream_discarded);
+                       do_write_streams_progress(progress_data,
+                                                 lte, stream_discarded);
                }
        }
        return ret;
@@ -755,11 +896,10 @@ do_write_stream_list(struct list_head *stream_list,
 static int
 do_write_stream_list_serial(struct list_head *stream_list,
                            struct wim_lookup_table *lookup_table,
-                           int out_fd,
+                           struct filedes *out_fd,
                            int out_ctype,
                            int write_resource_flags,
-                           wimlib_progress_func_t progress_func,
-                           union wimlib_progress_info *progress)
+                           struct write_streams_progress_data *progress_data)
 {
        struct serial_write_stream_ctx ctx = {
                .out_fd = out_fd,
@@ -770,8 +910,7 @@ do_write_stream_list_serial(struct list_head *stream_list,
                                    lookup_table,
                                    serial_write_stream,
                                    &ctx,
-                                   progress_func,
-                                   progress);
+                                   progress_data);
 }
 
 static inline int
@@ -780,59 +919,86 @@ write_flags_to_resource_flags(int write_flags)
        int resource_flags = 0;
 
        if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
-               resource_flags |= WIMLIB_RESOURCE_FLAG_RECOMPRESS;
+               resource_flags |= WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS;
+       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,
-                        int out_fd,
+                        struct filedes *out_fd,
                         int out_ctype,
                         int write_resource_flags,
-                        wimlib_progress_func_t progress_func,
-                        union wimlib_progress_info *progress)
+                        struct write_streams_progress_data *progress_data)
 {
-       DEBUG("Writing stream list (serial version)");
+       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_func)
-               progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
+       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,
                                           write_resource_flags,
-                                          progress_func,
-                                          progress);
+                                          progress_data);
 }
 
 #ifdef ENABLE_MULTITHREADED_COMPRESSION
 static int
-write_wim_chunks(struct message *msg, int out_fd,
-                struct chunk_table *chunk_tab)
+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 (full_writev(out_fd, msg->out_chunks,
-                       msg->num_chunks) != msg->total_out_bytes)
-       {
-               ERROR_WITH_ERRNO("Failed to write WIM chunks");
-               return WIMLIB_ERR_WRITE;
+
+       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;
+               }
        }
-       return 0;
+       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;
-       int out_fd;
+       struct filedes *out_fd;
+       off_t res_start_offset;
        int out_ctype;
        int write_resource_flags;
        struct shared_queue *res_to_compress_queue;
        struct shared_queue *compressed_res_queue;
        size_t num_messages;
-       wimlib_progress_func_t progress_func;
-       union wimlib_progress_info *progress;
+       struct write_streams_progress_data *progress_data;
 
        struct list_head available_msgs;
        struct list_head outstanding_streams;
@@ -979,21 +1145,35 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
        {
                list_move(&msg->list, &ctx->available_msgs);
                if (msg->begin_chunk == 0) {
-                       /* This is the first set of chunks.  Leave space
-                        * for the chunk table in the output file. */
-                       off_t cur_offset = filedes_offset(ctx->out_fd);
-                       if (cur_offset == -1)
-                               return WIMLIB_ERR_WRITE;
+                       /* 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,
-                                                          cur_offset,
-                                                          &ctx->cur_chunk_tab);
+                                                          &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);
+               ret = write_wim_chunks(msg, ctx->out_fd, ctx->cur_chunk_tab,
+                                      ctx->write_resource_flags);
                if (ret)
                        return ret;
 
@@ -1003,31 +1183,32 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                    msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
                {
                        u64 res_csize;
-                       off_t offset;
 
                        ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
                                                            ctx->out_fd,
-                                                           &res_csize);
+                                                           ctx->res_start_offset,
+                                                           ctx->write_resource_flags);
                        if (ret)
                                return ret;
 
                        list_del(&cur_lte->being_compressed_list);
 
-                       /* Grab the offset of this stream in the output file
-                        * from the chunk table before we free it. */
-                       offset = ctx->cur_chunk_tab->file_offset;
+                       res_csize = ctx->out_fd->offset - ctx->res_start_offset;
 
                        FREE(ctx->cur_chunk_tab);
                        ctx->cur_chunk_tab = NULL;
 
-                       if (res_csize >= wim_resource_size(cur_lte)) {
-                               /* Oops!  We compressed the resource to
-                                * larger than the original size.  Write
-                                * the resource uncompressed instead. */
+                       /* 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, offset);
+                               ret = seek_and_truncate(ctx->out_fd, ctx->res_start_offset);
                                if (ret)
                                        return ret;
                                ret = write_wim_resource(cur_lte,
@@ -1045,17 +1226,22 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                                        cur_lte->resource_entry.original_size;
 
                                cur_lte->output_resource_entry.offset =
-                                       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,
-                                                 ctx->progress_func,
-                                                 wim_resource_size(cur_lte),
-                                                 false);
+                       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
@@ -1069,8 +1255,7 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                                                                  ctx->out_fd,
                                                                  ctx->out_ctype,
                                                                  ctx->write_resource_flags,
-                                                                 ctx->progress_func,
-                                                                 ctx->progress);
+                                                                 ctx->progress_data);
                                if (ret)
                                        return ret;
                        }
@@ -1168,8 +1353,7 @@ main_writer_thread_finish(void *_ctx)
                                           ctx->out_fd,
                                           ctx->out_ctype,
                                           ctx->write_resource_flags,
-                                          ctx->progress_func,
-                                          ctx->progress);
+                                          ctx->progress_data);
 }
 
 static int
@@ -1190,11 +1374,10 @@ submit_stream_for_compression(struct wim_lookup_table_entry *lte,
        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 == 0) {
-               wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
-               ret = finalize_and_check_sha1(&ctx->next_sha_ctx, lte);
-       }
-       return ret;
+       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
@@ -1206,8 +1389,8 @@ main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
        if (wim_resource_size(lte) < 1000 ||
            ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
            (lte->resource_location == RESOURCE_IN_WIM &&
-            !(ctx->write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
-            wimlib_get_compression_type(lte->wim) == ctx->out_ctype))
+            !(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
@@ -1262,17 +1445,17 @@ get_default_num_threads(void)
 static int
 write_stream_list_parallel(struct list_head *stream_list,
                           struct wim_lookup_table *lookup_table,
-                          int out_fd,
+                          struct filedes *out_fd,
                           int out_ctype,
                           int write_resource_flags,
-                          wimlib_progress_func_t progress_func,
-                          union wimlib_progress_info *progress,
+                          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();
@@ -1286,8 +1469,9 @@ write_stream_list_parallel(struct list_head *stream_list,
                }
        }
 
-       DEBUG("Writing stream list (parallel version, num_threads=%u)",
-             num_threads);
+       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;
 
@@ -1329,8 +1513,10 @@ write_stream_list_parallel(struct list_head *stream_list,
                }
        }
 
-       if (progress_func)
-               progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
+       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;
@@ -1341,14 +1527,13 @@ write_stream_list_parallel(struct list_head *stream_list,
        ctx.compressed_res_queue  = &compressed_res_queue;
        ctx.num_messages          = queue_size;
        ctx.write_resource_flags  = write_resource_flags;
-       ctx.progress_func         = progress_func;
-       ctx.progress              = progress;
+       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_func, progress);
+                                  &ctx, progress_data);
        if (ret)
                goto out_destroy_ctx;
 
@@ -1387,8 +1572,7 @@ out_serial_quiet:
                                        out_fd,
                                        out_ctype,
                                        write_resource_flags,
-                                       progress_func,
-                                       progress);
+                                       progress_data);
 
 }
 #endif
@@ -1400,22 +1584,33 @@ out_serial_quiet:
 static int
 write_stream_list(struct list_head *stream_list,
                  struct wim_lookup_table *lookup_table,
-                 int out_fd, int out_ctype, int write_flags,
+                 struct filedes *out_fd, int out_ctype, 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;
-       union wimlib_progress_info progress;
+       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))
+       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
@@ -1425,18 +1620,32 @@ write_stream_list(struct list_head *stream_list,
                total_bytes += wim_resource_size(lte);
                if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
                       && (wim_resource_compression_type(lte) != out_ctype ||
-                          (write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS)))
+                          (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++;
+                       }
+               }
        }
-       progress.write_streams.total_bytes       = total_bytes;
-       progress.write_streams.total_streams     = num_streams;
-       progress.write_streams.completed_bytes   = 0;
-       progress.write_streams.completed_streams = 0;
-       progress.write_streams.num_threads       = num_threads;
-       progress.write_streams.compression_type  = out_ctype;
-       progress.write_streams._private          = 0;
+
+       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)
@@ -1445,8 +1654,7 @@ write_stream_list(struct list_head *stream_list,
                                                 out_fd,
                                                 out_ctype,
                                                 write_resource_flags,
-                                                progress_func,
-                                                &progress,
+                                                &progress_data,
                                                 num_threads);
        else
 #endif
@@ -1455,8 +1663,11 @@ write_stream_list(struct list_head *stream_list,
                                               out_fd,
                                               out_ctype,
                                               write_resource_flags,
-                                              progress_func,
-                                              &progress);
+                                              &progress_data);
+       if (ret == 0)
+               DEBUG("Successfully wrote stream list.");
+       else
+               DEBUG("Failed to write stream list.");
        return ret;
 }
 
@@ -1491,7 +1702,7 @@ stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
        struct wim_lookup_table_entry *same_size_lte;
        struct hlist_node *tmp;
 
-       pos = hash_u64(wim_resource_size(lte)) % tab->capacity;
+       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)) {
@@ -1506,480 +1717,984 @@ stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
        return 0;
 }
 
-
-struct lte_overwrite_prepare_args {
+struct find_streams_ctx {
        WIMStruct *wim;
-       off_t end_offset;
+       int write_flags;
        struct list_head stream_list;
        struct stream_size_table stream_size_tab;
 };
 
-/* First phase of preparing streams for an in-place overwrite.  This is called
- * on all streams, both hashed and unhashed, except the metadata resources. */
-static int
-lte_overwrite_prepare(struct wim_lookup_table_entry *lte, void *_args)
+static void
+lte_reference_for_logical_write(struct wim_lookup_table_entry *lte,
+                               struct find_streams_ctx *ctx,
+                               unsigned nref)
 {
-       struct lte_overwrite_prepare_args *args = _args;
+       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;
+}
 
-       wimlib_assert(!(lte->resource_entry.flags & WIM_RESHDR_FLAG_METADATA));
-       if (lte->resource_location != RESOURCE_IN_WIM || lte->wim != args->wim)
-               list_add_tail(&lte->write_streams_list, &args->stream_list);
-       lte->out_refcnt = lte->refcnt;
-       stream_size_table_insert(lte, &args->stream_size_tab);
+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;
 }
 
-/* Second phase of preparing streams for an in-place overwrite.  This is called
- * on existing metadata resources and hashed streams, but not unhashed streams.
- *
- * NOTE: lte->output_resource_entry is in union with lte->hash_list_2, so
- * lte_overwrite_prepare_2() must be called after lte_overwrite_prepare(), as
- * the latter uses lte->hash_list_2, while the former expects to set
- * lte->output_resource_entry. */
 static int
-lte_overwrite_prepare_2(struct wim_lookup_table_entry *lte, void *_args)
+inode_find_streams_to_write(struct wim_inode *inode,
+                           struct wim_lookup_table *table,
+                           struct find_streams_ctx *ctx)
 {
-       struct lte_overwrite_prepare_args *args = _args;
+       struct wim_lookup_table_entry *lte;
+       unsigned i;
 
-       if (lte->resource_location == RESOURCE_IN_WIM && lte->wim == args->wim) {
-               /* We can't do an in place overwrite on the WIM if there are
-                * streams after the XML data. */
-               if (lte->resource_entry.offset +
-                   lte->resource_entry.size > args->end_offset)
-               {
-                       if (wimlib_print_errors) {
-                               ERROR("The following resource is after the XML data:");
-                               print_lookup_table_entry(lte, stderr);
-                       }
-                       return WIMLIB_ERR_RESOURCE_ORDER;
-               }
-               copy_resource_entry(&lte->output_resource_entry,
-                                   &lte->resource_entry);
+       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;
 }
 
-/* Given a WIM that we are going to overwrite in place with zero or more
- * additional streams added, construct a list the list of new unique streams
- * ('struct wim_lookup_table_entry's) that must be written, plus any unhashed
- * streams that need to be added but may be identical to other hashed or
- * unhashed streams.  These unhashed streams are checksummed while the streams
- * are being written.  To aid this process, the member @unique_size is set to 1
- * on streams that have a unique size and therefore must be written.
- *
- * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
- * indicate the number of times the stream is referenced in only the streams
- * that are being written; this may still be adjusted later when unhashed
- * streams are being resolved.
- */
 static int
-prepare_streams_for_overwrite(WIMStruct *wim, off_t end_offset,
-                             struct list_head *stream_list)
+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;
-       struct lte_overwrite_prepare_args args;
-       unsigned i;
 
-       args.wim = wim;
-       args.end_offset = end_offset;
-       ret = init_stream_size_table(&args.stream_size_tab,
-                                    wim->lookup_table->capacity);
-       if (ret)
-               return ret;
-
-       INIT_LIST_HEAD(&args.stream_list);
-       for (i = 0; i < wim->hdr.image_count; i++) {
-               struct wim_image_metadata *imd;
-               struct wim_lookup_table_entry *lte;
+       ctx = wim->private;
+       imd = wim_get_current_image_metadata(wim);
 
-               imd = wim->image_metadata[i];
-               image_for_each_unhashed_stream(lte, imd)
-                       lte_overwrite_prepare(lte, &args);
-       }
-       for_lookup_table_entry(wim->lookup_table, lte_overwrite_prepare, &args);
-       list_transfer(&args.stream_list, stream_list);
+       image_for_each_unhashed_stream(lte, imd)
+               lte->out_refcnt = 0;
 
-       for (i = 0; i < wim->hdr.image_count; i++) {
-               ret = lte_overwrite_prepare_2(wim->image_metadata[i]->metadata_lte,
-                                             &args);
+       /* 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)
-                       goto out_destroy_stream_size_table;
+                       return ret;
        }
-       ret = for_lookup_table_entry(wim->lookup_table,
-                                    lte_overwrite_prepare_2, &args);
-out_destroy_stream_size_table:
-       destroy_stream_size_table(&args.stream_size_tab);
-       return ret;
+       return 0;
 }
 
-
-struct find_streams_ctx {
-       struct list_head stream_list;
-       struct stream_size_table stream_size_tab;
-};
-
-static void
-inode_find_streams_to_write(struct wim_inode *inode,
-                           struct wim_lookup_table *table,
-                           struct list_head *stream_list,
-                           struct stream_size_table *tab)
+/*
+ * 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;
-       for (unsigned i = 0; i <= inode->i_num_ads; i++) {
-               lte = inode_stream_lte(inode, i, table);
-               if (lte) {
-                       if (lte->out_refcnt == 0) {
-                               if (lte->unhashed)
-                                       stream_size_table_insert(lte, tab);
-                               list_add_tail(&lte->write_streams_list, stream_list);
-                       }
-                       lte->out_refcnt += inode->i_nlink;
+
+       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
-image_find_streams_to_write(WIMStruct *w)
+process_filtered_stream(struct wim_lookup_table_entry *lte, void *_ctx)
 {
-       struct find_streams_ctx *ctx;
-       struct wim_image_metadata *imd;
-       struct wim_inode *inode;
-       struct wim_lookup_table_entry *lte;
+       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;
+}
 
-       ctx = w->private;
-       imd = wim_get_current_image_metadata(w);
+static int
+mark_stream_not_filtered(struct wim_lookup_table_entry *lte, void *_ignore)
+{
+       lte->filtered = 0;
+       return 0;
+}
 
-       image_for_each_unhashed_stream(lte, imd)
-               lte->out_refcnt = 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;
 
-       /* Go through this image's inodes to find any streams that have not been
-        * found yet. */
-       image_for_each_inode(inode, imd) {
-               inode_find_streams_to_write(inode, w->lookup_table,
-                                           &ctx->stream_list,
-                                           &ctx->stream_size_tab);
+       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;
        }
-       return 0;
+
+       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);
 }
 
-/* Given a WIM that from which one or all of the images is being written, build
- * the list of unique streams ('struct wim_lookup_table_entry's) that must be
- * written, plus any unhashed streams that need to be written but may be
- * identical to other hashed or unhashed streams being written.  These unhashed
- * streams are checksummed while the streams are being written.  To aid this
- * process, the member @unique_size is set to 1 on streams that have a unique
- * size and therefore must be written.
+/* 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':
  *
- * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
- * indicate the number of times the stream is referenced in only the streams
- * that are being written; this may still be adjusted later when unhashed
- * streams are being resolved.
+ * - '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, struct list_head *stream_list)
+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;
 
-       for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
+       INIT_LIST_HEAD(&ctx.stream_list);
        ret = init_stream_size_table(&ctx.stream_size_tab,
                                     wim->lookup_table->capacity);
        if (ret)
                return ret;
-       for_lookup_table_entry(wim->lookup_table, stream_size_table_insert,
-                              &ctx.stream_size_tab);
-       INIT_LIST_HEAD(&ctx.stream_list);
-       wim->private = &ctx;
-       ret = for_image(wim, image, image_find_streams_to_write);
+       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);
-       if (ret == 0)
-               list_transfer(&ctx.stream_list, stream_list);
        return ret;
 }
 
-/* Writes the streams for the specified @image in @wim to @wim->out_fd.
- */
 static int
 write_wim_streams(WIMStruct *wim, int image, int write_flags,
                  unsigned num_threads,
-                 wimlib_progress_func_t progress_func)
+                 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 list_head *stream_list;
+       struct wim_lookup_table_entry *lte;
 
-       ret = prepare_stream_list(wim, image, &stream_list);
-       if (ret)
-               return ret;
-       return write_stream_list(&stream_list,
+       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,
-                                wimlib_get_compression_type(wim),
+                                &wim->out_fd,
+                                wim->compression_type,
                                 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);
+               }
+               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 writing a WIM file: write the lookup table, xml data, and integrity
- * table (optional), then overwrite the WIM header.
+ * table, then overwrite the WIM header.  Always closes the WIM file descriptor
+ * (wim->out_fd).
  *
  * write_flags is a bitwise OR of the following:
  *
- *     (public)  WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
- *             Include an integrity table.
+ *     (public) WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
+ *             Include an integrity table.
  *
- *     (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
- *             Don't write the lookup table.
+ *     (public) WIMLIB_WRITE_FLAG_FSYNC:
+ *             fsync() the output file before closing it.
  *
- *     (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
- *             When (if) writing the integrity table, re-use entries from the
- *             existing integrity table, if possible.
+ *     (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_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_NO_LOOKUP_TABLE:
+ *             Don't write the lookup table.
  *
- *     (private) WIMLIB_WRITE_FLAG_FSYNC:
- *             fsync() the output file before closing it.
+ *     (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_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.
  */
-int
-finish_write(WIMStruct *w, int image, int write_flags,
-            wimlib_progress_func_t progress_func)
+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;
-       struct wim_header hdr;
-
-       /* @hdr will be the header for the new WIM.  First copy all the data
-        * from the header in the WIMStruct; then set all the fields that may
-        * have changed, including the resource entries, boot index, and image
-        * count.  */
-       memcpy(&hdr, &w->hdr, sizeof(struct wim_header));
+       off_t hdr_offset;
+       int write_resource_flags;
+       off_t old_lookup_table_end;
+       off_t new_lookup_table_end;
+       u64 xml_totalbytes;
 
-       /* Set image count and boot index correctly for single image writes */
-       if (image != WIMLIB_ALL_IMAGES) {
-               hdr.image_count = 1;
-               if (hdr.boot_idx == image)
-                       hdr.boot_idx = 1;
-               else
-                       hdr.boot_idx = 0;
-       }
+       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 (hdr.boot_idx == 0) {
-               zero_resource_entry(&hdr.boot_metadata_res_entry);
+       if (wim->hdr.boot_idx == 0) {
+               zero_resource_entry(&wim->hdr.boot_metadata_res_entry);
        } else {
-               copy_resource_entry(&hdr.boot_metadata_res_entry,
-                           &w->image_metadata[ hdr.boot_idx- 1
+               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_lookup_table(w, image, &hdr.lookup_table_res_entry);
+               ret = write_wim_lookup_table(wim, image, write_flags,
+                                            &wim->hdr.lookup_table_res_entry,
+                                            stream_list_override);
                if (ret)
                        goto out_close_wim;
        }
 
-       ret = write_xml_data(w->wim_info, image, w->out_fd,
-                            (write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE) ?
-                             wim_info_get_total_bytes(w->wim_info) : 0,
-                            &hdr.xml_res_entry);
+       /* 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)
                goto out_close_wim;
 
        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, &hdr, sizeof(struct wim_header));
+                       memcpy(&checkpoint_hdr, &wim->hdr, sizeof(struct wim_header));
                        zero_resource_entry(&checkpoint_hdr.integrity);
-                       ret = write_header(&checkpoint_hdr, w->out_fd);
+                       checkpoint_hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
+                       ret = write_wim_header_at_offset(&checkpoint_hdr,
+                                                        &wim->out_fd, 0);
                        if (ret)
                                goto out_close_wim;
                }
 
-               off_t old_lookup_table_end;
-               off_t new_lookup_table_end;
-               if (write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE) {
-                       old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
-                                              w->hdr.lookup_table_res_entry.size;
-               } else {
+               if (!(write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE))
                        old_lookup_table_end = 0;
-               }
-               new_lookup_table_end = hdr.lookup_table_res_entry.offset +
-                                      hdr.lookup_table_res_entry.size;
 
-               ret = write_integrity_table(w->out_fd,
-                                           &hdr.integrity,
+               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)
                        goto out_close_wim;
        } else {
-               zero_resource_entry(&hdr.integrity);
+               zero_resource_entry(&wim->hdr.integrity);
        }
 
-       hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
-       ret = write_header(&hdr, w->out_fd);
+       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)
                goto out_close_wim;
 
        if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
-               if (fsync(w->out_fd)) {
+               if (fsync(wim->out_fd.fd)) {
                        ERROR_WITH_ERRNO("Error syncing data to WIM file");
                        ret = WIMLIB_ERR_WRITE;
+                       goto out_close_wim;
                }
        }
+
+       ret = 0;
 out_close_wim:
-       if (close(w->out_fd)) {
-               ERROR_WITH_ERRNO("Failed to close the output WIM file");
-               if (ret == 0)
+       if (close_wim_writable(wim, write_flags)) {
+               if (ret == 0) {
+                       ERROR_WITH_ERRNO("Failed to close the output WIM file");
                        ret = WIMLIB_ERR_WRITE;
+               }
        }
-       w->out_fd = -1;
        return ret;
 }
 
 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
 int
-lock_wim(WIMStruct *w, int fd)
+lock_wim(WIMStruct *wim, int fd)
 {
        int ret = 0;
-       if (fd != -1 && !w->wim_locked) {
+       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!", w->filename);
+                                     "        by another process!", wim->filename);
                                ret = WIMLIB_ERR_ALREADY_LOCKED;
                        } else {
                                WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
-                                                  w->filename);
+                                                  wim->filename);
                                ret = 0;
                        }
                } else {
-                       w->wim_locked = 1;
+                       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
-open_wim_writable(WIMStruct *w, const tchar *path, int open_flags)
+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)
 {
-       w->out_fd = topen(path, open_flags | O_BINARY, 0644);
-       if (w->out_fd == -1) {
-               ERROR_WITH_ERRNO("Failed to open `%"TS"' for writing", path);
-               return WIMLIB_ERR_OPEN;
-       }
-       return 0;
-}
+       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;
 
-void
-close_wim_writable(WIMStruct *w)
-{
-       if (w->out_fd != -1) {
-               if (close(w->out_fd))
-                       WARNING_WITH_ERRNO("Failed to close output WIM");
-               w->out_fd = -1;
-       }
+       /* 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().  */
 }
 
-/* Open file stream and write dummy header for WIM. */
+/* Write a standalone WIM or split WIM (SWM) part to a new file or to a file
+ * descriptor.  */
 int
-begin_write(WIMStruct *w, const tchar *path, int write_flags)
+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;
-       int open_flags = O_TRUNC | O_CREAT;
+       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)
-               open_flags |= O_RDWR;
+               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
-               open_flags |= O_WRONLY;
-       ret = open_wim_writable(w, path, open_flags);
-       if (ret)
-               return ret;
-       /* Write dummy header. It will be overwritten later. */
-       w->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
-       ret = write_header(&w->hdr, w->out_fd);
-       w->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
+               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)
-               return ret;
-       if (lseek(w->out_fd, WIM_HEADER_DISK_SIZE, SEEK_SET) == -1) {
-               ERROR_WITH_ERRNO("Failed to seek to end of WIM header");
-               return WIMLIB_ERR_WRITE;
+               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);
+               }
        }
-       return 0;
+
+       /* 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));
+       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);
 }
 
-/* Writes a stand-alone WIM to a file.  */
+/* API function documented in wimlib.h  */
 WIMLIBAPI int
-wimlib_write(WIMStruct *w, const tchar *path,
+wimlib_write(WIMStruct *wim, const tchar *path,
             int image, int write_flags, unsigned num_threads,
             wimlib_progress_func_t progress_func)
 {
-       int ret;
-
        if (!path)
                return WIMLIB_ERR_INVALID_PARAM;
 
        write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
 
-       if (image != WIMLIB_ALL_IMAGES &&
-            (image < 1 || image > w->hdr.image_count))
-               return WIMLIB_ERR_INVALID_IMAGE;
-
-       if (w->hdr.total_parts != 1) {
-               ERROR("Cannot call wimlib_write() on part of a split WIM");
-               return WIMLIB_ERR_SPLIT_UNSUPPORTED;
-       }
-
-       ret = begin_write(w, path, write_flags);
-       if (ret)
-               goto out_close_wim;
-
-       ret = write_wim_streams(w, image, write_flags, num_threads,
-                               progress_func);
-       if (ret)
-               goto out_close_wim;
-
-       if (progress_func)
-               progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
+       return write_standalone_wim(wim, path, image, write_flags,
+                                   num_threads, progress_func);
+}
 
-       ret = for_image(w, image, write_metadata_resource);
-       if (ret)
-               goto out_close_wim;
+/* 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;
 
-       if (progress_func)
-               progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
+       write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
+       write_flags |= WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR;
 
-       ret = finish_write(w, image, write_flags, progress_func);
-       /* finish_write() closed the WIM for us */
-       goto out;
-out_close_wim:
-       close_wim_writable(w);
-out:
-       DEBUG("wimlib_write(path=%"TS") = %d", path, ret);
-       return ret;
+       return write_standalone_wim(wim, &fd, image, write_flags,
+                                   num_threads, progress_func);
 }
 
 static bool
-any_images_modified(WIMStruct *w)
+any_images_modified(WIMStruct *wim)
 {
-       for (int i = 0; i < w->hdr.image_count; i++)
-               if (w->image_metadata[i]->modified)
+       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.
  *
@@ -2038,7 +2753,7 @@ any_images_modified(WIMStruct *w)
  * small amount of space compared to the streams, however.)
  */
 static int
-overwrite_wim_inplace(WIMStruct *w, int write_flags,
+overwrite_wim_inplace(WIMStruct *wim, int write_flags,
                      unsigned num_threads,
                      wimlib_progress_func_t progress_func)
 {
@@ -2046,25 +2761,35 @@ overwrite_wim_inplace(WIMStruct *w, int write_flags,
        struct list_head stream_list;
        off_t old_wim_end;
        u64 old_lookup_table_end, old_xml_begin, old_xml_end;
-       int open_flags;
 
-       DEBUG("Overwriting `%"TS"' in-place", w->filename);
+
+       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 = w->hdr.xml_res_entry.offset;
-       old_xml_end = old_xml_begin + w->hdr.xml_res_entry.size;
-       old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
-                              w->hdr.lookup_table_res_entry.size;
-       if (w->hdr.integrity.offset != 0 && w->hdr.integrity.offset < old_xml_end) {
-               ERROR("Didn't expect the integrity table to be before the XML data");
+       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) {
-               ERROR("Didn't expect the lookup table to be after the XML data");
+               WARNING("Didn't expect the lookup table to be after the XML data");
                return WIMLIB_ERR_RESOURCE_ORDER;
        }
 
@@ -2072,7 +2797,7 @@ overwrite_wim_inplace(WIMStruct *w, int write_flags,
         * 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 (!w->deletion_occurred && !any_images_modified(w)) {
+       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
@@ -2083,118 +2808,114 @@ overwrite_wim_inplace(WIMStruct *w, int write_flags,
                old_wim_end = old_lookup_table_end;
                write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
                               WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
-       } else if (w->hdr.integrity.offset) {
+       } else if (wim->hdr.integrity.offset) {
                /* Old WIM has an integrity table; begin writing new streams
                 * after it. */
-               old_wim_end = w->hdr.integrity.offset + w->hdr.integrity.size;
+               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 = prepare_streams_for_overwrite(w, old_wim_end, &stream_list);
+       ret = check_resource_offsets(wim, old_wim_end);
        if (ret)
                return ret;
 
-       open_flags = 0;
-       if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
-               open_flags |= O_RDWR;
-       else
-               open_flags |= O_WRONLY;
-       ret = open_wim_writable(w, w->filename, open_flags);
+       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(w, w->out_fd);
+       ret = lock_wim(wim, wim->out_fd.fd);
        if (ret) {
-               close_wim_writable(w);
+               close_wim_writable(wim, write_flags);
                return ret;
        }
 
-       /* Write header with write in progress flag set. */
-       w->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
-       ret = write_header(&w->hdr, w->out_fd);
-       w->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
-       if (ret)
-               return ret;
+       /* Set WIM_HDR_FLAG_WRITE_IN_PROGRESS flag in header. */
+       ret = write_wim_header_flags(wim->hdr.flags | WIM_HDR_FLAG_WRITE_IN_PROGRESS,
+                                    &wim->out_fd);
+       if (ret) {
+               ERROR_WITH_ERRNO("Error updating WIM header flags");
+               close_wim_writable(wim, write_flags);
+               goto out_unlock_wim;
+       }
 
-       if (lseek(w->out_fd, old_wim_end, SEEK_SET) == -1) {
+       if (filedes_seek(&wim->out_fd, old_wim_end) == -1) {
                ERROR_WITH_ERRNO("Can't seek to end of WIM");
-               close_wim_writable(w);
-               w->wim_locked = 0;
-               return WIMLIB_ERR_WRITE;
+               close_wim_writable(wim, write_flags);
+               ret = WIMLIB_ERR_WRITE;
+               goto out_unlock_wim;
        }
 
-       DEBUG("Writing newly added streams (offset = %"PRIu64")",
-             old_wim_end);
        ret = write_stream_list(&stream_list,
-                               w->lookup_table,
-                               w->out_fd,
-                               wimlib_get_compression_type(w),
+                               wim->lookup_table,
+                               &wim->out_fd,
+                               wim->compression_type,
                                write_flags,
                                num_threads,
                                progress_func);
        if (ret)
                goto out_truncate;
 
-       for (int i = 0; i < w->hdr.image_count; i++) {
-               if (w->image_metadata[i]->modified) {
-                       select_wim_image(w, i + 1);
-                       ret = write_metadata_resource(w);
-                       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(w, WIMLIB_ALL_IMAGES, write_flags,
-                          progress_func);
+       ret = finish_write(wim, WIMLIB_ALL_IMAGES, write_flags,
+                          progress_func, NULL);
 out_truncate:
-       close_wim_writable(w);
-       if (ret != 0 && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
+       close_wim_writable(wim, write_flags);
+       if (ret && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
                WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
-                       w->filename, old_wim_end);
+                       wim->filename, old_wim_end);
                /* Return value of truncate() is ignored because this is already
                 * an error path. */
-               (void)ttruncate(w->filename, old_wim_end);
+               (void)ttruncate(wim->filename, old_wim_end);
        }
-       w->wim_locked = 0;
+out_unlock_wim:
+       wim->wim_locked = 0;
        return ret;
 }
 
 static int
-overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
+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", w->filename);
+       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(w->filename);
+       wim_name_len = tstrlen(wim->filename);
        tchar tmpfile[wim_name_len + 10];
-       tmemcpy(tmpfile, w->filename, wim_name_len);
+       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(w, tmpfile, WIMLIB_ALL_IMAGES,
+       ret = wimlib_write(wim, tmpfile, WIMLIB_ALL_IMAGES,
                           write_flags | WIMLIB_WRITE_FLAG_FSYNC,
                           num_threads, progress_func);
-       if (ret) {
-               ERROR("Failed to write the WIM file `%"TS"'", tmpfile);
+       if (ret)
                goto out_unlink;
-       }
 
-       close_wim(w);
+       close_wim(wim);
 
-       DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, w->filename);
+       DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, wim->filename);
        /* Rename the new file to the old file .*/
-       if (trename(tmpfile, w->filename) != 0) {
+       if (trename(tmpfile, wim->filename) != 0) {
                ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
-                                tmpfile, w->filename);
+                                tmpfile, wim->filename);
                ret = WIMLIB_ERR_RENAME;
                goto out_unlink;
        }
@@ -2202,48 +2923,53 @@ overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
        if (progress_func) {
                union wimlib_progress_info progress;
                progress.rename.from = tmpfile;
-               progress.rename.to = w->filename;
+               progress.rename.to = wim->filename;
                progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
        }
-       goto out;
+       return 0;
+
 out_unlink:
        /* Remove temporary file. */
-       if (tunlink(tmpfile) != 0)
-               WARNING_WITH_ERRNO("Failed to remove `%"TS"'", tmpfile);
-out:
+       tunlink(tmpfile);
        return ret;
 }
 
-/*
- * Writes a WIM file to the original file that it was read from, overwriting it.
- */
+/* API function documented in wimlib.h  */
 WIMLIBAPI int
-wimlib_overwrite(WIMStruct *w, int write_flags,
+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 (!w->filename)
+       if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
+               return WIMLIB_ERR_INVALID_PARAM;
+
+       if (!wim->filename)
                return WIMLIB_ERR_NO_FILENAME;
 
-       ret = can_modify_wim(w);
+       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 ((!w->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
-           && !(write_flags & WIMLIB_WRITE_FLAG_REBUILD))
+       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)))
        {
-               int ret;
-               ret = overwrite_wim_inplace(w, write_flags, num_threads,
+               ret = overwrite_wim_inplace(wim, write_flags, num_threads,
                                            progress_func);
-               if (ret == WIMLIB_ERR_RESOURCE_ORDER)
-                       WARNING("Falling back to re-building entire WIM");
-               else
+               if (ret != WIMLIB_ERR_RESOURCE_ORDER)
                        return ret;
+               WARNING("Falling back to re-building entire WIM");
        }
-       return overwrite_wim_via_tmpfile(w, write_flags, num_threads,
+       return overwrite_wim_via_tmpfile(wim, write_flags, num_threads,
                                         progress_func);
 }