]> wimlib.net Git - wimlib/blobdiff - src/write.c
wim.c: Cleanup
[wimlib] / src / write.c
index 42b9fcfff84e38f682d05b5406dcfa0b58d7118a..c63f0f706dae9f3ad63cc7a927f8721b77587449 100644 (file)
 #  include <sys/uio.h> /* for `struct iovec' */
 #endif
 
+/* Return true if the specified resource is compressed and the compressed data
+ * can be reused with the specified output parameters.  */
+static bool
+can_raw_copy(const struct wim_lookup_table_entry *lte,
+            int write_resource_flags, int out_ctype, u32 out_chunk_size)
+{
+       return (out_ctype == wim_resource_compression_type(lte)
+               && out_chunk_size == wim_resource_chunk_size(lte)
+               && out_ctype != WIMLIB_COMPRESSION_TYPE_NONE);
+}
+
+
+/* Return true if the specified resource must be recompressed when the specified
+ * output parameters are used.  */
+static bool
+must_compress_stream(const struct wim_lookup_table_entry *lte,
+                    int write_resource_flags, int out_ctype, u32 out_chunk_size)
+{
+       return (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
+               && ((write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS)
+                   || !can_raw_copy(lte, write_resource_flags,
+                                    out_ctype, out_chunk_size)));
+}
+
+static unsigned
+compress_chunk(const void * uncompressed_data,
+              unsigned uncompressed_len,
+              void *compressed_data,
+              int out_ctype,
+              struct wimlib_lzx_context *comp_ctx)
+{
+       switch (out_ctype) {
+       case WIMLIB_COMPRESSION_TYPE_XPRESS:
+               return wimlib_xpress_compress(uncompressed_data,
+                                             uncompressed_len,
+                                             compressed_data);
+       case WIMLIB_COMPRESSION_TYPE_LZX:
+               return wimlib_lzx_compress2(uncompressed_data,
+                                           uncompressed_len,
+                                           compressed_data,
+                                           comp_ctx);
+       default:
+               wimlib_assert(0);
+               return 0;
+       }
+}
+
 /* Chunk table that's located at the beginning of each compressed resource in
  * the WIM.  (This is not the on-disk format; the on-disk format just has an
  * array of offsets.) */
@@ -92,6 +139,7 @@ struct chunk_table {
 static int
 begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
                             struct filedes *out_fd,
+                            u32 out_chunk_size,
                             struct chunk_table **chunk_tab_ret,
                             int resource_flags)
 {
@@ -103,7 +151,7 @@ begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
        int ret;
 
        size = wim_resource_size(lte);
-       num_chunks = wim_resource_chunks(lte);
+       num_chunks = DIV_ROUND_UP(size, out_chunk_size);
        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);
@@ -158,41 +206,6 @@ chunk_tab_record_chunk(struct chunk_table *chunk_tab, unsigned out_chunk_size)
        }
 }
 
-/*
- * compress_func_t- Pointer to a function to compresses a chunk
- *                  of a WIM resource.  This may be either
- *                  wimlib_xpress_compress() (xpress-compress.c) or
- *                  wimlib_lzx_compress() (lzx-compress.c).
- *
- * @chunk:       Uncompressed data of the chunk.
- * @chunk_size:          Size of the uncompressed chunk, in bytes.
- * @out:         Pointer to output buffer of size at least (@chunk_size - 1) bytes.
- *
- * Returns the size of the compressed data written to @out in bytes, or 0 if the
- * data could not be compressed to (@chunk_size - 1) bytes or fewer.
- *
- * As a special requirement, the compression code is optimized for the WIM
- * format and therefore requires (@chunk_size <= 32768).
- *
- * As another special requirement, the compression code will read up to 8 bytes
- * off the end of the @chunk array for performance reasons.  The values of these
- * bytes will not affect the output of the compression, but the calling code
- * must make sure that the buffer holding the uncompressed chunk is actually at
- * least (@chunk_size + 8) bytes, or at least that these extra bytes are in
- * mapped memory that will not cause a memory access violation if accessed.
- */
-typedef unsigned (*compress_func_t)(const void *chunk, unsigned chunk_size,
-                                   void *out);
-
-static compress_func_t
-get_compress_func(int out_ctype)
-{
-       if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX)
-               return wimlib_lzx_compress;
-       else
-               return wimlib_xpress_compress;
-}
-
 /* Finishes a WIM chunk table and writes it to the output file at the correct
  * offset.  */
 static int
@@ -209,16 +222,14 @@ finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
                                         chunk_tab->bytes_per_chunk_entry,
                                 chunk_tab->table_disk_size);
        } else {
-               ret  = full_pwrite(out_fd,
-                                  chunk_tab->offsets +
-                                          chunk_tab->bytes_per_chunk_entry,
-                                  chunk_tab->table_disk_size,
-                                  res_start_offset);
-       }
-       if (ret) {
-               ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
-                                "file resource");
+               ret = full_pwrite(out_fd,
+                                 chunk_tab->offsets +
+                                         chunk_tab->bytes_per_chunk_entry,
+                                 chunk_tab->table_disk_size,
+                                 res_start_offset);
        }
+       if (ret)
+               ERROR_WITH_ERRNO("Write error");
        return ret;
 }
 
@@ -284,7 +295,9 @@ finalize_and_check_sha1(SHA_CTX *sha_ctx, struct wim_lookup_table_entry *lte)
 }
 
 struct write_resource_ctx {
-       compress_func_t compress;
+       int out_ctype;
+       u32 out_chunk_size;
+       struct wimlib_lzx_context *comp_ctx;
        struct chunk_table *chunk_tab;
        struct filedes *out_fd;
        SHA_CTX sha_ctx;
@@ -299,21 +312,32 @@ write_resource_cb(const void *chunk, size_t chunk_size, void *_ctx)
        const void *out_chunk;
        unsigned out_chunk_size;
        int ret;
+       void *compressed_chunk = NULL;
+       unsigned compressed_size;
+       bool compressed_chunk_malloced = false;
+       size_t stack_max = 32768;
 
        if (ctx->doing_sha)
                sha1_update(&ctx->sha_ctx, chunk, chunk_size);
 
        out_chunk = chunk;
        out_chunk_size = chunk_size;
-       if (ctx->compress) {
-               void *compressed_chunk;
-               unsigned compressed_size;
+       if (ctx->out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
 
                /* Compress the chunk.  */
-               compressed_chunk = alloca(chunk_size);
-               compressed_size = (*ctx->compress)(chunk, chunk_size,
-                                                  compressed_chunk);
+               if (chunk_size <= stack_max) {
+                       compressed_chunk = alloca(chunk_size);
+               } else {
+                       compressed_chunk = MALLOC(chunk_size);
+                       if (compressed_chunk == NULL)
+                               return WIMLIB_ERR_NOMEM;
+                       compressed_chunk_malloced = true;
+               }
 
+               compressed_size = compress_chunk(chunk, chunk_size,
+                                                compressed_chunk,
+                                                ctx->out_ctype,
+                                                ctx->comp_ctx);
                /* Use compressed data if compression to less than input size
                 * was successful.  */
                if (compressed_size) {
@@ -344,11 +368,15 @@ write_resource_cb(const void *chunk, size_t chunk_size, void *_ctx)
        ret = full_write(ctx->out_fd, out_chunk, out_chunk_size);
        if (ret)
                goto error;
-       return 0;
+
+out_free_memory:
+       if (compressed_chunk_malloced)
+               FREE(compressed_chunk);
+       return ret;
 
 error:
        ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
-       return ret;
+       goto out_free_memory;
 }
 
 /*
@@ -367,30 +395,40 @@ error:
  *     One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate which
  *     compression algorithm to use.
  *
+ * @out_chunk_size:
+ *     Compressed chunk size 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.
  *
- * @write_resource_flags:
+ * @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).
  *
+ * @comp_ctx:
+ *     Location of LZX compression context pointer, which will be allocated or
+ *     updated if needed.  (Initialize to NULL.)
+ *
  * Additional notes:  The SHA1 message digest of the uncompressed data is
  * calculated (except when doing a raw copy --- see below).  If the @unhashed
  * flag is set on the lookup table entry, this message digest is simply copied
  * to it; otherwise, the message digest is compared with the existing one, and
- * the function will fail if they do not match.
+ * this function will fail if they do not match.
  */
-int
+static int
 write_wim_resource(struct wim_lookup_table_entry *lte,
                   struct filedes *out_fd, int out_ctype,
+                  u32 out_chunk_size,
                   struct resource_entry *out_res_entry,
-                  int resource_flags)
+                  int resource_flags,
+                  struct wimlib_lzx_context **comp_ctx)
 {
        struct write_resource_ctx write_ctx;
        off_t res_start_offset;
+       u32 in_chunk_size;
        u64 read_size;
        int ret;
 
@@ -402,45 +440,55 @@ write_wim_resource(struct wim_lookup_table_entry *lte,
        res_start_offset = out_fd->offset;
 
        /* If we are not forcing the data to be recompressed, and the input
-        * resource is located in a WIM with the same compression type as that
-        * desired other than no compression, we can simply copy the compressed
-        * data without recompressing it.  This also means we must skip
-        * calculating the SHA1, as we never will see the uncompressed data.  */
-       if (lte->resource_location == RESOURCE_IN_WIM &&
-           out_ctype == wim_resource_compression_type(lte) &&
-           out_ctype != WIMLIB_COMPRESSION_TYPE_NONE &&
-           !(resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS))
-       {
-               /* Normally we can request a RAW_FULL read, but if we're reading
-                * from a pipable resource and writing a non-pipable resource or
-                * vice versa, then a RAW_CHUNKS read needs to be requested so
-                * that the written resource can be appropriately formatted.
-                * However, in neither case is any actual decompression needed.
-                */
+        * resource is located in a WIM with a compression mode compatible with
+        * the output, 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 (can_raw_copy(lte, resource_flags, out_ctype, out_chunk_size)) {
+               /* Normally, for raw copies 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
+                       read_size = lte->resource_entry.size;
+               } else {
                        resource_flags |= WIMLIB_READ_RESOURCE_FLAG_RAW_CHUNKS;
+                       read_size = lte->resource_entry.original_size;
+               }
                write_ctx.doing_sha = false;
-               read_size = lte->resource_entry.size;
        } else {
                write_ctx.doing_sha = true;
                sha1_init(&write_ctx.sha_ctx);
                read_size = lte->resource_entry.original_size;
        }
 
-       /* If the output resource is to be compressed, initialize the chunk
-        * table and set the function to use for chunk compression.  Exceptions:
-        * no compression function is needed if doing a raw copy; also, no chunk
-        * table is needed if doing a *full* (not per-chunk) raw copy.  */
-       write_ctx.compress = NULL;
+       /* Set the output compression mode and initialize chunk table if needed.
+        */
+       write_ctx.out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
+       write_ctx.out_chunk_size = out_chunk_size;
        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);
+               wimlib_assert(out_chunk_size > 0);
+               if (!(resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW)) {
+                       /* Compression needed.  */
+                       write_ctx.out_ctype = out_ctype;
+                       if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX) {
+                               ret = wimlib_lzx_alloc_context(out_chunk_size,
+                                                              NULL, comp_ctx);
+                               if (ret)
+                                       goto out;
+                       }
+                       write_ctx.comp_ctx = *comp_ctx;
+               }
                if (!(resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW_FULL)) {
+                       /* Chunk table needed.  */
                        ret = begin_wim_resource_chunk_tab(lte, out_fd,
+                                                          out_chunk_size,
                                                           &write_ctx.chunk_tab,
                                                           resource_flags);
                        if (ret)
@@ -461,12 +509,17 @@ write_wim_resource(struct wim_lookup_table_entry *lte,
        }
 
        /* Write the entire resource by reading the entire resource and feeding
-        * the data through the write_resource_cb function. */
+        * the data through write_resource_cb().  */
        write_ctx.out_fd = out_fd;
        write_ctx.resource_flags = resource_flags;
 try_write_again:
+       if (write_ctx.out_ctype == WIMLIB_COMPRESSION_TYPE_NONE)
+               in_chunk_size = wim_resource_chunk_size(lte);
+       else
+               in_chunk_size = out_chunk_size;
        ret = read_resource_prefix(lte, read_size,
-                                  write_resource_cb, &write_ctx, resource_flags);
+                                  write_resource_cb,
+                                  in_chunk_size, &write_ctx, resource_flags);
        if (ret)
                goto out_free_chunk_tab;
 
@@ -514,12 +567,12 @@ try_write_again:
                        goto out_free_chunk_tab;
                out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
                FREE(write_ctx.chunk_tab);
-               write_ctx.compress = NULL;
+               write_ctx.out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
                write_ctx.chunk_tab = NULL;
                write_ctx.doing_sha = false;
                goto try_write_again;
        }
-       if (resource_flags & (WIMLIB_READ_RESOURCE_FLAG_RAW)) {
+       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,
@@ -543,39 +596,50 @@ out:
 }
 
 /* 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.  */
+ * uncompressed data rather a lookup table entry.  Also writes the SHA1 message
+ * digest of the buffer to @hash_ret if it is non-NULL.  */
 int
 write_wim_resource_from_buffer(const void *buf, size_t buf_size,
                               int reshdr_flags, struct filedes *out_fd,
                               int out_ctype,
+                              u32 out_chunk_size,
                               struct resource_entry *out_res_entry,
-                              u8 *hash_ret, int write_resource_flags)
+                              u8 *hash_ret, int write_resource_flags,
+                              struct wimlib_lzx_context **comp_ctx)
 {
-       /* Set up a temporary lookup table entry to provide to
-        * write_wim_resource(). */
-       struct wim_lookup_table_entry lte;
        int ret;
+       struct wim_lookup_table_entry *lte;
 
-       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;
+       /* Set up a temporary lookup table entry to provide to
+        * write_wim_resource().  */
+
+       lte = new_lookup_table_entry();
+       if (lte == NULL)
+               return WIMLIB_ERR_NOMEM;
+
+       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;
+               sha1_buffer(buf, buf_size, lte->hash);
+               lte->unhashed = 0;
        } else {
-               lte.unhashed = 1;
+               lte->unhashed = 1;
        }
 
-       ret = write_wim_resource(&lte, out_fd, out_ctype, out_res_entry,
-                                write_resource_flags);
+       ret = write_wim_resource(lte, out_fd, out_ctype, out_chunk_size,
+                                out_res_entry, write_resource_flags, comp_ctx);
        if (ret)
-               return ret;
+               goto out_free_lte;
        if (hash_ret)
-               copy_hash(hash_ret, lte.hash);
-       return 0;
+               copy_hash(hash_ret, lte->hash);
+       ret = 0;
+out_free_lte:
+       lte->resource_location = RESOURCE_NONEXISTENT;
+       free_lookup_table_entry(lte);
+       return ret;
 }
 
 
@@ -671,13 +735,15 @@ shared_queue_get(struct shared_queue *q)
 struct compressor_thread_params {
        struct shared_queue *res_to_compress_queue;
        struct shared_queue *compressed_res_queue;
-       compress_func_t compress;
+       int out_ctype;
+       struct wimlib_lzx_context *comp_ctx;
 };
 
 #define MAX_CHUNKS_PER_MSG 2
 
 struct message {
        struct wim_lookup_table_entry *lte;
+       u32 out_chunk_size;
        u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
        u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
        unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
@@ -689,12 +755,18 @@ struct message {
 };
 
 static void
-compress_chunks(struct message *msg, compress_func_t compress)
+compress_chunks(struct message *msg, int out_ctype,
+               struct wimlib_lzx_context *comp_ctx)
 {
        for (unsigned i = 0; i < msg->num_chunks; i++) {
-               unsigned len = compress(msg->uncompressed_chunks[i],
-                                       msg->uncompressed_chunk_sizes[i],
-                                       msg->compressed_chunks[i]);
+               unsigned len;
+
+               len = compress_chunk(msg->uncompressed_chunks[i],
+                                    msg->uncompressed_chunk_sizes[i],
+                                    msg->compressed_chunks[i],
+                                    out_ctype,
+                                    comp_ctx);
+
                void *out_chunk;
                unsigned out_len;
                if (len) {
@@ -722,12 +794,11 @@ compressor_thread_proc(void *arg)
        struct compressor_thread_params *params = arg;
        struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
        struct shared_queue *compressed_res_queue = params->compressed_res_queue;
-       compress_func_t compress = params->compress;
        struct message *msg;
 
        DEBUG("Compressor thread ready");
        while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
-               compress_chunks(msg, compress);
+               compress_chunks(msg, params->out_ctype, params->comp_ctx);
                shared_queue_put(compressed_res_queue, msg);
        }
        DEBUG("Compressor thread terminating");
@@ -791,6 +862,8 @@ do_write_streams_progress(struct write_streams_progress_data *progress_data,
 struct serial_write_stream_ctx {
        struct filedes *out_fd;
        int out_ctype;
+       u32 out_chunk_size;
+       struct wimlib_lzx_context **comp_ctx;
        int write_resource_flags;
 };
 
@@ -799,8 +872,11 @@ serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
 {
        struct serial_write_stream_ctx *ctx = _ctx;
        return write_wim_resource(lte, ctx->out_fd,
-                                 ctx->out_ctype, &lte->output_resource_entry,
-                                 ctx->write_resource_flags);
+                                 ctx->out_ctype,
+                                 ctx->out_chunk_size,
+                                 &lte->output_resource_entry,
+                                 ctx->write_resource_flags,
+                                 ctx->comp_ctx);
 }
 
 
@@ -876,7 +952,7 @@ do_write_stream_list(struct list_head *stream_list,
                if (ret)
                        break;
                /* In parallel mode, some streams are deferred for later,
-                * serialized processing; ignore them here. */
+                * serialized processing; ignore them here.  */
                if (lte->deferred)
                        continue;
                if (lte->unhashed) {
@@ -898,13 +974,17 @@ do_write_stream_list_serial(struct list_head *stream_list,
                            struct wim_lookup_table *lookup_table,
                            struct filedes *out_fd,
                            int out_ctype,
+                           u32 out_chunk_size,
+                           struct wimlib_lzx_context **comp_ctx,
                            int write_resource_flags,
                            struct write_streams_progress_data *progress_data)
 {
        struct serial_write_stream_ctx ctx = {
                .out_fd = out_fd,
                .out_ctype = out_ctype,
+               .out_chunk_size = out_chunk_size,
                .write_resource_flags = write_resource_flags,
+               .comp_ctx = comp_ctx,
        };
        return do_write_stream_list(stream_list,
                                    lookup_table,
@@ -930,6 +1010,8 @@ write_stream_list_serial(struct list_head *stream_list,
                         struct wim_lookup_table *lookup_table,
                         struct filedes *out_fd,
                         int out_ctype,
+                        u32 out_chunk_size,
+                        struct wimlib_lzx_context **comp_ctx,
                         int write_resource_flags,
                         struct write_streams_progress_data *progress_data)
 {
@@ -945,6 +1027,8 @@ write_stream_list_serial(struct list_head *stream_list,
                                           lookup_table,
                                           out_fd,
                                           out_ctype,
+                                          out_chunk_size,
+                                          comp_ctx,
                                           write_resource_flags,
                                           progress_data);
 }
@@ -984,7 +1068,7 @@ write_wim_chunks(struct message *msg, struct filedes *out_fd,
        }
        ret = full_writev(out_fd, vecs, nvecs);
        if (ret)
-               ERROR_WITH_ERRNO("Failed to write WIM chunks");
+               ERROR_WITH_ERRNO("Write error");
        return ret;
 }
 
@@ -994,6 +1078,8 @@ struct main_writer_thread_ctx {
        struct filedes *out_fd;
        off_t res_start_offset;
        int out_ctype;
+       u32 out_chunk_size;
+       struct wimlib_lzx_context **comp_ctx;
        int write_resource_flags;
        struct shared_queue *res_to_compress_queue;
        struct shared_queue *compressed_res_queue;
@@ -1016,11 +1102,12 @@ struct main_writer_thread_ctx {
 };
 
 static int
-init_message(struct message *msg)
+init_message(struct message *msg, u32 out_chunk_size)
 {
+       msg->out_chunk_size = out_chunk_size;
        for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
-               msg->compressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
-               msg->uncompressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
+               msg->compressed_chunks[i] = MALLOC(out_chunk_size);
+               msg->uncompressed_chunks[i] = MALLOC(out_chunk_size);
                if (msg->compressed_chunks[i] == NULL ||
                    msg->uncompressed_chunks[i] == NULL)
                        return WIMLIB_ERR_NOMEM;
@@ -1048,15 +1135,15 @@ free_messages(struct message *msgs, size_t num_messages)
 }
 
 static struct message *
-allocate_messages(size_t num_messages)
+allocate_messages(size_t num_messages, u32 out_chunk_size)
 {
        struct message *msgs;
 
        msgs = CALLOC(num_messages, sizeof(struct message));
-       if (!msgs)
+       if (msgs == NULL)
                return NULL;
        for (size_t i = 0; i < num_messages; i++) {
-               if (init_message(&msgs[i])) {
+               if (init_message(&msgs[i], out_chunk_size)) {
                        free_messages(msgs, num_messages);
                        return NULL;
                }
@@ -1078,8 +1165,8 @@ main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
 {
        /* Pre-allocate all the buffers that will be needed to do the chunk
         * compression. */
-       ctx->msgs = allocate_messages(ctx->num_messages);
-       if (!ctx->msgs)
+       ctx->msgs = allocate_messages(ctx->num_messages, ctx->out_chunk_size);
+       if (ctx->msgs == NULL)
                return WIMLIB_ERR_NOMEM;
 
        /* Initially, all the messages are available to use. */
@@ -1099,12 +1186,14 @@ main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
        INIT_LIST_HEAD(&ctx->outstanding_streams);
        ctx->num_outstanding_messages = 0;
 
+       /* Message currently being prepared.  */
        ctx->next_msg = NULL;
 
        /* Resources that don't need any chunks compressed are added to this
-        * list and written directly by the main thread. */
+        * list and written directly by the main thread.  */
        INIT_LIST_HEAD(&ctx->serial_streams);
 
+       /* Pointer to chunk table for stream currently being written.  */
        ctx->cur_chunk_tab = NULL;
 
        return 0;
@@ -1126,17 +1215,17 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
 
        /* Get the next message from the queue and process it.
         * The message will contain 1 or more data chunks that have been
-        * compressed. */
+        * compressed.  */
        msg = shared_queue_get(ctx->compressed_res_queue);
        msg->complete = true;
        --ctx->num_outstanding_messages;
 
-       /* Is this the next chunk in the current resource?  If it's not
-        * (i.e., an earlier chunk in a same or different resource
-        * hasn't been compressed yet), do nothing, and keep this
-        * message around until all earlier chunks are received.
+       /* Is this the next chunk in the current resource?  If it's not (i.e.,
+        * an earlier chunk in a same or different resource hasn't been
+        * compressed yet), do nothing, and keep this message around until all
+        * earlier chunks are received.
         *
-        * Otherwise, write all the chunks we can. */
+        * Otherwise, write all the chunks we can.  */
        while (cur_lte != NULL &&
               !list_empty(&cur_lte->msg_list)
               && (msg = container_of(cur_lte->msg_list.next,
@@ -1164,21 +1253,21 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                         * it if needed.  */
                        ret = begin_wim_resource_chunk_tab(cur_lte,
                                                           ctx->out_fd,
+                                                          ctx->out_chunk_size,
                                                           &ctx->cur_chunk_tab,
                                                           ctx->write_resource_flags);
                        if (ret)
                                return ret;
-
                }
 
-               /* Write the compressed chunks from the message. */
+               /* Write the compressed chunks from the message.  */
                ret = write_wim_chunks(msg, ctx->out_fd, ctx->cur_chunk_tab,
                                       ctx->write_resource_flags);
                if (ret)
                        return ret;
 
-               /* Was this the last chunk of the stream?  If so, finish
-                * it. */
+               /* Was this the last chunk of the stream?  If so, finish the
+                * stream by writing the chunk table.  */
                if (list_empty(&cur_lte->msg_list) &&
                    msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
                {
@@ -1214,8 +1303,10 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                                ret = write_wim_resource(cur_lte,
                                                         ctx->out_fd,
                                                         WIMLIB_COMPRESSION_TYPE_NONE,
+                                                        0,
                                                         &cur_lte->output_resource_entry,
-                                                        ctx->write_resource_flags);
+                                                        ctx->write_resource_flags,
+                                                        ctx->comp_ctx);
                                if (ret)
                                        return ret;
                        } else {
@@ -1248,19 +1339,21 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                         * list for direct writing by the main thread (e.g.
                         * resources that don't need to be compressed because
                         * the desired compression type is the same as the
-                        * previous compression type). */
+                        * previous compression type).  */
                        if (!list_empty(&ctx->serial_streams)) {
                                ret = do_write_stream_list_serial(&ctx->serial_streams,
                                                                  ctx->lookup_table,
                                                                  ctx->out_fd,
                                                                  ctx->out_ctype,
+                                                                 ctx->out_chunk_size,
+                                                                 ctx->comp_ctx,
                                                                  ctx->write_resource_flags,
                                                                  ctx->progress_data);
                                if (ret)
                                        return ret;
                        }
 
-                       /* Advance to the next stream to write. */
+                       /* Advance to the next stream to write.  */
                        if (list_empty(&ctx->outstanding_streams)) {
                                cur_lte = NULL;
                        } else {
@@ -1273,7 +1366,7 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
        return 0;
 }
 
-/* Called when the main thread has read a new chunk of data. */
+/* Called when the main thread has read a new chunk of data.  */
 static int
 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
 {
@@ -1352,6 +1445,8 @@ main_writer_thread_finish(void *_ctx)
                                           ctx->lookup_table,
                                           ctx->out_fd,
                                           ctx->out_ctype,
+                                          ctx->out_chunk_size,
+                                          ctx->comp_ctx,
                                           ctx->write_resource_flags,
                                           ctx->progress_data);
 }
@@ -1365,15 +1460,17 @@ submit_stream_for_compression(struct wim_lookup_table_entry *lte,
        /* Read the entire stream @lte, feeding its data chunks to the
         * compressor threads.  Also SHA1-sum the stream; this is required in
         * the case that @lte is unhashed, and a nice additional verification
-        * when @lte is already hashed. */
+        * when @lte is already hashed.  */
        sha1_init(&ctx->next_sha_ctx);
        ctx->next_chunk = 0;
-       ctx->next_num_chunks = wim_resource_chunks(lte);
+       ctx->next_num_chunks = DIV_ROUND_UP(wim_resource_size(lte),
+                                           ctx->out_chunk_size);
        ctx->next_lte = lte;
        INIT_LIST_HEAD(&lte->msg_list);
        list_add_tail(&lte->being_compressed_list, &ctx->outstanding_streams);
        ret = read_resource_prefix(lte, wim_resource_size(lte),
-                                  main_writer_thread_cb, ctx, 0);
+                                  main_writer_thread_cb,
+                                  ctx->out_chunk_size, ctx, 0);
        if (ret)
                return ret;
        wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
@@ -1387,15 +1484,13 @@ main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
        int ret;
 
        if (wim_resource_size(lte) < 1000 ||
-           ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
-           (lte->resource_location == RESOURCE_IN_WIM &&
-            !(ctx->write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS) &&
-            lte->wim->compression_type == ctx->out_ctype))
+           !must_compress_stream(lte, ctx->write_resource_flags,
+                                 ctx->out_ctype, ctx->out_chunk_size))
        {
                /* Stream is too small or isn't being compressed.  Process it by
                 * the main thread when we have a chance.  We can't necessarily
                 * process it right here, as the main thread could be in the
-                * middle of writing a different stream. */
+                * middle of writing a different stream.  */
                list_add_tail(&lte->write_streams_list, &ctx->serial_streams);
                lte->deferred = 1;
                ret = 0;
@@ -1422,15 +1517,15 @@ get_default_num_threads(void)
  * create the number of threads requested.
  *
  * High level description of the algorithm for writing compressed streams in
- * parallel:  We perform compression on chunks of size WIM_CHUNK_SIZE bytes
- * rather than on full files.  The currently executing thread becomes the main
- * thread and is entirely in charge of reading the data to compress (which may
- * be in any location understood by the resource code--- such as in an external
- * file being captured, or in another WIM file from which an image is being
- * exported) and actually writing the compressed data to the output file.
- * Additional threads are "compressor threads" and all execute the
- * compressor_thread_proc, where they repeatedly retrieve buffers of data from
- * the main thread, compress them, and hand them back to the main thread.
+ * parallel:  We perform compression on chunks rather than on full files.  The
+ * currently executing thread becomes the main thread and is entirely in charge
+ * of reading the data to compress (which may be in any location understood by
+ * the resource code--- such as in an external file being captured, or in
+ * another WIM file from which an image is being exported) and actually writing
+ * the compressed data to the output file.  Additional threads are "compressor
+ * threads" and all execute the compressor_thread_proc, where they repeatedly
+ * retrieve buffers of data from the main thread, compress them, and hand them
+ * back to the main thread.
  *
  * Certain streams, such as streams that do not need to be compressed (e.g.
  * input compression type same as output compression type) or streams of very
@@ -1440,13 +1535,14 @@ get_default_num_threads(void)
  * At any given point in time, multiple streams may be having chunks compressed
  * concurrently.  The stream that the main thread is currently *reading* may be
  * later in the list that the stream that the main thread is currently
- * *writing*.
- */
+ * *writing*.  */
 static int
 write_stream_list_parallel(struct list_head *stream_list,
                           struct wim_lookup_table *lookup_table,
                           struct filedes *out_fd,
                           int out_ctype,
+                          u32 out_chunk_size,
+                          struct wimlib_lzx_context **comp_ctx,
                           int write_resource_flags,
                           struct write_streams_progress_data *progress_data,
                           unsigned num_threads)
@@ -1456,12 +1552,14 @@ write_stream_list_parallel(struct list_head *stream_list,
        struct shared_queue compressed_res_queue;
        pthread_t *compressor_threads = NULL;
        union wimlib_progress_info *progress = &progress_data->progress;
+       unsigned num_started_threads;
+       bool can_retry = true;
 
        if (num_threads == 0) {
                long nthreads = get_default_num_threads();
                if (nthreads < 1 || nthreads > UINT_MAX) {
                        WARNING("Could not determine number of processors! Assuming 1");
-                       goto out_serial;
+                       goto out_serial_quiet;
                } else if (nthreads == 1) {
                        goto out_serial_quiet;
                } else {
@@ -1488,30 +1586,47 @@ write_stream_list_parallel(struct list_head *stream_list,
        if (ret)
                goto out_destroy_res_to_compress_queue;
 
-       struct compressor_thread_params params;
-       params.res_to_compress_queue = &res_to_compress_queue;
-       params.compressed_res_queue = &compressed_res_queue;
-       params.compress = get_compress_func(out_ctype);
+       struct compressor_thread_params *params;
 
-       compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
-       if (!compressor_threads) {
+       params = CALLOC(num_threads, sizeof(params[0]));
+       if (params == NULL) {
                ret = WIMLIB_ERR_NOMEM;
                goto out_destroy_compressed_res_queue;
        }
 
+       for (unsigned i = 0; i < num_threads; i++) {
+               params[i].res_to_compress_queue = &res_to_compress_queue;
+               params[i].compressed_res_queue = &compressed_res_queue;
+               params[i].out_ctype = out_ctype;
+               if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX) {
+                       ret = wimlib_lzx_alloc_context(out_chunk_size,
+                                                      NULL, &params[i].comp_ctx);
+                       if (ret)
+                               goto out_free_params;
+               }
+       }
+
+       compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
+       if (compressor_threads == NULL) {
+               ret = WIMLIB_ERR_NOMEM;
+               goto out_free_params;
+       }
+
        for (unsigned i = 0; i < num_threads; i++) {
                DEBUG("pthread_create thread %u of %u", i + 1, num_threads);
                ret = pthread_create(&compressor_threads[i], NULL,
-                                    compressor_thread_proc, &params);
-               if (ret != 0) {
+                                    compressor_thread_proc, &params[i]);
+               if (ret) {
+                       errno = ret;
                        ret = -1;
                        ERROR_WITH_ERRNO("Failed to create compressor "
                                         "thread %u of %u",
                                         i + 1, num_threads);
-                       num_threads = i;
+                       num_started_threads = i;
                        goto out_join;
                }
        }
+       num_started_threads = num_threads;
 
        if (progress_data->progress_func) {
                progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
@@ -1519,10 +1634,15 @@ write_stream_list_parallel(struct list_head *stream_list,
        }
 
        struct main_writer_thread_ctx ctx;
+
+       memset(&ctx, 0, sizeof(ctx));
+
        ctx.stream_list           = stream_list;
        ctx.lookup_table          = lookup_table;
        ctx.out_fd                = out_fd;
        ctx.out_ctype             = out_ctype;
+       ctx.out_chunk_size        = out_chunk_size;
+       ctx.comp_ctx              = comp_ctx;
        ctx.res_to_compress_queue = &res_to_compress_queue;
        ctx.compressed_res_queue  = &compressed_res_queue;
        ctx.num_messages          = queue_size;
@@ -1531,6 +1651,8 @@ write_stream_list_parallel(struct list_head *stream_list,
        ret = main_writer_thread_init_ctx(&ctx);
        if (ret)
                goto out_join;
+
+       can_retry = false;
        ret = do_write_stream_list(stream_list, lookup_table,
                                   main_thread_process_next_stream,
                                   &ctx, progress_data);
@@ -1542,15 +1664,15 @@ write_stream_list_parallel(struct list_head *stream_list,
         * chunks to be compressed so that the remaining streams can actually be
         * written to the output file.  Furthermore, any remaining streams that
         * had processing deferred to the main thread need to be handled.  These
-        * tasks are done by the main_writer_thread_finish() function. */
+        * tasks are done by the main_writer_thread_finish() function.  */
        ret = main_writer_thread_finish(&ctx);
 out_destroy_ctx:
        main_writer_thread_destroy_ctx(&ctx);
 out_join:
-       for (unsigned i = 0; i < num_threads; i++)
+       for (unsigned i = 0; i < num_started_threads; i++)
                shared_queue_put(&res_to_compress_queue, NULL);
 
-       for (unsigned i = 0; i < num_threads; i++) {
+       for (unsigned i = 0; i < num_started_threads; i++) {
                if (pthread_join(compressor_threads[i], NULL)) {
                        WARNING_WITH_ERRNO("Failed to join compressor "
                                           "thread %u of %u",
@@ -1558,11 +1680,15 @@ out_join:
                }
        }
        FREE(compressor_threads);
+out_free_params:
+       for (unsigned i = 0; i < num_threads; i++)
+               wimlib_lzx_free_context(params[i].comp_ctx);
+       FREE(params);
 out_destroy_compressed_res_queue:
        shared_queue_destroy(&compressed_res_queue);
 out_destroy_res_to_compress_queue:
        shared_queue_destroy(&res_to_compress_queue);
-       if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
+       if (!can_retry || (ret >= 0 && ret != WIMLIB_ERR_NOMEM))
                return ret;
 out_serial:
        WARNING("Falling back to single-threaded compression");
@@ -1571,31 +1697,35 @@ out_serial_quiet:
                                        lookup_table,
                                        out_fd,
                                        out_ctype,
+                                       out_chunk_size,
+                                       comp_ctx,
                                        write_resource_flags,
                                        progress_data);
 
 }
 #endif
 
-/*
- * Write a list of streams to a WIM (@out_fd) using the compression type
- * @out_ctype and up to @num_threads compressor threads.
- */
+/* Write a list of streams to a WIM (@out_fd) using the compression type
+ * @out_ctype, chunk size @out_chunk_size, and up to @num_threads compressor
+ * threads.  */
 static int
 write_stream_list(struct list_head *stream_list,
                  struct wim_lookup_table *lookup_table,
-                 struct filedes *out_fd, int out_ctype, int write_flags,
+                 struct filedes *out_fd, int out_ctype,
+                 u32 out_chunk_size,
+                 struct wimlib_lzx_context **comp_ctx,
+                 int write_flags,
                  unsigned num_threads, wimlib_progress_func_t progress_func)
 {
-       struct wim_lookup_table_entry *lte;
-       size_t num_streams = 0;
-       u64 total_bytes = 0;
-       u64 total_compression_bytes = 0;
-       struct write_streams_progress_data progress_data;
        int ret;
        int write_resource_flags;
-       unsigned total_parts = 0;
-       WIMStruct *prev_wim_part = NULL;
+       u64 total_bytes;
+       u64 total_compression_bytes;
+       unsigned total_parts;
+       WIMStruct *prev_wim_part;
+       size_t num_streams;
+       struct wim_lookup_table_entry *lte;
+       struct write_streams_progress_data progress_data;
 
        if (list_empty(stream_list)) {
                DEBUG("No streams to write.");
@@ -1607,23 +1737,28 @@ write_stream_list(struct list_head *stream_list,
        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));
+       /* Sort the stream list into a good order for reading.  */
+       ret = sort_stream_list_by_sequential_order(stream_list,
+                                                  offsetof(struct wim_lookup_table_entry,
+                                                           write_streams_list));
+       if (ret)
+               return ret;
 
        /* Calculate the total size of the streams to be written.  Note: this
         * will be the uncompressed size, as we may not know the compressed size
         * yet, and also this will assume that every unhashed stream will be
-        * written (which will not necessarily be the case). */
+        * written (which will not necessarily be the case).  */
+       total_bytes = 0;
+       total_compression_bytes = 0;
+       num_streams = 0;
+       total_parts = 0;
+       prev_wim_part = NULL;
        list_for_each_entry(lte, stream_list, write_streams_list) {
                num_streams++;
                total_bytes += wim_resource_size(lte);
-               if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
-                      && (wim_resource_compression_type(lte) != out_ctype ||
-                          (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS)))
-               {
+               if (must_compress_stream(lte, write_resource_flags,
+                                        out_ctype, out_chunk_size))
                        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;
@@ -1653,6 +1788,8 @@ write_stream_list(struct list_head *stream_list,
                                                 lookup_table,
                                                 out_fd,
                                                 out_ctype,
+                                                out_chunk_size,
+                                                comp_ctx,
                                                 write_resource_flags,
                                                 &progress_data,
                                                 num_threads);
@@ -1662,12 +1799,14 @@ write_stream_list(struct list_head *stream_list,
                                               lookup_table,
                                               out_fd,
                                               out_ctype,
+                                              out_chunk_size,
+                                              comp_ctx,
                                               write_resource_flags,
                                               &progress_data);
        if (ret == 0)
                DEBUG("Successfully wrote stream list.");
        else
-               DEBUG("Failed to write stream list.");
+               DEBUG("Failed to write stream list (ret=%d).", ret);
        return ret;
 }
 
@@ -1983,7 +2122,9 @@ write_wim_streams(WIMStruct *wim, int image, int write_flags,
        return write_stream_list(stream_list,
                                 wim->lookup_table,
                                 &wim->out_fd,
-                                wim->compression_type,
+                                wim->out_compression_type,
+                                wim->out_chunk_size,
+                                &wim->lzx_context,
                                 write_flags,
                                 num_threads,
                                 progress_func);
@@ -2042,9 +2183,11 @@ write_wim_metadata_resources(WIMStruct *wim, int image, int write_flags,
                              "metadata resource.", i);
                        ret = write_wim_resource(imd->metadata_lte,
                                                 &wim->out_fd,
-                                                wim->compression_type,
+                                                wim->out_compression_type,
+                                                wim->out_chunk_size,
                                                 &imd->metadata_lte->output_resource_entry,
-                                                write_resource_flags);
+                                                write_resource_flags,
+                                                &wim->lzx_context);
                }
                if (ret)
                        return ret;
@@ -2074,18 +2217,22 @@ close_wim_writable(WIMStruct *wim, int write_flags)
 {
        int ret = 0;
 
-       if (!(write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR))
+       if (!(write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)) {
+               DEBUG("Closing WIM file.");
                if (filedes_valid(&wim->out_fd))
                        if (filedes_close(&wim->out_fd))
                                ret = WIMLIB_ERR_WRITE;
+       }
        filedes_invalidate(&wim->out_fd);
        return ret;
 }
 
 /*
+ * finish_write():
+ *
  * Finish writing a WIM file: write the lookup table, xml data, and integrity
- * table, then overwrite the WIM header.  Always closes the WIM file descriptor
- * (wim->out_fd).
+ * table, then overwrite the WIM header.  By default, closes the WIM file
+ * descriptor (@wim->out_fd) if successful.
  *
  * write_flags is a bitwise OR of the following:
  *
@@ -2095,7 +2242,7 @@ close_wim_writable(WIMStruct *wim, int write_flags)
  *     (public) WIMLIB_WRITE_FLAG_FSYNC:
  *             fsync() the output file before closing it.
  *
- *     (public)  WIMLIB_WRITE_FLAG_PIPABLE:
+ *     (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
@@ -2118,6 +2265,9 @@ close_wim_writable(WIMStruct *wim, int write_flags)
  *             Instead of overwriting the WIM header at the beginning of the
  *             file, simply append it to the end of the file.  (Used when
  *             writing to pipe.)
+ *     (private) WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR:
+ *             Do not close the file descriptor @wim->out_fd on either success
+ *             on failure.
  *     (private) WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES:
  *             Use the existing <TOTALBYTES> stored in the in-memory XML
  *             information, rather than setting it to the offset of the XML
@@ -2135,6 +2285,8 @@ finish_write(WIMStruct *wim, int image, int write_flags,
        off_t new_lookup_table_end;
        u64 xml_totalbytes;
 
+       DEBUG("image=%d, write_flags=%08x", image, write_flags);
+
        write_resource_flags = write_flags_to_resource_flags(write_flags);
 
        /* In the WIM header, there is room for the resource entry for a
@@ -2158,7 +2310,7 @@ finish_write(WIMStruct *wim, int image, int write_flags,
                                             &wim->hdr.lookup_table_res_entry,
                                             stream_list_override);
                if (ret)
-                       goto out_close_wim;
+                       return ret;
        }
 
        /* Write XML data.  */
@@ -2169,8 +2321,9 @@ finish_write(WIMStruct *wim, int image, int write_flags,
                                 &wim->hdr.xml_res_entry,
                                 write_resource_flags);
        if (ret)
-               goto out_close_wim;
+               return ret;
 
+       /* Write integrity table (optional).  */
        if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
                if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
                        struct wim_header checkpoint_hdr;
@@ -2180,7 +2333,7 @@ finish_write(WIMStruct *wim, int image, int write_flags,
                        ret = write_wim_header_at_offset(&checkpoint_hdr,
                                                         &wim->out_fd, 0);
                        if (ret)
-                               goto out_close_wim;
+                               return ret;
                }
 
                if (!(write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE))
@@ -2194,36 +2347,45 @@ finish_write(WIMStruct *wim, int image, int write_flags,
                                            old_lookup_table_end,
                                            progress_func);
                if (ret)
-                       goto out_close_wim;
+                       return ret;
        } else {
+               /* No integrity table.  */
                zero_resource_entry(&wim->hdr.integrity);
        }
 
+       /* Now that all information in the WIM header has been determined, the
+        * preliminary header written earlier can be overwritten, the header of
+        * the existing WIM file can be overwritten, or the final header can be
+        * written to the end of the pipable WIM.  */
        wim->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
        hdr_offset = 0;
        if (write_flags & WIMLIB_WRITE_FLAG_HEADER_AT_END)
                hdr_offset = wim->out_fd.offset;
+       DEBUG("Writing new header @ %"PRIu64".", hdr_offset);
        ret = write_wim_header_at_offset(&wim->hdr, &wim->out_fd, hdr_offset);
        if (ret)
-               goto out_close_wim;
+               return ret;
 
+       /* Possibly sync file data to disk before closing.  On POSIX systems, it
+        * is necessary to do this before using rename() to overwrite an
+        * existing file with a new file.  Otherwise, data loss would occur if
+        * the system is abruptly terminated when the metadata for the rename
+        * operation has been written to disk, but the new file data has not.
+        */
        if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
+               DEBUG("Syncing WIM file.");
                if (fsync(wim->out_fd.fd)) {
                        ERROR_WITH_ERRNO("Error syncing data to WIM file");
-                       ret = WIMLIB_ERR_WRITE;
-                       goto out_close_wim;
+                       return WIMLIB_ERR_WRITE;
                }
        }
 
-       ret = 0;
-out_close_wim:
        if (close_wim_writable(wim, write_flags)) {
-               if (ret == 0) {
-                       ERROR_WITH_ERRNO("Failed to close the output WIM file");
-                       ret = WIMLIB_ERR_WRITE;
-               }
+               ERROR_WITH_ERRNO("Failed to close the output WIM file");
+               return WIMLIB_ERR_WRITE;
        }
-       return ret;
+
+       return 0;
 }
 
 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
@@ -2253,8 +2415,10 @@ lock_wim(WIMStruct *wim, int fd)
 #endif
 
 /*
+ * write_pipable_wim():
+ *
  * Perform the intermediate stages of creating a "pipable" WIM (i.e. a WIM
- * capable of being applied from a pipe).  Such a WIM looks like:
+ * 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
@@ -2263,7 +2427,7 @@ lock_wim(WIMStruct *wim, int fd)
  *
  * - 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
- *   should stop other software from trying to read the file as a normal WIM.
+ *   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
@@ -2285,10 +2449,13 @@ lock_wim(WIMStruct *wim, int fd)
  *   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 have no 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.  However, the offsets are
- *   given in the chunk table as if these chunk headers were not present.
+ *   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,
@@ -2305,15 +2472,20 @@ lock_wim(WIMStruct *wim, int fd)
  *
  *   Layout of pipable WIM:
  *
- * ----------+----------+--------------------+----------------+--------------+------------+--------+
- * | Header  | XML data | Metadata resources | File resources | Lookup table | XML data   | Header |
- * ----------+----------+--------------------+----------------+--------------+------------+--------+
+ * ---------+----------+--------------------+----------------+--------------+-----------+--------+
+ * | Header | XML data | Metadata resources | File resources | Lookup table | XML data  | Header |
+ * ---------+----------+--------------------+----------------+--------------+-----------+--------+
  *
  *   Layout of normal WIM:
  *
- * +---------+--------------------+----------------+--------------+----------+
- * | Header  | Metadata resources | File resources | Lookup table | XML data |
- * +---------+--------------------+----------------+--------------+----------+
+ * +--------+-----------------------------+-------------------------+
+ * | 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
@@ -2495,6 +2667,15 @@ write_wim_part(WIMStruct *wim,
        wim->hdr.part_number = part_number;
        wim->hdr.total_parts = total_parts;
 
+       /* Set compression type if different.  */
+       if (wim->compression_type != wim->out_compression_type) {
+               ret = set_wim_hdr_cflags(wim->out_compression_type, &wim->hdr);
+               wimlib_assert(ret == 0);
+       }
+
+       /* Set chunk size if different.  */
+       wim->hdr.chunk_size = wim->out_chunk_size;
+
        /* Use GUID if specified; otherwise generate a new one.  */
        if (guid)
                memcpy(wim->hdr.guid, guid, WIMLIB_GUID_LEN);
@@ -2597,7 +2778,8 @@ write_wim_part(WIMStruct *wim,
                           stream_list_override);
 out_restore_hdr:
        memcpy(&wim->hdr, &hdr_save, sizeof(struct wim_header));
-       close_wim_writable(wim, write_flags);
+       (void)close_wim_writable(wim, write_flags);
+       DEBUG("ret=%d", ret);
        return ret;
 }
 
@@ -2751,7 +2933,7 @@ overwrite_wim_inplace(WIMStruct *wim, int write_flags,
        struct list_head stream_list;
        off_t old_wim_end;
        u64 old_lookup_table_end, old_xml_begin, old_xml_end;
-
+       struct wim_header hdr_save;
 
        DEBUG("Overwriting `%"TS"' in-place", wim->filename);
 
@@ -2822,31 +3004,32 @@ overwrite_wim_inplace(WIMStruct *wim, int write_flags,
                return ret;
 
        ret = lock_wim(wim, wim->out_fd.fd);
-       if (ret) {
-               close_wim_writable(wim, write_flags);
-               return ret;
-       }
+       if (ret)
+               goto out_close_wim;
+
+       /* Save original header so it can be restored in case of error  */
+       memcpy(&hdr_save, &wim->hdr, sizeof(struct wim_header));
 
        /* Set WIM_HDR_FLAG_WRITE_IN_PROGRESS flag in header. */
-       ret = write_wim_header_flags(wim->hdr.flags | WIM_HDR_FLAG_WRITE_IN_PROGRESS,
-                                    &wim->out_fd);
+       wim->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
+       ret = write_wim_header_flags(wim->hdr.flags, &wim->out_fd);
        if (ret) {
                ERROR_WITH_ERRNO("Error updating WIM header flags");
-               close_wim_writable(wim, write_flags);
-               goto out_unlock_wim;
+               goto out_restore_memory_hdr;
        }
 
        if (filedes_seek(&wim->out_fd, old_wim_end) == -1) {
                ERROR_WITH_ERRNO("Can't seek to end of WIM");
-               close_wim_writable(wim, write_flags);
                ret = WIMLIB_ERR_WRITE;
-               goto out_unlock_wim;
+               goto out_restore_physical_hdr;
        }
 
        ret = write_stream_list(&stream_list,
                                wim->lookup_table,
                                &wim->out_fd,
                                wim->compression_type,
+                               wim->chunk_size,
+                               &wim->lzx_context,
                                write_flags,
                                num_threads,
                                progress_func);
@@ -2861,15 +3044,25 @@ overwrite_wim_inplace(WIMStruct *wim, int write_flags,
        write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
        ret = finish_write(wim, WIMLIB_ALL_IMAGES, write_flags,
                           progress_func, NULL);
+       if (ret)
+               goto out_truncate;
+
+       goto out_unlock_wim;
+
 out_truncate:
-       close_wim_writable(wim, write_flags);
-       if (ret && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
+       if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
                WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
                        wim->filename, old_wim_end);
-               /* Return value of truncate() is ignored because this is already
-                * an error path. */
-               (void)ttruncate(wim->filename, old_wim_end);
+               /* Return value of ftruncate() is ignored because this is
+                * already an error path.  */
+               (void)ftruncate(wim->out_fd.fd, old_wim_end);
        }
+out_restore_physical_hdr:
+       (void)write_wim_header_flags(hdr_save.flags, &wim->out_fd);
+out_restore_memory_hdr:
+       memcpy(&wim->hdr, &hdr_save, sizeof(struct wim_header));
+out_close_wim:
+       (void)close_wim_writable(wim, write_flags);
 out_unlock_wim:
        wim->wim_locked = 0;
        return ret;
@@ -2896,18 +3089,28 @@ overwrite_wim_via_tmpfile(WIMStruct *wim, int write_flags,
        ret = wimlib_write(wim, tmpfile, WIMLIB_ALL_IMAGES,
                           write_flags | WIMLIB_WRITE_FLAG_FSYNC,
                           num_threads, progress_func);
-       if (ret)
-               goto out_unlink;
+       if (ret) {
+               tunlink(tmpfile);
+               return ret;
+       }
 
        close_wim(wim);
 
+       /* Rename the new WIM file to the original WIM file.  Note: on Windows
+        * this actually calls win32_rename_replacement(), not _wrename(), so
+        * that removing the existing destination file can be handled.  */
        DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, wim->filename);
-       /* Rename the new file to the old file .*/
-       if (trename(tmpfile, wim->filename) != 0) {
+       ret = trename(tmpfile, wim->filename);
+       if (ret) {
                ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
                                 tmpfile, wim->filename);
-               ret = WIMLIB_ERR_RENAME;
-               goto out_unlink;
+       #ifdef __WIN32__
+               if (ret < 0)
+       #endif
+               {
+                       tunlink(tmpfile);
+               }
+               return WIMLIB_ERR_RENAME;
        }
 
        if (progress_func) {
@@ -2917,11 +3120,6 @@ overwrite_wim_via_tmpfile(WIMStruct *wim, int write_flags,
                progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
        }
        return 0;
-
-out_unlink:
-       /* Remove temporary file. */
-       tunlink(tmpfile);
-       return ret;
 }
 
 /* API function documented in wimlib.h  */
@@ -2952,7 +3150,9 @@ wimlib_overwrite(WIMStruct *wim, int write_flags,
        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)))
+           && !(wim_is_pipable(wim))
+           && wim->compression_type == wim->out_compression_type
+           && wim->chunk_size == wim->out_chunk_size)
        {
                ret = overwrite_wim_inplace(wim, write_flags, num_threads,
                                            progress_func);