]> wimlib.net Git - wimlib/blobdiff - src/write.c
Win32: Do not create lookup table entries for 0 length streams
[wimlib] / src / write.c
index bf4f2396e0c28b74d387eaa672eeabad8b0a9930..8b09449036b2ac52ad016278fe6f0f4180a91700 100644 (file)
 
 #include <limits.h>
 
-#if defined(__WIN32__) && !defined(INVALID_HANDLE_VALUE)
-#  define INVALID_HANDLE_VALUE ((HANDLE)(-1))
-#endif
-
 /* 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.) */
@@ -97,13 +93,13 @@ begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
        u64 num_chunks = (size + WIM_CHUNK_SIZE - 1) / WIM_CHUNK_SIZE;
        size_t alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
        struct chunk_table *chunk_tab = CALLOC(1, alloc_size);
-       int ret;
+
+       DEBUG("Begin chunk table for stream with size %"PRIu64, size);
 
        if (!chunk_tab) {
                ERROR("Failed to allocate chunk table for %"PRIu64" byte "
                      "resource", size);
-               ret = WIMLIB_ERR_NOMEM;
-               goto out;
+               return WIMLIB_ERR_NOMEM;
        }
        chunk_tab->file_offset = file_offset;
        chunk_tab->num_chunks = num_chunks;
@@ -119,14 +115,10 @@ begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
                ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
                                 "file resource");
                FREE(chunk_tab);
-               ret = WIMLIB_ERR_WRITE;
-               goto out;
+               return WIMLIB_ERR_WRITE;
        }
-
-       ret = 0;
        *chunk_tab_ret = chunk_tab;
-out:
-       return ret;
+       return 0;
 }
 
 /*
@@ -155,7 +147,7 @@ out:
 typedef unsigned (*compress_func_t)(const void *chunk, unsigned chunk_size,
                                    void *out);
 
-compress_func_t
+static compress_func_t
 get_compress_func(int out_ctype)
 {
        if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX)
@@ -178,16 +170,18 @@ get_compress_func(int out_ctype)
  * Returns 0 on success; nonzero on failure.
  */
 static int
-write_wim_resource_chunk(const void *chunk, unsigned chunk_size,
-                        FILE *out_fp, compress_func_t compress,
-                        struct chunk_table *chunk_tab)
+write_wim_resource_chunk(const void * restrict chunk,
+                        unsigned chunk_size,
+                        FILE * restrict out_fp,
+                        compress_func_t compress,
+                        struct chunk_table * restrict chunk_tab)
 {
-       const u8 *out_chunk;
+       const void *out_chunk;
        unsigned out_chunk_size;
        if (compress) {
-               u8 *compressed_chunk = alloca(chunk_size);
+               void *compressed_chunk = alloca(chunk_size);
 
-               out_chunk_size = compress(chunk, chunk_size, compressed_chunk);
+               out_chunk_size = (*compress)(chunk, chunk_size, compressed_chunk);
                if (out_chunk_size) {
                        /* Write compressed */
                        out_chunk = compressed_chunk;
@@ -218,8 +212,9 @@ write_wim_resource_chunk(const void *chunk, unsigned chunk_size,
  * @compressed_size_p.
  */
 static int
-finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
-                             FILE *out_fp, u64 *compressed_size_p)
+finish_wim_resource_chunk_tab(struct chunk_table * restrict chunk_tab,
+                             FILE * restrict out_fp,
+                             u64 * restrict compressed_size_p)
 {
        size_t bytes_written;
        if (fseeko(out_fp, chunk_tab->file_offset, SEEK_SET) != 0) {
@@ -252,7 +247,23 @@ finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
 }
 
 static int
-finalize_and_check_sha1(SHA_CTX *sha_ctx, struct wim_lookup_table_entry *lte)
+fflush_and_ftruncate(FILE *out_fp, off_t offset)
+{
+       if (fseeko(out_fp, offset, SEEK_SET) ||
+           fflush(out_fp) ||
+           ftruncate(fileno(out_fp), offset))
+       {
+               ERROR_WITH_ERRNO("Failed to flush and/or truncate "
+                                "output WIM file");
+               return WIMLIB_ERR_WRITE;
+       } else {
+               return 0;
+       }
+}
+
+static int
+finalize_and_check_sha1(SHA_CTX * restrict sha_ctx,
+                       struct wim_lookup_table_entry * restrict lte)
 {
        u8 md[SHA1_HASH_SIZE];
        sha1_final(md, sha_ctx);
@@ -280,25 +291,16 @@ struct write_resource_ctx {
 };
 
 static int
-write_resource_cb(const void *chunk, size_t chunk_size, void *_ctx)
+write_resource_cb(const void *restrict chunk, size_t chunk_size,
+                 void *restrict _ctx)
 {
        struct write_resource_ctx *ctx = _ctx;
 
        if (ctx->doing_sha)
                sha1_update(&ctx->sha_ctx, chunk, chunk_size);
-
-       if (ctx->compress) {
-               return write_wim_resource_chunk(chunk, chunk_size,
-                                               ctx->out_fp, ctx->compress,
-                                               ctx->chunk_tab);
-       } else {
-               if (fwrite(chunk, 1, chunk_size, ctx->out_fp) != chunk_size) {
-                       ERROR_WITH_ERRNO("Error writing to output WIM");
-                       return WIMLIB_ERR_WRITE;
-               } else {
-                       return 0;
-               }
-       }
+       return write_wim_resource_chunk(chunk, chunk_size,
+                                       ctx->out_fp, ctx->compress,
+                                       ctx->chunk_tab);
 }
 
 /*
@@ -337,8 +339,6 @@ write_wim_resource(struct wim_lookup_table_entry *lte,
        off_t offset;
        int ret;
 
-       DEBUG2("wim_resource_size(lte)=%"PRIu64, wim_resource_size(lte));
-
        flags &= ~WIMLIB_RESOURCE_FLAG_RECOMPRESS;
 
        /* Get current position in output WIM */
@@ -422,19 +422,12 @@ try_write_again:
                if (new_size >= wim_resource_size(lte)) {
                        /* Oops!  We compressed the resource to larger than the original
                         * size.  Write the resource uncompressed instead. */
-                       if (fseeko(out_fp, offset, SEEK_SET) ||
-                           fflush(out_fp) ||
-                           ftruncate(fileno(out_fp),
-                                     offset + wim_resource_size(lte)))
-                       {
-                               ERROR_WITH_ERRNO("Failed to flush and/or truncate "
-                                                "output WIM file");
-                               ret = WIMLIB_ERR_WRITE;
-                               goto out_free_chunk_tab;
-                       }
                        DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
                              "writing uncompressed instead",
                              wim_resource_size(lte), new_size);
+                       ret = fflush_and_ftruncate(out_fp, offset);
+                       if (ret)
+                               goto out_free_chunk_tab;
                        write_ctx.compress = NULL;
                        write_ctx.doing_sha = false;
                        out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
@@ -563,7 +556,6 @@ static void
 compress_chunks(struct message *msg, compress_func_t compress)
 {
        for (unsigned i = 0; i < msg->num_chunks; i++) {
-               DEBUG2("compress chunk %u of %u", i, msg->num_chunks);
                unsigned len = compress(msg->uncompressed_chunks[i],
                                        msg->uncompressed_chunk_sizes[i],
                                        msg->compressed_chunks[i]);
@@ -642,6 +634,10 @@ 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
+ * passed the context @write_stream_ctx. */
 static int
 do_write_stream_list(struct list_head *stream_list,
                     struct wim_lookup_table *lookup_table,
@@ -650,7 +646,7 @@ do_write_stream_list(struct list_head *stream_list,
                     wimlib_progress_func_t progress_func,
                     union wimlib_progress_info *progress)
 {
-       int ret;
+       int ret = 0;
        struct wim_lookup_table_entry *lte;
 
        /* For each stream in @stream_list ... */
@@ -666,11 +662,9 @@ do_write_stream_list(struct list_head *stream_list,
                        struct wim_lookup_table_entry *tmp;
                        u32 orig_refcnt = lte->out_refcnt;
 
-                       ret = hash_unhashed_stream(lte,
-                                                  lookup_table,
-                                                  &tmp);
+                       ret = hash_unhashed_stream(lte, lookup_table, &tmp);
                        if (ret)
-                               goto out;
+                               break;
                        if (tmp != lte) {
                                lte = tmp;
                                /* We found a duplicate stream. */
@@ -680,6 +674,7 @@ do_write_stream_list(struct list_head *stream_list,
                                         * just skip to the next stream. */
                                        DEBUG("Discarding duplicate stream of length %"PRIu64,
                                              wim_resource_size(lte));
+                                       lte->no_progress = 0;
                                        goto skip_to_progress;
                                }
                        }
@@ -693,23 +688,27 @@ do_write_stream_list(struct list_head *stream_list,
                 * while in the latter case this is done because we do not have
                 * the SHA1 message digest yet.  */
                wimlib_assert(lte->out_refcnt != 0);
+               lte->deferred = 0;
+               lte->no_progress = 0;
                ret = (*write_stream_cb)(lte, write_stream_ctx);
                if (ret)
-                       goto out;
+                       break;
+               /* In parallel mode, some streams are deferred for later,
+                * serialized processing; ignore them here. */
+               if (lte->deferred)
+                       continue;
                if (lte->unhashed) {
                        list_del(&lte->unhashed_list);
                        lookup_table_insert(lookup_table, lte);
                        lte->unhashed = 0;
                }
        skip_to_progress:
-               if (progress_func) {
+               if (!lte->no_progress) {
                        do_write_streams_progress(progress,
                                                  progress_func,
                                                  wim_resource_size(lte));
                }
        }
-       ret = 0;
-out:
        return ret;
 }
 
@@ -738,8 +737,11 @@ do_write_stream_list_serial(struct list_head *stream_list,
 static inline int
 write_flags_to_resource_flags(int write_flags)
 {
-       return (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS) ?
-                       WIMLIB_RESOURCE_FLAG_RECOMPRESS : 0;
+       int resource_flags = 0;
+
+       if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
+               resource_flags |= WIMLIB_RESOURCE_FLAG_RECOMPRESS;
+       return resource_flags;
 }
 
 static int
@@ -751,6 +753,7 @@ write_stream_list_serial(struct list_head *stream_list,
                         wimlib_progress_func_t progress_func,
                         union wimlib_progress_info *progress)
 {
+       DEBUG("Writing stream list (serial version)");
        progress->write_streams.num_threads = 1;
        if (progress_func)
                progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
@@ -771,9 +774,6 @@ write_wim_chunks(struct message *msg, FILE *out_fp,
        for (unsigned i = 0; i < msg->num_chunks; i++) {
                unsigned chunk_csize = msg->compressed_chunk_sizes[i];
 
-               DEBUG2("Write wim chunk %u of %u (csize = %u)",
-                     i, msg->num_chunks, chunk_csize);
-
                if (fwrite(msg->out_compressed_chunks[i], 1, chunk_csize, out_fp)
                    != chunk_csize)
                {
@@ -802,6 +802,7 @@ struct main_writer_thread_ctx {
        struct list_head available_msgs;
        struct list_head outstanding_streams;
        struct list_head serial_streams;
+       size_t num_outstanding_messages;
 
        SHA_CTX next_sha_ctx;
        u64 next_chunk;
@@ -865,23 +866,12 @@ allocate_messages(size_t num_messages)
 static void
 main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
 {
-       size_t num_available_msgs;
-       size_t num_outstanding_msgs;
-       struct list_head *cur;
-
-       num_available_msgs = 0;
-       list_for_each(cur, &ctx->available_msgs)
-               num_available_msgs++;
-       
-       num_outstanding_msgs = ctx->num_messages - num_available_msgs;
-       while (num_outstanding_msgs--)
+       while (ctx->num_outstanding_messages--)
                shared_queue_get(ctx->compressed_res_queue);
-
        free_messages(ctx->msgs, ctx->num_messages);
        FREE(ctx->cur_chunk_tab);
 }
 
-
 static int
 main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
 {
@@ -906,11 +896,16 @@ main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
         * currently being read and having chunks fed to the compressor threads.
         * */
        INIT_LIST_HEAD(&ctx->outstanding_streams);
+       ctx->num_outstanding_messages = 0;
+
+       ctx->next_msg = NULL;
 
        /* Resources that don't need any chunks compressed are added to this
         * list and written directly by the main thread. */
        INIT_LIST_HEAD(&ctx->serial_streams);
 
+       ctx->cur_chunk_tab = NULL;
+
        return 0;
 }
 
@@ -922,7 +917,8 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
        int ret;
 
        wimlib_assert(!list_empty(&ctx->outstanding_streams));
-       DEBUG2("Receiving more compressed chunks");
+       wimlib_assert(ctx->num_outstanding_messages != 0);
+
        cur_lte = container_of(ctx->outstanding_streams.next,
                               struct wim_lookup_table_entry,
                               being_compressed_list);
@@ -932,8 +928,7 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
         * compressed. */
        msg = shared_queue_get(ctx->compressed_res_queue);
        msg->complete = true;
-
-       DEBUG2("recved msg %p", msg);
+       --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
@@ -941,7 +936,8 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
         * message around until all earlier chunks are received.
         *
         * Otherwise, write all the chunks we can. */
-       while (!list_empty(&cur_lte->msg_list) 
+       while (cur_lte != NULL &&
+              !list_empty(&cur_lte->msg_list)
               && (msg = container_of(cur_lte->msg_list.next,
                                      struct message,
                                      list))->complete)
@@ -971,8 +967,9 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                if (list_empty(&cur_lte->msg_list) &&
                    msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
                {
-                       DEBUG2("Finish wim chunk tab");
                        u64 res_csize;
+                       off_t offset;
+
                        ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
                                                            ctx->out_fp,
                                                            &res_csize);
@@ -980,21 +977,32 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                                return ret;
 
                        list_del(&cur_lte->being_compressed_list);
-#if 0
+
+                       /* 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;
+
+                       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. */
-                               ret = write_uncompressed_resource_and_truncate(
-                                                cur_lte,
-                                                ctx->out_fp,
-                                                ctx->cur_chunk_tab->file_offset,
-                                                &cur_lte->output_resource_entry);
+                               DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
+                                     "writing uncompressed instead",
+                                     wim_resource_size(cur_lte), res_csize);
+                               ret = fflush_and_ftruncate(ctx->out_fp, offset);
                                if (ret)
-                                       goto out;
-                       } else
-#endif
-                       {
+                                       return ret;
+                               ret = write_wim_resource(cur_lte,
+                                                        ctx->out_fp,
+                                                        WIMLIB_COMPRESSION_TYPE_NONE,
+                                                        &cur_lte->output_resource_entry,
+                                                        ctx->write_resource_flags);
+                               if (ret)
+                                       return ret;
+                       } else {
                                cur_lte->output_resource_entry.size =
                                        res_csize;
 
@@ -1002,16 +1010,16 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                                        cur_lte->resource_entry.original_size;
 
                                cur_lte->output_resource_entry.offset =
-                                       ctx->cur_chunk_tab->file_offset;
+                                       offset;
 
                                cur_lte->output_resource_entry.flags =
                                        cur_lte->resource_entry.flags |
                                                WIM_RESHDR_FLAG_COMPRESSED;
                        }
-                       do_write_streams_progress(ctx->progress, ctx->progress_func,
+
+                       do_write_streams_progress(ctx->progress,
+                                                 ctx->progress_func,
                                                  wim_resource_size(cur_lte));
-                       FREE(ctx->cur_chunk_tab);
-                       ctx->cur_chunk_tab = NULL;
 
                        /* Since we just finished writing a stream, write any
                         * streams that have been added to the serial_streams
@@ -1019,52 +1027,63 @@ receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
                         * resources that don't need to be compressed because
                         * the desired compression type is the same as the
                         * previous compression type). */
-#if 0
-                       ret = do_write_stream_list_serial(&ctx->serial_streams,
-                                                         ctx->lookup_table,
-                                                         ctx->out_fp,
-                                                         ctx->out_ctype,
-                                                         ctx->write_resource_flags,
-                                                         ctx->progress_func,
-                                                         ctx->progress);
-#endif
-                       if (ret)
-                               return ret;
-                       if (list_empty(&ctx->outstanding_streams))
-                               return 0;
-                       cur_lte = container_of(ctx->outstanding_streams.next,
-                                              struct wim_lookup_table_entry,
-                                              being_compressed_list);
-               #ifdef ENABLE_MORE_DEBUG
-                       DEBUG2("Advance to stream:");
-                       print_lookup_table_entry(cur_lte, stderr);
-               #endif
+                       if (!list_empty(&ctx->serial_streams)) {
+                               ret = do_write_stream_list_serial(&ctx->serial_streams,
+                                                                 ctx->lookup_table,
+                                                                 ctx->out_fp,
+                                                                 ctx->out_ctype,
+                                                                 ctx->write_resource_flags,
+                                                                 ctx->progress_func,
+                                                                 ctx->progress);
+                               if (ret)
+                                       return ret;
+                       }
+
+                       /* Advance to the next stream to write. */
+                       if (list_empty(&ctx->outstanding_streams)) {
+                               cur_lte = NULL;
+                       } else {
+                               cur_lte = container_of(ctx->outstanding_streams.next,
+                                                      struct wim_lookup_table_entry,
+                                                      being_compressed_list);
+                       }
                }
        }
        return 0;
 }
 
+/* Called when the main thread has read a new chunk of data. */
 static int
 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
 {
        struct main_writer_thread_ctx *ctx = _ctx;
        int ret;
        struct message *next_msg;
+       u64 next_chunk_in_msg;
 
-       DEBUG2("chunk_size=%zu, wim_resource_size(next_lte)=%"PRIu64,
-              chunk_size, wim_resource_size(ctx->next_lte));
-
+       /* Update SHA1 message digest for the stream currently being read by the
+        * main thread. */
        sha1_update(&ctx->next_sha_ctx, chunk, chunk_size);
+
+       /* We send chunks of data to the compressor chunks in batches which we
+        * refer to as "messages".  @next_msg is the message that is currently
+        * being prepared to send off.  If it is NULL, that indicates that we
+        * need to start a new message. */
        next_msg = ctx->next_msg;
        if (!next_msg) {
-               /* Start filling in a new message */
-
-               DEBUG2("Start new msg");
-
+               /* We need to start a new message.  First check to see if there
+                * is a message available in the list of available messages.  If
+                * so, we can just take one.  If not, all the messages (there is
+                * a fixed number of them, proportional to the number of
+                * threads) have been sent off to the compressor threads, so we
+                * receive messages from the compressor threads containing
+                * compressed chunks of data.
+                *
+                * We may need to receive multiple messages before one is
+                * actually available to use because messages received that are
+                * *not* for the very next set of chunks to compress must be
+                * buffered until it's time to write those chunks. */
                while (list_empty(&ctx->available_msgs)) {
-                       /* No message available; receive messages, writing
-                        * compressed data. */
-                       DEBUG2("No msgs available!");
                        ret = receive_compressed_chunks(ctx);
                        if (ret)
                                return ret;
@@ -1077,23 +1096,21 @@ main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
                next_msg->begin_chunk = ctx->next_chunk;
                next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
                                           ctx->next_num_chunks - ctx->next_chunk);
-               DEBUG2("next_msg {begin_chunk=%"PRIu64", num_chunks=%"PRIu64"}",
-                      next_msg->begin_chunk, next_msg->num_chunks);
                ctx->next_msg = next_msg;
        }
 
-       u64 next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;
-
        /* Fill in the next chunk to compress */
+       next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;
+
        next_msg->uncompressed_chunk_sizes[next_chunk_in_msg] = chunk_size;
        memcpy(next_msg->uncompressed_chunks[next_chunk_in_msg],
               chunk, chunk_size);
        ctx->next_chunk++;
        if (++next_chunk_in_msg == next_msg->num_chunks) {
-               DEBUG2("Sending message %p", next_msg);
                /* Send off an array of chunks to compress */
                list_add_tail(&next_msg->list, &ctx->next_lte->msg_list);
                shared_queue_put(ctx->res_to_compress_queue, next_msg);
+               ++ctx->num_outstanding_messages;
                ctx->next_msg = NULL;
        }
        return 0;
@@ -1103,13 +1120,13 @@ static int
 main_writer_thread_finish(void *_ctx)
 {
        struct main_writer_thread_ctx *ctx = _ctx;
-       int ret = 0;
-       DEBUG2("finishing");
-       while (!list_empty(&ctx->outstanding_streams)) {
+       int ret;
+       while (ctx->num_outstanding_messages != 0) {
                ret = receive_compressed_chunks(ctx);
                if (ret)
-                       break;
+                       return ret;
        }
+       wimlib_assert(list_empty(&ctx->outstanding_streams));
        return do_write_stream_list_serial(&ctx->serial_streams,
                                           ctx->lookup_table,
                                           ctx->out_fp,
@@ -1125,11 +1142,10 @@ submit_stream_for_compression(struct wim_lookup_table_entry *lte,
 {
        int ret;
 
-#ifdef ENABLE_MORE_DEBUG
-       DEBUG2("Submit for compression:");
-       print_lookup_table_entry(lte, stderr);
-#endif
-
+       /* Read the entire stream @lte, feeding its data chunks to the
+        * compressor threads.  Also SHA1-sum the stream; this is required in
+        * the case that @lte is unhashed, and a nice additional verification
+        * when @lte is already hashed. */
        sha1_init(&ctx->next_sha_ctx);
        ctx->next_chunk = 0;
        ctx->next_num_chunks = wim_resource_chunks(lte);
@@ -1154,14 +1170,20 @@ 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))
        {
-               list_add_tail(&lte->write_streams_list,
-                             &ctx->serial_streams);
+               /* Stream is too small or isn't being compressed.  Process it by
+                * the main thread when we have a chance.  We can't necessarily
+                * process it right here, as the main thread could be in the
+                * middle of writing a different stream. */
+               list_add_tail(&lte->write_streams_list, &ctx->serial_streams);
+               lte->deferred = 1;
                ret = 0;
        } else {
                ret = submit_stream_for_compression(lte, ctx);
        }
+       lte->no_progress = 1;
        return ret;
 }
 
@@ -1175,15 +1197,41 @@ get_default_num_threads()
 #endif
 }
 
+/* Equivalent to write_stream_list_serial(), except this takes a @num_threads
+ * parameter and will perform compression using that many threads.  Falls
+ * back to write_stream_list_serial() on certain errors, such as a failure to
+ * create the number of threads requested.
+ *
+ * High level description of the algorithm for writing compressed streams in
+ * parallel:  We perform compression on chunks of size WIM_CHUNK_SIZE bytes
+ * rather than on full files.  The currently executing thread becomes the main
+ * thread and is entirely in charge of reading the data to compress (which may
+ * be in any location understood by the resource code--- such as in an external
+ * file being captured, or in another WIM file from which an image is being
+ * exported) and actually writing the compressed data to the output file.
+ * Additional threads are "compressor threads" and all execute the
+ * compressor_thread_proc, where they repeatedly retrieve buffers of data from
+ * the main thread, compress them, and hand them back to the main thread.
+ *
+ * Certain streams, such as streams that do not need to be compressed (e.g.
+ * input compression type same as output compression type) or streams of very
+ * small size are placed in a list (main_writer_thread_ctx.serial_list) and
+ * handled entirely by the main thread at an appropriate time.
+ *
+ * At any given point in time, multiple streams may be having chunks compressed
+ * concurrently.  The stream that the main thread is currently *reading* may be
+ * later in the list that the stream that the main thread is currently
+ * *writing*.
+ */
 static int
 write_stream_list_parallel(struct list_head *stream_list,
                           struct wim_lookup_table *lookup_table,
                           FILE *out_fp,
                           int out_ctype,
                           int write_resource_flags,
-                          unsigned num_threads,
                           wimlib_progress_func_t progress_func,
-                          union wimlib_progress_info *progress)
+                          union wimlib_progress_info *progress,
+                          unsigned num_threads)
 {
        int ret;
        struct shared_queue res_to_compress_queue;
@@ -1195,11 +1243,16 @@ write_stream_list_parallel(struct list_head *stream_list,
                if (nthreads < 1 || nthreads > UINT_MAX) {
                        WARNING("Could not determine number of processors! Assuming 1");
                        goto out_serial;
+               } else if (nthreads == 1) {
+                       goto out_serial_quiet;
                } else {
                        num_threads = nthreads;
                }
        }
 
+       DEBUG("Writing stream list (parallel version, num_threads=%u)",
+             num_threads);
+
        progress->write_streams.num_threads = num_threads;
 
        static const size_t MESSAGES_PER_THREAD = 2;
@@ -1244,7 +1297,6 @@ write_stream_list_parallel(struct list_head *stream_list,
                progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
 
        struct main_writer_thread_ctx ctx;
-       memset(&ctx, 0, sizeof(ctx));
        ctx.stream_list           = stream_list;
        ctx.lookup_table          = lookup_table;
        ctx.out_fp                = out_fp;
@@ -1252,20 +1304,24 @@ write_stream_list_parallel(struct list_head *stream_list,
        ctx.res_to_compress_queue = &res_to_compress_queue;
        ctx.compressed_res_queue  = &compressed_res_queue;
        ctx.num_messages          = queue_size;
-       ctx.write_resource_flags  = write_resource_flags;
+       ctx.write_resource_flags  = write_resource_flags | WIMLIB_RESOURCE_FLAG_THREADSAFE_READ;
        ctx.progress_func         = progress_func;
        ctx.progress              = progress;
        ret = main_writer_thread_init_ctx(&ctx);
        if (ret)
-               goto out;
-       ret = do_write_stream_list(stream_list,
-                                  lookup_table,
+               goto out_join;
+       ret = do_write_stream_list(stream_list, lookup_table,
                                   main_thread_process_next_stream,
-                                  &ctx,
-                                  NULL,
-                                  NULL);
+                                  &ctx, progress_func, progress);
        if (ret)
                goto out_destroy_ctx;
+
+       /* The main thread has finished reading all streams that are going to be
+        * compressed in parallel, and it now needs to wait for all remaining
+        * chunks to be compressed so that the remaining streams can actually be
+        * written to the output file.  Furthermore, any remaining streams that
+        * had processing deferred to the main thread need to be handled.  These
+        * tasks are done by the main_writer_thread_finish() function. */
        ret = main_writer_thread_finish(&ctx);
 out_destroy_ctx:
        main_writer_thread_destroy_ctx(&ctx);
@@ -1285,11 +1341,11 @@ out_destroy_compressed_res_queue:
        shared_queue_destroy(&compressed_res_queue);
 out_destroy_res_to_compress_queue:
        shared_queue_destroy(&res_to_compress_queue);
-out:
        if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
                return ret;
 out_serial:
        WARNING("Falling back to single-threaded compression");
+out_serial_quiet:
        return write_stream_list_serial(stream_list,
                                        lookup_table,
                                        out_fp,
@@ -1308,7 +1364,7 @@ out_serial:
 static int
 write_stream_list(struct list_head *stream_list,
                  struct wim_lookup_table *lookup_table,
-                 FILE *out_fp, int out_ctype, int write_resource_flags,
+                 FILE *out_fp, int out_ctype, int write_flags,
                  unsigned num_threads, wimlib_progress_func_t progress_func)
 {
        struct wim_lookup_table_entry *lte;
@@ -1317,10 +1373,13 @@ write_stream_list(struct list_head *stream_list,
        u64 total_compression_bytes = 0;
        union wimlib_progress_info progress;
        int ret;
+       int write_resource_flags;
 
        if (list_empty(stream_list))
                return 0;
 
+       write_resource_flags = write_flags_to_resource_flags(write_flags);
+
        /* 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
@@ -1350,9 +1409,9 @@ write_stream_list(struct list_head *stream_list,
                                                 out_fp,
                                                 out_ctype,
                                                 write_resource_flags,
-                                                num_threads,
                                                 progress_func,
-                                                &progress);
+                                                &progress,
+                                                num_threads);
        else
 #endif
                ret = write_stream_list_serial(stream_list,
@@ -1619,7 +1678,7 @@ write_wim_streams(WIMStruct *wim, int image, int write_flags,
                                 wim->lookup_table,
                                 wim->out_fp,
                                 wimlib_get_compression_type(wim),
-                                write_flags_to_resource_flags(write_flags),
+                                write_flags,
                                 num_threads,
                                 progress_func);
 }