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;
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;
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;
}
/*
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)
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 */
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]);
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,
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 ... */
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. */
* 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;
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(<e->unhashed_list);
lookup_table_insert(lookup_table, lte);
wim_resource_size(lte));
}
}
- ret = 0;
-out:
return ret;
}
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)
{
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;
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)
{
* 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;
}
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);
* 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
* 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)
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;
ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
ctx->out_fp,
* 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->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
+
+ /* 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;
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;
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,
{
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);
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(<e->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(<e->write_streams_list, &ctx->serial_streams);
+ lte->deferred = 1;
ret = 0;
} else {
ret = submit_stream_for_compression(lte, ctx);
#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;
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;
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, NULL, NULL);
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);
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:
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;
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
out_fp,
out_ctype,
write_resource_flags,
- num_threads,
progress_func,
- &progress);
+ &progress,
+ num_threads);
else
#endif
ret = write_stream_list_serial(stream_list,
wim->lookup_table,
wim->out_fp,
wimlib_get_compression_type(wim),
- write_flags_to_resource_flags(write_flags),
+ write_flags,
num_threads,
progress_func);
}
git://wimlib.net / wimlib / blobdiff