4 * Support for writing WIM files; write a WIM file, overwrite a WIM file, write
5 * compressed file resources, etc.
9 * Copyright (C) 2012, 2013 Eric Biggers
11 * This file is part of wimlib, a library for working with WIM files.
13 * wimlib is free software; you can redistribute it and/or modify it under the
14 * terms of the GNU General Public License as published by the Free
15 * Software Foundation; either version 3 of the License, or (at your option)
18 * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
19 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
20 * A PARTICULAR PURPOSE. See the GNU General Public License for more
23 * You should have received a copy of the GNU General Public License
24 * along with wimlib; if not, see http://www.gnu.org/licenses/.
31 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
32 /* On BSD, this should be included before "wimlib/list.h" so that "wimlib/list.h" can
33 * overwrite the LIST_HEAD macro. */
34 # include <sys/file.h>
37 #include "wimlib/endianness.h"
38 #include "wimlib/error.h"
39 #include "wimlib/file_io.h"
40 #include "wimlib/header.h"
41 #include "wimlib/integrity.h"
42 #include "wimlib/lookup_table.h"
43 #include "wimlib/metadata.h"
44 #include "wimlib/resource.h"
45 #include "wimlib/write.h"
46 #include "wimlib/xml.h"
49 # include "wimlib/win32.h" /* win32_get_number_of_processors() */
52 #ifdef ENABLE_MULTITHREADED_COMPRESSION
62 # include <ntfs-3g/attrib.h>
63 # include <ntfs-3g/inode.h>
64 # include <ntfs-3g/dir.h>
76 # include <sys/uio.h> /* for `struct iovec' */
79 /* Chunk table that's located at the beginning of each compressed resource in
80 * the WIM. (This is not the on-disk format; the on-disk format just has an
81 * array of offsets.) */
85 u64 original_resource_size;
86 u64 bytes_per_chunk_entry;
94 * Allocates and initializes a chunk table, and reserves space for it in the
98 begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
101 struct chunk_table **chunk_tab_ret)
103 u64 size = wim_resource_size(lte);
104 u64 num_chunks = (size + WIM_CHUNK_SIZE - 1) / WIM_CHUNK_SIZE;
105 size_t alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
106 struct chunk_table *chunk_tab = CALLOC(1, alloc_size);
108 DEBUG("Begin chunk table for stream with size %"PRIu64, size);
111 ERROR("Failed to allocate chunk table for %"PRIu64" byte "
113 return WIMLIB_ERR_NOMEM;
115 chunk_tab->file_offset = file_offset;
116 chunk_tab->num_chunks = num_chunks;
117 chunk_tab->original_resource_size = size;
118 chunk_tab->bytes_per_chunk_entry = (size >= (1ULL << 32)) ? 8 : 4;
119 chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
121 chunk_tab->cur_offset = 0;
122 chunk_tab->cur_offset_p = chunk_tab->offsets;
124 if (full_write(out_fd, chunk_tab,
125 chunk_tab->table_disk_size) != chunk_tab->table_disk_size)
127 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
130 return WIMLIB_ERR_WRITE;
132 *chunk_tab_ret = chunk_tab;
137 * compress_func_t- Pointer to a function to compresses a chunk
138 * of a WIM resource. This may be either
139 * wimlib_xpress_compress() (xpress-compress.c) or
140 * wimlib_lzx_compress() (lzx-compress.c).
142 * @chunk: Uncompressed data of the chunk.
143 * @chunk_size: Size of the uncompressed chunk, in bytes.
144 * @out: Pointer to output buffer of size at least (@chunk_size - 1) bytes.
146 * Returns the size of the compressed data written to @out in bytes, or 0 if the
147 * data could not be compressed to (@chunk_size - 1) bytes or fewer.
149 * As a special requirement, the compression code is optimized for the WIM
150 * format and therefore requires (@chunk_size <= 32768).
152 * As another special requirement, the compression code will read up to 8 bytes
153 * off the end of the @chunk array for performance reasons. The values of these
154 * bytes will not affect the output of the compression, but the calling code
155 * must make sure that the buffer holding the uncompressed chunk is actually at
156 * least (@chunk_size + 8) bytes, or at least that these extra bytes are in
157 * mapped memory that will not cause a memory access violation if accessed.
159 typedef unsigned (*compress_func_t)(const void *chunk, unsigned chunk_size,
162 static compress_func_t
163 get_compress_func(int out_ctype)
165 if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX)
166 return wimlib_lzx_compress;
168 return wimlib_xpress_compress;
172 * Writes a chunk of a WIM resource to an output file.
174 * @chunk: Uncompressed data of the chunk.
175 * @chunk_size: Size of the chunk (<= WIM_CHUNK_SIZE)
176 * @out_fd: File descriptor to write the chunk to.
177 * @compress: Compression function to use (NULL if writing uncompressed
179 * @chunk_tab: Pointer to chunk table being created. It is updated with the
180 * offset of the chunk we write.
182 * Returns 0 on success; nonzero on failure.
185 write_wim_resource_chunk(const void * restrict chunk,
188 compress_func_t compress,
189 struct chunk_table * restrict chunk_tab)
191 const void *out_chunk;
192 unsigned out_chunk_size;
194 void *compressed_chunk = alloca(chunk_size);
196 out_chunk_size = (*compress)(chunk, chunk_size, compressed_chunk);
197 if (out_chunk_size) {
198 /* Write compressed */
199 out_chunk = compressed_chunk;
201 /* Write uncompressed */
203 out_chunk_size = chunk_size;
205 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
206 chunk_tab->cur_offset += out_chunk_size;
208 /* Write uncompressed */
210 out_chunk_size = chunk_size;
212 if (full_write(out_fd, out_chunk, out_chunk_size) != out_chunk_size) {
213 ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
214 return WIMLIB_ERR_WRITE;
220 * Finishes a WIM chunk table and writes it to the output file at the correct
223 * The final size of the full compressed resource is returned in the
224 * @compressed_size_p.
227 finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
228 int out_fd, u64 *compressed_size_p)
230 size_t bytes_written;
232 if (chunk_tab->bytes_per_chunk_entry == 8) {
233 array_cpu_to_le64(chunk_tab->offsets, chunk_tab->num_chunks);
235 for (u64 i = 0; i < chunk_tab->num_chunks; i++)
236 ((u32*)chunk_tab->offsets)[i] =
237 cpu_to_le32(chunk_tab->offsets[i]);
239 bytes_written = full_pwrite(out_fd,
240 (u8*)chunk_tab->offsets + chunk_tab->bytes_per_chunk_entry,
241 chunk_tab->table_disk_size,
242 chunk_tab->file_offset);
243 if (bytes_written != chunk_tab->table_disk_size) {
244 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
246 return WIMLIB_ERR_WRITE;
248 *compressed_size_p = chunk_tab->cur_offset + chunk_tab->table_disk_size;
253 seek_and_truncate(int out_fd, off_t offset)
255 if (lseek(out_fd, offset, SEEK_SET) == -1 ||
256 ftruncate(out_fd, offset))
258 ERROR_WITH_ERRNO("Failed to truncate output WIM file");
259 return WIMLIB_ERR_WRITE;
266 finalize_and_check_sha1(SHA_CTX * restrict sha_ctx,
267 struct wim_lookup_table_entry * restrict lte)
269 u8 md[SHA1_HASH_SIZE];
270 sha1_final(md, sha_ctx);
272 copy_hash(lte->hash, md);
273 } else if (!hashes_equal(md, lte->hash)) {
274 ERROR("WIM resource has incorrect hash!");
275 if (lte_filename_valid(lte)) {
276 ERROR("We were reading it from \"%"TS"\"; maybe "
277 "it changed while we were reading it.",
280 return WIMLIB_ERR_INVALID_RESOURCE_HASH;
286 struct write_resource_ctx {
287 compress_func_t compress;
288 struct chunk_table *chunk_tab;
295 write_resource_cb(const void *restrict chunk, size_t chunk_size,
298 struct write_resource_ctx *ctx = _ctx;
301 sha1_update(&ctx->sha_ctx, chunk, chunk_size);
302 return write_wim_resource_chunk(chunk, chunk_size,
303 ctx->out_fd, ctx->compress,
308 * Write a resource to an output WIM.
310 * @lte: Lookup table entry for the resource, which could be in another WIM,
311 * in an external file, or in another location.
313 * @out_fd: File descriptor opened to the output WIM.
315 * @out_ctype: One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate
316 * which compression algorithm to use.
318 * @out_res_entry: On success, this is filled in with the offset, flags,
319 * compressed size, and uncompressed size of the resource
322 * @flags: WIMLIB_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed
323 * even if it could otherwise be copied directly from the input.
325 * Additional notes: The SHA1 message digest of the uncompressed data is
326 * calculated (except when doing a raw copy --- see below). If the @unhashed
327 * flag is set on the lookup table entry, this message digest is simply copied
328 * to it; otherwise, the message digest is compared with the existing one, and
329 * the function will fail if they do not match.
332 write_wim_resource(struct wim_lookup_table_entry *lte,
333 int out_fd, int out_ctype,
334 struct resource_entry *out_res_entry,
337 struct write_resource_ctx write_ctx;
343 flags &= ~WIMLIB_RESOURCE_FLAG_RECOMPRESS;
345 /* Get current position in output WIM */
346 offset = filedes_offset(out_fd);
348 ERROR_WITH_ERRNO("Can't get position in output WIM");
349 return WIMLIB_ERR_WRITE;
352 /* If we are not forcing the data to be recompressed, and the input
353 * resource is located in a WIM with the same compression type as that
354 * desired other than no compression, we can simply copy the compressed
355 * data without recompressing it. This also means we must skip
356 * calculating the SHA1, as we never will see the uncompressed data. */
357 if (!(flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
358 lte->resource_location == RESOURCE_IN_WIM &&
359 out_ctype != WIMLIB_COMPRESSION_TYPE_NONE &&
360 wimlib_get_compression_type(lte->wim) == out_ctype)
362 flags |= WIMLIB_RESOURCE_FLAG_RAW;
363 write_ctx.doing_sha = false;
364 read_size = lte->resource_entry.size;
366 write_ctx.doing_sha = true;
367 sha1_init(&write_ctx.sha_ctx);
368 read_size = lte->resource_entry.original_size;
371 /* Initialize the chunk table and set the compression function if
372 * compressing the resource. */
373 if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
374 (flags & WIMLIB_RESOURCE_FLAG_RAW)) {
375 write_ctx.compress = NULL;
376 write_ctx.chunk_tab = NULL;
378 write_ctx.compress = get_compress_func(out_ctype);
379 ret = begin_wim_resource_chunk_tab(lte, out_fd,
381 &write_ctx.chunk_tab);
386 /* Write the entire resource by reading the entire resource and feeding
387 * the data through the write_resource_cb function. */
388 write_ctx.out_fd = out_fd;
390 ret = read_resource_prefix(lte, read_size,
391 write_resource_cb, &write_ctx, flags);
393 goto out_free_chunk_tab;
395 /* Verify SHA1 message digest of the resource, or set the hash for the
397 if (write_ctx.doing_sha) {
398 ret = finalize_and_check_sha1(&write_ctx.sha_ctx, lte);
400 goto out_free_chunk_tab;
403 out_res_entry->flags = lte->resource_entry.flags;
404 out_res_entry->original_size = wim_resource_size(lte);
405 out_res_entry->offset = offset;
406 if (flags & WIMLIB_RESOURCE_FLAG_RAW) {
407 /* Doing a raw write: The new compressed size is the same as
408 * the compressed size in the other WIM. */
409 new_size = lte->resource_entry.size;
410 } else if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE) {
411 /* Using WIMLIB_COMPRESSION_TYPE_NONE: The new compressed size
412 * is the original size. */
413 new_size = lte->resource_entry.original_size;
414 out_res_entry->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
416 /* Using a different compression type: Call
417 * finish_wim_resource_chunk_tab() and it will provide the new
418 * compressed size. */
419 ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab, out_fd,
422 goto out_free_chunk_tab;
423 if (new_size >= wim_resource_size(lte)) {
424 /* Oops! We compressed the resource to larger than the original
425 * size. Write the resource uncompressed instead. */
426 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
427 "writing uncompressed instead",
428 wim_resource_size(lte), new_size);
429 ret = seek_and_truncate(out_fd, offset);
431 goto out_free_chunk_tab;
432 write_ctx.compress = NULL;
433 write_ctx.doing_sha = false;
434 out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
435 goto try_write_again;
437 out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
439 out_res_entry->size = new_size;
442 FREE(write_ctx.chunk_tab);
446 #ifdef ENABLE_MULTITHREADED_COMPRESSION
448 /* Blocking shared queue (solves the producer-consumer problem) */
449 struct shared_queue {
453 unsigned filled_slots;
455 pthread_mutex_t lock;
456 pthread_cond_t msg_avail_cond;
457 pthread_cond_t space_avail_cond;
461 shared_queue_init(struct shared_queue *q, unsigned size)
463 wimlib_assert(size != 0);
464 q->array = CALLOC(sizeof(q->array[0]), size);
471 if (pthread_mutex_init(&q->lock, NULL)) {
472 ERROR_WITH_ERRNO("Failed to initialize mutex");
475 if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
476 ERROR_WITH_ERRNO("Failed to initialize condition variable");
477 goto err_destroy_lock;
479 if (pthread_cond_init(&q->space_avail_cond, NULL)) {
480 ERROR_WITH_ERRNO("Failed to initialize condition variable");
481 goto err_destroy_msg_avail_cond;
484 err_destroy_msg_avail_cond:
485 pthread_cond_destroy(&q->msg_avail_cond);
487 pthread_mutex_destroy(&q->lock);
489 return WIMLIB_ERR_NOMEM;
493 shared_queue_destroy(struct shared_queue *q)
496 pthread_mutex_destroy(&q->lock);
497 pthread_cond_destroy(&q->msg_avail_cond);
498 pthread_cond_destroy(&q->space_avail_cond);
502 shared_queue_put(struct shared_queue *q, void *obj)
504 pthread_mutex_lock(&q->lock);
505 while (q->filled_slots == q->size)
506 pthread_cond_wait(&q->space_avail_cond, &q->lock);
508 q->back = (q->back + 1) % q->size;
509 q->array[q->back] = obj;
512 pthread_cond_broadcast(&q->msg_avail_cond);
513 pthread_mutex_unlock(&q->lock);
517 shared_queue_get(struct shared_queue *q)
521 pthread_mutex_lock(&q->lock);
522 while (q->filled_slots == 0)
523 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
525 obj = q->array[q->front];
526 q->array[q->front] = NULL;
527 q->front = (q->front + 1) % q->size;
530 pthread_cond_broadcast(&q->space_avail_cond);
531 pthread_mutex_unlock(&q->lock);
535 struct compressor_thread_params {
536 struct shared_queue *res_to_compress_queue;
537 struct shared_queue *compressed_res_queue;
538 compress_func_t compress;
541 #define MAX_CHUNKS_PER_MSG 2
544 struct wim_lookup_table_entry *lte;
545 u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
546 u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
547 unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
548 struct iovec out_chunks[MAX_CHUNKS_PER_MSG];
549 size_t total_out_bytes;
551 struct list_head list;
557 compress_chunks(struct message *msg, compress_func_t compress)
559 msg->total_out_bytes = 0;
560 for (unsigned i = 0; i < msg->num_chunks; i++) {
561 unsigned len = compress(msg->uncompressed_chunks[i],
562 msg->uncompressed_chunk_sizes[i],
563 msg->compressed_chunks[i]);
567 /* To be written compressed */
568 out_chunk = msg->compressed_chunks[i];
571 /* To be written uncompressed */
572 out_chunk = msg->uncompressed_chunks[i];
573 out_len = msg->uncompressed_chunk_sizes[i];
575 msg->out_chunks[i].iov_base = out_chunk;
576 msg->out_chunks[i].iov_len = out_len;
577 msg->total_out_bytes += out_len;
581 /* Compressor thread routine. This is a lot simpler than the main thread
582 * routine: just repeatedly get a group of chunks from the
583 * res_to_compress_queue, compress them, and put them in the
584 * compressed_res_queue. A NULL pointer indicates that the thread should stop.
587 compressor_thread_proc(void *arg)
589 struct compressor_thread_params *params = arg;
590 struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
591 struct shared_queue *compressed_res_queue = params->compressed_res_queue;
592 compress_func_t compress = params->compress;
595 DEBUG("Compressor thread ready");
596 while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
597 compress_chunks(msg, compress);
598 shared_queue_put(compressed_res_queue, msg);
600 DEBUG("Compressor thread terminating");
603 #endif /* ENABLE_MULTITHREADED_COMPRESSION */
606 do_write_streams_progress(union wimlib_progress_info *progress,
607 wimlib_progress_func_t progress_func,
609 bool stream_discarded)
611 if (stream_discarded) {
612 progress->write_streams.total_bytes -= size_added;
613 if (progress->write_streams._private != ~(uint64_t)0 &&
614 progress->write_streams._private > progress->write_streams.total_bytes)
616 progress->write_streams._private = progress->write_streams.total_bytes;
619 progress->write_streams.completed_bytes += size_added;
621 progress->write_streams.completed_streams++;
623 progress->write_streams.completed_bytes >= progress->write_streams._private)
625 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
627 if (progress->write_streams._private == progress->write_streams.total_bytes) {
628 progress->write_streams._private = ~(uint64_t)0;
630 progress->write_streams._private =
631 min(progress->write_streams.total_bytes,
632 progress->write_streams.completed_bytes +
633 progress->write_streams.total_bytes / 100);
638 struct serial_write_stream_ctx {
641 int write_resource_flags;
645 serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
647 struct serial_write_stream_ctx *ctx = _ctx;
648 return write_wim_resource(lte, ctx->out_fd,
649 ctx->out_ctype, <e->output_resource_entry,
650 ctx->write_resource_flags);
653 /* Write a list of streams, taking into account that some streams may be
654 * duplicates that are checksummed and discarded on the fly, and also delegating
655 * the actual writing of a stream to a function @write_stream_cb, which is
656 * passed the context @write_stream_ctx. */
658 do_write_stream_list(struct list_head *stream_list,
659 struct wim_lookup_table *lookup_table,
660 int (*write_stream_cb)(struct wim_lookup_table_entry *, void *),
661 void *write_stream_ctx,
662 wimlib_progress_func_t progress_func,
663 union wimlib_progress_info *progress)
666 struct wim_lookup_table_entry *lte;
667 bool stream_discarded;
669 /* For each stream in @stream_list ... */
670 while (!list_empty(stream_list)) {
671 stream_discarded = false;
672 lte = container_of(stream_list->next,
673 struct wim_lookup_table_entry,
675 list_del(<e->write_streams_list);
676 if (lte->unhashed && !lte->unique_size) {
677 /* Unhashed stream that shares a size with some other
678 * stream in the WIM we are writing. The stream must be
679 * checksummed to know if we need to write it or not. */
680 struct wim_lookup_table_entry *tmp;
681 u32 orig_refcnt = lte->out_refcnt;
683 ret = hash_unhashed_stream(lte, lookup_table, &tmp);
688 /* We found a duplicate stream. */
689 if (orig_refcnt != tmp->out_refcnt) {
690 /* We have already written, or are going
691 * to write, the duplicate stream. So
692 * just skip to the next stream. */
693 DEBUG("Discarding duplicate stream of length %"PRIu64,
694 wim_resource_size(lte));
695 lte->no_progress = 0;
696 stream_discarded = true;
697 goto skip_to_progress;
702 /* Here, @lte is either a hashed stream or an unhashed stream
703 * with a unique size. In either case we know that the stream
704 * has to be written. In either case the SHA1 message digest
705 * will be calculated over the stream while writing it; however,
706 * in the former case this is done merely to check the data,
707 * while in the latter case this is done because we do not have
708 * the SHA1 message digest yet. */
709 wimlib_assert(lte->out_refcnt != 0);
711 lte->no_progress = 0;
712 ret = (*write_stream_cb)(lte, write_stream_ctx);
715 /* In parallel mode, some streams are deferred for later,
716 * serialized processing; ignore them here. */
720 list_del(<e->unhashed_list);
721 lookup_table_insert(lookup_table, lte);
725 if (!lte->no_progress) {
726 do_write_streams_progress(progress,
728 wim_resource_size(lte),
736 do_write_stream_list_serial(struct list_head *stream_list,
737 struct wim_lookup_table *lookup_table,
740 int write_resource_flags,
741 wimlib_progress_func_t progress_func,
742 union wimlib_progress_info *progress)
744 struct serial_write_stream_ctx ctx = {
746 .out_ctype = out_ctype,
747 .write_resource_flags = write_resource_flags,
749 return do_write_stream_list(stream_list,
758 write_flags_to_resource_flags(int write_flags)
760 int resource_flags = 0;
762 if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
763 resource_flags |= WIMLIB_RESOURCE_FLAG_RECOMPRESS;
764 return resource_flags;
768 write_stream_list_serial(struct list_head *stream_list,
769 struct wim_lookup_table *lookup_table,
772 int write_resource_flags,
773 wimlib_progress_func_t progress_func,
774 union wimlib_progress_info *progress)
776 DEBUG("Writing stream list (serial version)");
777 progress->write_streams.num_threads = 1;
779 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
780 return do_write_stream_list_serial(stream_list,
784 write_resource_flags,
789 #ifdef ENABLE_MULTITHREADED_COMPRESSION
791 write_wim_chunks(struct message *msg, int out_fd,
792 struct chunk_table *chunk_tab)
794 for (unsigned i = 0; i < msg->num_chunks; i++) {
795 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
796 chunk_tab->cur_offset += msg->out_chunks[i].iov_len;
798 if (full_writev(out_fd, msg->out_chunks,
799 msg->num_chunks) != msg->total_out_bytes)
801 ERROR_WITH_ERRNO("Failed to write WIM chunks");
802 return WIMLIB_ERR_WRITE;
807 struct main_writer_thread_ctx {
808 struct list_head *stream_list;
809 struct wim_lookup_table *lookup_table;
812 int write_resource_flags;
813 struct shared_queue *res_to_compress_queue;
814 struct shared_queue *compressed_res_queue;
816 wimlib_progress_func_t progress_func;
817 union wimlib_progress_info *progress;
819 struct list_head available_msgs;
820 struct list_head outstanding_streams;
821 struct list_head serial_streams;
822 size_t num_outstanding_messages;
824 SHA_CTX next_sha_ctx;
827 struct wim_lookup_table_entry *next_lte;
829 struct message *msgs;
830 struct message *next_msg;
831 struct chunk_table *cur_chunk_tab;
835 init_message(struct message *msg)
837 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
838 msg->compressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
839 msg->uncompressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
840 if (msg->compressed_chunks[i] == NULL ||
841 msg->uncompressed_chunks[i] == NULL)
842 return WIMLIB_ERR_NOMEM;
848 destroy_message(struct message *msg)
850 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
851 FREE(msg->compressed_chunks[i]);
852 FREE(msg->uncompressed_chunks[i]);
857 free_messages(struct message *msgs, size_t num_messages)
860 for (size_t i = 0; i < num_messages; i++)
861 destroy_message(&msgs[i]);
866 static struct message *
867 allocate_messages(size_t num_messages)
869 struct message *msgs;
871 msgs = CALLOC(num_messages, sizeof(struct message));
874 for (size_t i = 0; i < num_messages; i++) {
875 if (init_message(&msgs[i])) {
876 free_messages(msgs, num_messages);
884 main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
886 while (ctx->num_outstanding_messages--)
887 shared_queue_get(ctx->compressed_res_queue);
888 free_messages(ctx->msgs, ctx->num_messages);
889 FREE(ctx->cur_chunk_tab);
893 main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
895 /* Pre-allocate all the buffers that will be needed to do the chunk
897 ctx->msgs = allocate_messages(ctx->num_messages);
899 return WIMLIB_ERR_NOMEM;
901 /* Initially, all the messages are available to use. */
902 INIT_LIST_HEAD(&ctx->available_msgs);
903 for (size_t i = 0; i < ctx->num_messages; i++)
904 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
906 /* outstanding_streams is the list of streams that currently have had
907 * chunks sent off for compression.
909 * The first stream in outstanding_streams is the stream that is
910 * currently being written.
912 * The last stream in outstanding_streams is the stream that is
913 * currently being read and having chunks fed to the compressor threads.
915 INIT_LIST_HEAD(&ctx->outstanding_streams);
916 ctx->num_outstanding_messages = 0;
918 ctx->next_msg = NULL;
920 /* Resources that don't need any chunks compressed are added to this
921 * list and written directly by the main thread. */
922 INIT_LIST_HEAD(&ctx->serial_streams);
924 ctx->cur_chunk_tab = NULL;
930 receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
933 struct wim_lookup_table_entry *cur_lte;
936 wimlib_assert(!list_empty(&ctx->outstanding_streams));
937 wimlib_assert(ctx->num_outstanding_messages != 0);
939 cur_lte = container_of(ctx->outstanding_streams.next,
940 struct wim_lookup_table_entry,
941 being_compressed_list);
943 /* Get the next message from the queue and process it.
944 * The message will contain 1 or more data chunks that have been
946 msg = shared_queue_get(ctx->compressed_res_queue);
947 msg->complete = true;
948 --ctx->num_outstanding_messages;
950 /* Is this the next chunk in the current resource? If it's not
951 * (i.e., an earlier chunk in a same or different resource
952 * hasn't been compressed yet), do nothing, and keep this
953 * message around until all earlier chunks are received.
955 * Otherwise, write all the chunks we can. */
956 while (cur_lte != NULL &&
957 !list_empty(&cur_lte->msg_list)
958 && (msg = container_of(cur_lte->msg_list.next,
962 list_move(&msg->list, &ctx->available_msgs);
963 if (msg->begin_chunk == 0) {
964 /* This is the first set of chunks. Leave space
965 * for the chunk table in the output file. */
966 off_t cur_offset = filedes_offset(ctx->out_fd);
967 if (cur_offset == -1)
968 return WIMLIB_ERR_WRITE;
969 ret = begin_wim_resource_chunk_tab(cur_lte,
972 &ctx->cur_chunk_tab);
977 /* Write the compressed chunks from the message. */
978 ret = write_wim_chunks(msg, ctx->out_fd, ctx->cur_chunk_tab);
982 /* Was this the last chunk of the stream? If so, finish
984 if (list_empty(&cur_lte->msg_list) &&
985 msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
990 ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
996 list_del(&cur_lte->being_compressed_list);
998 /* Grab the offset of this stream in the output file
999 * from the chunk table before we free it. */
1000 offset = ctx->cur_chunk_tab->file_offset;
1002 FREE(ctx->cur_chunk_tab);
1003 ctx->cur_chunk_tab = NULL;
1005 if (res_csize >= wim_resource_size(cur_lte)) {
1006 /* Oops! We compressed the resource to
1007 * larger than the original size. Write
1008 * the resource uncompressed instead. */
1009 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
1010 "writing uncompressed instead",
1011 wim_resource_size(cur_lte), res_csize);
1012 ret = seek_and_truncate(ctx->out_fd, offset);
1015 ret = write_wim_resource(cur_lte,
1017 WIMLIB_COMPRESSION_TYPE_NONE,
1018 &cur_lte->output_resource_entry,
1019 ctx->write_resource_flags);
1023 cur_lte->output_resource_entry.size =
1026 cur_lte->output_resource_entry.original_size =
1027 cur_lte->resource_entry.original_size;
1029 cur_lte->output_resource_entry.offset =
1032 cur_lte->output_resource_entry.flags =
1033 cur_lte->resource_entry.flags |
1034 WIM_RESHDR_FLAG_COMPRESSED;
1037 do_write_streams_progress(ctx->progress,
1039 wim_resource_size(cur_lte),
1042 /* Since we just finished writing a stream, write any
1043 * streams that have been added to the serial_streams
1044 * list for direct writing by the main thread (e.g.
1045 * resources that don't need to be compressed because
1046 * the desired compression type is the same as the
1047 * previous compression type). */
1048 if (!list_empty(&ctx->serial_streams)) {
1049 ret = do_write_stream_list_serial(&ctx->serial_streams,
1053 ctx->write_resource_flags,
1060 /* Advance to the next stream to write. */
1061 if (list_empty(&ctx->outstanding_streams)) {
1064 cur_lte = container_of(ctx->outstanding_streams.next,
1065 struct wim_lookup_table_entry,
1066 being_compressed_list);
1073 /* Called when the main thread has read a new chunk of data. */
1075 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
1077 struct main_writer_thread_ctx *ctx = _ctx;
1079 struct message *next_msg;
1080 u64 next_chunk_in_msg;
1082 /* Update SHA1 message digest for the stream currently being read by the
1084 sha1_update(&ctx->next_sha_ctx, chunk, chunk_size);
1086 /* We send chunks of data to the compressor chunks in batches which we
1087 * refer to as "messages". @next_msg is the message that is currently
1088 * being prepared to send off. If it is NULL, that indicates that we
1089 * need to start a new message. */
1090 next_msg = ctx->next_msg;
1092 /* We need to start a new message. First check to see if there
1093 * is a message available in the list of available messages. If
1094 * so, we can just take one. If not, all the messages (there is
1095 * a fixed number of them, proportional to the number of
1096 * threads) have been sent off to the compressor threads, so we
1097 * receive messages from the compressor threads containing
1098 * compressed chunks of data.
1100 * We may need to receive multiple messages before one is
1101 * actually available to use because messages received that are
1102 * *not* for the very next set of chunks to compress must be
1103 * buffered until it's time to write those chunks. */
1104 while (list_empty(&ctx->available_msgs)) {
1105 ret = receive_compressed_chunks(ctx);
1110 next_msg = container_of(ctx->available_msgs.next,
1111 struct message, list);
1112 list_del(&next_msg->list);
1113 next_msg->complete = false;
1114 next_msg->begin_chunk = ctx->next_chunk;
1115 next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
1116 ctx->next_num_chunks - ctx->next_chunk);
1117 ctx->next_msg = next_msg;
1120 /* Fill in the next chunk to compress */
1121 next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;
1123 next_msg->uncompressed_chunk_sizes[next_chunk_in_msg] = chunk_size;
1124 memcpy(next_msg->uncompressed_chunks[next_chunk_in_msg],
1127 if (++next_chunk_in_msg == next_msg->num_chunks) {
1128 /* Send off an array of chunks to compress */
1129 list_add_tail(&next_msg->list, &ctx->next_lte->msg_list);
1130 shared_queue_put(ctx->res_to_compress_queue, next_msg);
1131 ++ctx->num_outstanding_messages;
1132 ctx->next_msg = NULL;
1138 main_writer_thread_finish(void *_ctx)
1140 struct main_writer_thread_ctx *ctx = _ctx;
1142 while (ctx->num_outstanding_messages != 0) {
1143 ret = receive_compressed_chunks(ctx);
1147 wimlib_assert(list_empty(&ctx->outstanding_streams));
1148 return do_write_stream_list_serial(&ctx->serial_streams,
1152 ctx->write_resource_flags,
1158 submit_stream_for_compression(struct wim_lookup_table_entry *lte,
1159 struct main_writer_thread_ctx *ctx)
1163 /* Read the entire stream @lte, feeding its data chunks to the
1164 * compressor threads. Also SHA1-sum the stream; this is required in
1165 * the case that @lte is unhashed, and a nice additional verification
1166 * when @lte is already hashed. */
1167 sha1_init(&ctx->next_sha_ctx);
1168 ctx->next_chunk = 0;
1169 ctx->next_num_chunks = wim_resource_chunks(lte);
1170 ctx->next_lte = lte;
1171 INIT_LIST_HEAD(<e->msg_list);
1172 list_add_tail(<e->being_compressed_list, &ctx->outstanding_streams);
1173 ret = read_resource_prefix(lte, wim_resource_size(lte),
1174 main_writer_thread_cb, ctx, 0);
1176 wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
1177 ret = finalize_and_check_sha1(&ctx->next_sha_ctx, lte);
1183 main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
1185 struct main_writer_thread_ctx *ctx = _ctx;
1188 if (wim_resource_size(lte) < 1000 ||
1189 ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
1190 (lte->resource_location == RESOURCE_IN_WIM &&
1191 !(ctx->write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
1192 wimlib_get_compression_type(lte->wim) == ctx->out_ctype))
1194 /* Stream is too small or isn't being compressed. Process it by
1195 * the main thread when we have a chance. We can't necessarily
1196 * process it right here, as the main thread could be in the
1197 * middle of writing a different stream. */
1198 list_add_tail(<e->write_streams_list, &ctx->serial_streams);
1202 ret = submit_stream_for_compression(lte, ctx);
1204 lte->no_progress = 1;
1209 get_default_num_threads(void)
1212 return win32_get_number_of_processors();
1214 return sysconf(_SC_NPROCESSORS_ONLN);
1218 /* Equivalent to write_stream_list_serial(), except this takes a @num_threads
1219 * parameter and will perform compression using that many threads. Falls
1220 * back to write_stream_list_serial() on certain errors, such as a failure to
1221 * create the number of threads requested.
1223 * High level description of the algorithm for writing compressed streams in
1224 * parallel: We perform compression on chunks of size WIM_CHUNK_SIZE bytes
1225 * rather than on full files. The currently executing thread becomes the main
1226 * thread and is entirely in charge of reading the data to compress (which may
1227 * be in any location understood by the resource code--- such as in an external
1228 * file being captured, or in another WIM file from which an image is being
1229 * exported) and actually writing the compressed data to the output file.
1230 * Additional threads are "compressor threads" and all execute the
1231 * compressor_thread_proc, where they repeatedly retrieve buffers of data from
1232 * the main thread, compress them, and hand them back to the main thread.
1234 * Certain streams, such as streams that do not need to be compressed (e.g.
1235 * input compression type same as output compression type) or streams of very
1236 * small size are placed in a list (main_writer_thread_ctx.serial_list) and
1237 * handled entirely by the main thread at an appropriate time.
1239 * At any given point in time, multiple streams may be having chunks compressed
1240 * concurrently. The stream that the main thread is currently *reading* may be
1241 * later in the list that the stream that the main thread is currently
1245 write_stream_list_parallel(struct list_head *stream_list,
1246 struct wim_lookup_table *lookup_table,
1249 int write_resource_flags,
1250 wimlib_progress_func_t progress_func,
1251 union wimlib_progress_info *progress,
1252 unsigned num_threads)
1255 struct shared_queue res_to_compress_queue;
1256 struct shared_queue compressed_res_queue;
1257 pthread_t *compressor_threads = NULL;
1259 if (num_threads == 0) {
1260 long nthreads = get_default_num_threads();
1261 if (nthreads < 1 || nthreads > UINT_MAX) {
1262 WARNING("Could not determine number of processors! Assuming 1");
1264 } else if (nthreads == 1) {
1265 goto out_serial_quiet;
1267 num_threads = nthreads;
1271 DEBUG("Writing stream list (parallel version, num_threads=%u)",
1274 progress->write_streams.num_threads = num_threads;
1276 static const size_t MESSAGES_PER_THREAD = 2;
1277 size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1279 DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1281 ret = shared_queue_init(&res_to_compress_queue, queue_size);
1285 ret = shared_queue_init(&compressed_res_queue, queue_size);
1287 goto out_destroy_res_to_compress_queue;
1289 struct compressor_thread_params params;
1290 params.res_to_compress_queue = &res_to_compress_queue;
1291 params.compressed_res_queue = &compressed_res_queue;
1292 params.compress = get_compress_func(out_ctype);
1294 compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1295 if (!compressor_threads) {
1296 ret = WIMLIB_ERR_NOMEM;
1297 goto out_destroy_compressed_res_queue;
1300 for (unsigned i = 0; i < num_threads; i++) {
1301 DEBUG("pthread_create thread %u of %u", i + 1, num_threads);
1302 ret = pthread_create(&compressor_threads[i], NULL,
1303 compressor_thread_proc, ¶ms);
1306 ERROR_WITH_ERRNO("Failed to create compressor "
1308 i + 1, num_threads);
1315 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
1317 struct main_writer_thread_ctx ctx;
1318 ctx.stream_list = stream_list;
1319 ctx.lookup_table = lookup_table;
1320 ctx.out_fd = out_fd;
1321 ctx.out_ctype = out_ctype;
1322 ctx.res_to_compress_queue = &res_to_compress_queue;
1323 ctx.compressed_res_queue = &compressed_res_queue;
1324 ctx.num_messages = queue_size;
1325 ctx.write_resource_flags = write_resource_flags;
1326 ctx.progress_func = progress_func;
1327 ctx.progress = progress;
1328 ret = main_writer_thread_init_ctx(&ctx);
1331 ret = do_write_stream_list(stream_list, lookup_table,
1332 main_thread_process_next_stream,
1333 &ctx, progress_func, progress);
1335 goto out_destroy_ctx;
1337 /* The main thread has finished reading all streams that are going to be
1338 * compressed in parallel, and it now needs to wait for all remaining
1339 * chunks to be compressed so that the remaining streams can actually be
1340 * written to the output file. Furthermore, any remaining streams that
1341 * had processing deferred to the main thread need to be handled. These
1342 * tasks are done by the main_writer_thread_finish() function. */
1343 ret = main_writer_thread_finish(&ctx);
1345 main_writer_thread_destroy_ctx(&ctx);
1347 for (unsigned i = 0; i < num_threads; i++)
1348 shared_queue_put(&res_to_compress_queue, NULL);
1350 for (unsigned i = 0; i < num_threads; i++) {
1351 if (pthread_join(compressor_threads[i], NULL)) {
1352 WARNING_WITH_ERRNO("Failed to join compressor "
1354 i + 1, num_threads);
1357 FREE(compressor_threads);
1358 out_destroy_compressed_res_queue:
1359 shared_queue_destroy(&compressed_res_queue);
1360 out_destroy_res_to_compress_queue:
1361 shared_queue_destroy(&res_to_compress_queue);
1362 if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
1365 WARNING("Falling back to single-threaded compression");
1367 return write_stream_list_serial(stream_list,
1371 write_resource_flags,
1379 * Write a list of streams to a WIM (@out_fd) using the compression type
1380 * @out_ctype and up to @num_threads compressor threads.
1383 write_stream_list(struct list_head *stream_list,
1384 struct wim_lookup_table *lookup_table,
1385 int out_fd, int out_ctype, int write_flags,
1386 unsigned num_threads, wimlib_progress_func_t progress_func)
1388 struct wim_lookup_table_entry *lte;
1389 size_t num_streams = 0;
1390 u64 total_bytes = 0;
1391 u64 total_compression_bytes = 0;
1392 union wimlib_progress_info progress;
1394 int write_resource_flags;
1396 if (list_empty(stream_list))
1399 write_resource_flags = write_flags_to_resource_flags(write_flags);
1401 /* Calculate the total size of the streams to be written. Note: this
1402 * will be the uncompressed size, as we may not know the compressed size
1403 * yet, and also this will assume that every unhashed stream will be
1404 * written (which will not necessarily be the case). */
1405 list_for_each_entry(lte, stream_list, write_streams_list) {
1407 total_bytes += wim_resource_size(lte);
1408 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
1409 && (wim_resource_compression_type(lte) != out_ctype ||
1410 (write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS)))
1412 total_compression_bytes += wim_resource_size(lte);
1415 progress.write_streams.total_bytes = total_bytes;
1416 progress.write_streams.total_streams = num_streams;
1417 progress.write_streams.completed_bytes = 0;
1418 progress.write_streams.completed_streams = 0;
1419 progress.write_streams.num_threads = num_threads;
1420 progress.write_streams.compression_type = out_ctype;
1421 progress.write_streams._private = 0;
1423 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1424 if (total_compression_bytes >= 2000000 && num_threads != 1)
1425 ret = write_stream_list_parallel(stream_list,
1429 write_resource_flags,
1435 ret = write_stream_list_serial(stream_list,
1439 write_resource_flags,
1445 struct stream_size_table {
1446 struct hlist_head *array;
1452 init_stream_size_table(struct stream_size_table *tab, size_t capacity)
1454 tab->array = CALLOC(capacity, sizeof(tab->array[0]));
1456 return WIMLIB_ERR_NOMEM;
1457 tab->num_entries = 0;
1458 tab->capacity = capacity;
1463 destroy_stream_size_table(struct stream_size_table *tab)
1469 stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
1471 struct stream_size_table *tab = _tab;
1473 struct wim_lookup_table_entry *same_size_lte;
1474 struct hlist_node *tmp;
1476 pos = hash_u64(wim_resource_size(lte)) % tab->capacity;
1477 lte->unique_size = 1;
1478 hlist_for_each_entry(same_size_lte, tmp, &tab->array[pos], hash_list_2) {
1479 if (wim_resource_size(same_size_lte) == wim_resource_size(lte)) {
1480 lte->unique_size = 0;
1481 same_size_lte->unique_size = 0;
1486 hlist_add_head(<e->hash_list_2, &tab->array[pos]);
1492 struct lte_overwrite_prepare_args {
1495 struct list_head stream_list;
1496 struct stream_size_table stream_size_tab;
1499 /* First phase of preparing streams for an in-place overwrite. This is called
1500 * on all streams, both hashed and unhashed, except the metadata resources. */
1502 lte_overwrite_prepare(struct wim_lookup_table_entry *lte, void *_args)
1504 struct lte_overwrite_prepare_args *args = _args;
1506 wimlib_assert(!(lte->resource_entry.flags & WIM_RESHDR_FLAG_METADATA));
1507 if (lte->resource_location != RESOURCE_IN_WIM || lte->wim != args->wim)
1508 list_add_tail(<e->write_streams_list, &args->stream_list);
1509 lte->out_refcnt = lte->refcnt;
1510 stream_size_table_insert(lte, &args->stream_size_tab);
1514 /* Second phase of preparing streams for an in-place overwrite. This is called
1515 * on existing metadata resources and hashed streams, but not unhashed streams.
1517 * NOTE: lte->output_resource_entry is in union with lte->hash_list_2, so
1518 * lte_overwrite_prepare_2() must be called after lte_overwrite_prepare(), as
1519 * the latter uses lte->hash_list_2, while the former expects to set
1520 * lte->output_resource_entry. */
1522 lte_overwrite_prepare_2(struct wim_lookup_table_entry *lte, void *_args)
1524 struct lte_overwrite_prepare_args *args = _args;
1526 if (lte->resource_location == RESOURCE_IN_WIM && lte->wim == args->wim) {
1527 /* We can't do an in place overwrite on the WIM if there are
1528 * streams after the XML data. */
1529 if (lte->resource_entry.offset +
1530 lte->resource_entry.size > args->end_offset)
1532 #ifdef ENABLE_ERROR_MESSAGES
1533 ERROR("The following resource is after the XML data:");
1534 print_lookup_table_entry(lte, stderr);
1536 return WIMLIB_ERR_RESOURCE_ORDER;
1538 copy_resource_entry(<e->output_resource_entry,
1539 <e->resource_entry);
1544 /* Given a WIM that we are going to overwrite in place with zero or more
1545 * additional streams added, construct a list the list of new unique streams
1546 * ('struct wim_lookup_table_entry's) that must be written, plus any unhashed
1547 * streams that need to be added but may be identical to other hashed or
1548 * unhashed streams. These unhashed streams are checksummed while the streams
1549 * are being written. To aid this process, the member @unique_size is set to 1
1550 * on streams that have a unique size and therefore must be written.
1552 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1553 * indicate the number of times the stream is referenced in only the streams
1554 * that are being written; this may still be adjusted later when unhashed
1555 * streams are being resolved.
1558 prepare_streams_for_overwrite(WIMStruct *wim, off_t end_offset,
1559 struct list_head *stream_list)
1562 struct lte_overwrite_prepare_args args;
1566 args.end_offset = end_offset;
1567 ret = init_stream_size_table(&args.stream_size_tab,
1568 wim->lookup_table->capacity);
1572 INIT_LIST_HEAD(&args.stream_list);
1573 for (i = 0; i < wim->hdr.image_count; i++) {
1574 struct wim_image_metadata *imd;
1575 struct wim_lookup_table_entry *lte;
1577 imd = wim->image_metadata[i];
1578 image_for_each_unhashed_stream(lte, imd)
1579 lte_overwrite_prepare(lte, &args);
1581 for_lookup_table_entry(wim->lookup_table, lte_overwrite_prepare, &args);
1582 list_transfer(&args.stream_list, stream_list);
1584 for (i = 0; i < wim->hdr.image_count; i++) {
1585 ret = lte_overwrite_prepare_2(wim->image_metadata[i]->metadata_lte,
1588 goto out_destroy_stream_size_table;
1590 ret = for_lookup_table_entry(wim->lookup_table,
1591 lte_overwrite_prepare_2, &args);
1592 out_destroy_stream_size_table:
1593 destroy_stream_size_table(&args.stream_size_tab);
1598 struct find_streams_ctx {
1599 struct list_head stream_list;
1600 struct stream_size_table stream_size_tab;
1604 inode_find_streams_to_write(struct wim_inode *inode,
1605 struct wim_lookup_table *table,
1606 struct list_head *stream_list,
1607 struct stream_size_table *tab)
1609 struct wim_lookup_table_entry *lte;
1610 for (unsigned i = 0; i <= inode->i_num_ads; i++) {
1611 lte = inode_stream_lte(inode, i, table);
1613 if (lte->out_refcnt == 0) {
1615 stream_size_table_insert(lte, tab);
1616 list_add_tail(<e->write_streams_list, stream_list);
1618 lte->out_refcnt += inode->i_nlink;
1624 image_find_streams_to_write(WIMStruct *w)
1626 struct find_streams_ctx *ctx;
1627 struct wim_image_metadata *imd;
1628 struct wim_inode *inode;
1629 struct wim_lookup_table_entry *lte;
1632 imd = wim_get_current_image_metadata(w);
1634 image_for_each_unhashed_stream(lte, imd)
1635 lte->out_refcnt = 0;
1637 /* Go through this image's inodes to find any streams that have not been
1639 image_for_each_inode(inode, imd) {
1640 inode_find_streams_to_write(inode, w->lookup_table,
1642 &ctx->stream_size_tab);
1647 /* Given a WIM that from which one or all of the images is being written, build
1648 * the list of unique streams ('struct wim_lookup_table_entry's) that must be
1649 * written, plus any unhashed streams that need to be written but may be
1650 * identical to other hashed or unhashed streams being written. These unhashed
1651 * streams are checksummed while the streams are being written. To aid this
1652 * process, the member @unique_size is set to 1 on streams that have a unique
1653 * size and therefore must be written.
1655 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1656 * indicate the number of times the stream is referenced in only the streams
1657 * that are being written; this may still be adjusted later when unhashed
1658 * streams are being resolved.
1661 prepare_stream_list(WIMStruct *wim, int image, struct list_head *stream_list)
1664 struct find_streams_ctx ctx;
1666 for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
1667 ret = init_stream_size_table(&ctx.stream_size_tab,
1668 wim->lookup_table->capacity);
1671 for_lookup_table_entry(wim->lookup_table, stream_size_table_insert,
1672 &ctx.stream_size_tab);
1673 INIT_LIST_HEAD(&ctx.stream_list);
1674 wim->private = &ctx;
1675 ret = for_image(wim, image, image_find_streams_to_write);
1676 destroy_stream_size_table(&ctx.stream_size_tab);
1678 list_transfer(&ctx.stream_list, stream_list);
1682 /* Writes the streams for the specified @image in @wim to @wim->out_fd.
1685 write_wim_streams(WIMStruct *wim, int image, int write_flags,
1686 unsigned num_threads,
1687 wimlib_progress_func_t progress_func)
1690 struct list_head stream_list;
1692 ret = prepare_stream_list(wim, image, &stream_list);
1695 return write_stream_list(&stream_list,
1698 wimlib_get_compression_type(wim),
1705 * Finish writing a WIM file: write the lookup table, xml data, and integrity
1706 * table (optional), then overwrite the WIM header.
1708 * write_flags is a bitwise OR of the following:
1710 * (public) WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
1711 * Include an integrity table.
1713 * (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
1714 * Don't write the lookup table.
1716 * (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
1717 * When (if) writing the integrity table, re-use entries from the
1718 * existing integrity table, if possible.
1720 * (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
1721 * After writing the XML data but before writing the integrity
1722 * table, write a temporary WIM header and flush the stream so that
1723 * the WIM is less likely to become corrupted upon abrupt program
1726 * (private) WIMLIB_WRITE_FLAG_FSYNC:
1727 * fsync() the output file before closing it.
1731 finish_write(WIMStruct *w, int image, int write_flags,
1732 wimlib_progress_func_t progress_func)
1735 struct wim_header hdr;
1737 /* @hdr will be the header for the new WIM. First copy all the data
1738 * from the header in the WIMStruct; then set all the fields that may
1739 * have changed, including the resource entries, boot index, and image
1741 memcpy(&hdr, &w->hdr, sizeof(struct wim_header));
1743 /* Set image count and boot index correctly for single image writes */
1744 if (image != WIMLIB_ALL_IMAGES) {
1745 hdr.image_count = 1;
1746 if (hdr.boot_idx == image)
1752 /* In the WIM header, there is room for the resource entry for a
1753 * metadata resource labeled as the "boot metadata". This entry should
1754 * be zeroed out if there is no bootable image (boot_idx 0). Otherwise,
1755 * it should be a copy of the resource entry for the image that is
1756 * marked as bootable. This is not well documented... */
1757 if (hdr.boot_idx == 0) {
1758 zero_resource_entry(&hdr.boot_metadata_res_entry);
1760 copy_resource_entry(&hdr.boot_metadata_res_entry,
1761 &w->image_metadata[ hdr.boot_idx- 1
1762 ]->metadata_lte->output_resource_entry);
1765 if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
1766 ret = write_lookup_table(w, image, &hdr.lookup_table_res_entry);
1771 ret = write_xml_data(w->wim_info, image, w->out_fd,
1772 (write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE) ?
1773 wim_info_get_total_bytes(w->wim_info) : 0,
1774 &hdr.xml_res_entry);
1778 if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
1779 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
1780 struct wim_header checkpoint_hdr;
1781 memcpy(&checkpoint_hdr, &hdr, sizeof(struct wim_header));
1782 zero_resource_entry(&checkpoint_hdr.integrity);
1783 ret = write_header(&checkpoint_hdr, w->out_fd);
1788 off_t old_lookup_table_end;
1789 off_t new_lookup_table_end;
1790 if (write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE) {
1791 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
1792 w->hdr.lookup_table_res_entry.size;
1794 old_lookup_table_end = 0;
1796 new_lookup_table_end = hdr.lookup_table_res_entry.offset +
1797 hdr.lookup_table_res_entry.size;
1799 ret = write_integrity_table(w->out_fd,
1801 new_lookup_table_end,
1802 old_lookup_table_end,
1807 zero_resource_entry(&hdr.integrity);
1810 ret = write_header(&hdr, w->out_fd);
1814 if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
1815 if (fsync(w->out_fd)) {
1816 ERROR_WITH_ERRNO("Error syncing data to WIM file");
1817 ret = WIMLIB_ERR_WRITE;
1821 if (close(w->out_fd)) {
1822 ERROR_WITH_ERRNO("Failed to close the output WIM file");
1824 ret = WIMLIB_ERR_WRITE;
1830 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
1832 lock_wim(WIMStruct *w, int fd)
1835 if (fd != -1 && !w->wim_locked) {
1836 ret = flock(fd, LOCK_EX | LOCK_NB);
1838 if (errno == EWOULDBLOCK) {
1839 ERROR("`%"TS"' is already being modified or has been "
1840 "mounted read-write\n"
1841 " by another process!", w->filename);
1842 ret = WIMLIB_ERR_ALREADY_LOCKED;
1844 WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
1857 open_wim_writable(WIMStruct *w, const tchar *path, int open_flags)
1859 w->out_fd = topen(path, open_flags | O_BINARY, 0644);
1860 if (w->out_fd == -1) {
1861 ERROR_WITH_ERRNO("Failed to open `%"TS"' for writing", path);
1862 return WIMLIB_ERR_OPEN;
1869 close_wim_writable(WIMStruct *w)
1871 if (w->out_fd != -1) {
1872 if (close(w->out_fd))
1873 WARNING_WITH_ERRNO("Failed to close output WIM");
1878 /* Open file stream and write dummy header for WIM. */
1880 begin_write(WIMStruct *w, const tchar *path, int write_flags)
1883 int open_flags = O_TRUNC | O_CREAT;
1884 if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
1885 open_flags |= O_RDWR;
1887 open_flags |= O_WRONLY;
1888 ret = open_wim_writable(w, path, open_flags);
1891 /* Write dummy header. It will be overwritten later. */
1892 ret = write_header(&w->hdr, w->out_fd);
1895 if (lseek(w->out_fd, WIM_HEADER_DISK_SIZE, SEEK_SET) == -1) {
1896 ERROR_WITH_ERRNO("Failed to seek to end of WIM header");
1897 return WIMLIB_ERR_WRITE;
1902 /* Writes a stand-alone WIM to a file. */
1904 wimlib_write(WIMStruct *w, const tchar *path,
1905 int image, int write_flags, unsigned num_threads,
1906 wimlib_progress_func_t progress_func)
1911 return WIMLIB_ERR_INVALID_PARAM;
1913 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1915 if (image != WIMLIB_ALL_IMAGES &&
1916 (image < 1 || image > w->hdr.image_count))
1917 return WIMLIB_ERR_INVALID_IMAGE;
1919 if (w->hdr.total_parts != 1) {
1920 ERROR("Cannot call wimlib_write() on part of a split WIM");
1921 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1924 ret = begin_write(w, path, write_flags);
1928 ret = write_wim_streams(w, image, write_flags, num_threads,
1934 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
1936 ret = for_image(w, image, write_metadata_resource);
1941 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
1943 ret = finish_write(w, image, write_flags, progress_func);
1944 /* finish_write() closed the WIM for us */
1947 close_wim_writable(w);
1949 DEBUG("wimlib_write(path=%"TS") = %d", path, ret);
1954 any_images_modified(WIMStruct *w)
1956 for (int i = 0; i < w->hdr.image_count; i++)
1957 if (w->image_metadata[i]->modified)
1963 * Overwrite a WIM, possibly appending streams to it.
1965 * A WIM looks like (or is supposed to look like) the following:
1967 * Header (212 bytes)
1968 * Streams and metadata resources (variable size)
1969 * Lookup table (variable size)
1970 * XML data (variable size)
1971 * Integrity table (optional) (variable size)
1973 * If we are not adding any streams or metadata resources, the lookup table is
1974 * unchanged--- so we only need to overwrite the XML data, integrity table, and
1975 * header. This operation is potentially unsafe if the program is abruptly
1976 * terminated while the XML data or integrity table are being overwritten, but
1977 * before the new header has been written. To partially alleviate this problem,
1978 * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
1979 * finish_write() to cause a temporary WIM header to be written after the XML
1980 * data has been written. This may prevent the WIM from becoming corrupted if
1981 * the program is terminated while the integrity table is being calculated (but
1982 * no guarantees, due to write re-ordering...).
1984 * If we are adding new streams or images (metadata resources), the lookup table
1985 * needs to be changed, and those streams need to be written. In this case, we
1986 * try to perform a safe update of the WIM file by writing the streams *after*
1987 * the end of the previous WIM, then writing the new lookup table, XML data, and
1988 * (optionally) integrity table following the new streams. This will produce a
1989 * layout like the following:
1991 * Header (212 bytes)
1992 * (OLD) Streams and metadata resources (variable size)
1993 * (OLD) Lookup table (variable size)
1994 * (OLD) XML data (variable size)
1995 * (OLD) Integrity table (optional) (variable size)
1996 * (NEW) Streams and metadata resources (variable size)
1997 * (NEW) Lookup table (variable size)
1998 * (NEW) XML data (variable size)
1999 * (NEW) Integrity table (optional) (variable size)
2001 * At all points, the WIM is valid as nothing points to the new data yet. Then,
2002 * the header is overwritten to point to the new lookup table, XML data, and
2003 * integrity table, to produce the following layout:
2005 * Header (212 bytes)
2006 * Streams and metadata resources (variable size)
2007 * Nothing (variable size)
2008 * More Streams and metadata resources (variable size)
2009 * Lookup table (variable size)
2010 * XML data (variable size)
2011 * Integrity table (optional) (variable size)
2013 * This method allows an image to be appended to a large WIM very quickly, and
2014 * is is crash-safe except in the case of write re-ordering, but the
2015 * disadvantage is that a small hole is left in the WIM where the old lookup
2016 * table, xml data, and integrity table were. (These usually only take up a
2017 * small amount of space compared to the streams, however.)
2020 overwrite_wim_inplace(WIMStruct *w, int write_flags,
2021 unsigned num_threads,
2022 wimlib_progress_func_t progress_func)
2025 struct list_head stream_list;
2027 u64 old_lookup_table_end, old_xml_begin, old_xml_end;
2030 DEBUG("Overwriting `%"TS"' in-place", w->filename);
2032 /* Make sure that the integrity table (if present) is after the XML
2033 * data, and that there are no stream resources, metadata resources, or
2034 * lookup tables after the XML data. Otherwise, these data would be
2036 old_xml_begin = w->hdr.xml_res_entry.offset;
2037 old_xml_end = old_xml_begin + w->hdr.xml_res_entry.size;
2038 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
2039 w->hdr.lookup_table_res_entry.size;
2040 if (w->hdr.integrity.offset != 0 && w->hdr.integrity.offset < old_xml_end) {
2041 ERROR("Didn't expect the integrity table to be before the XML data");
2042 return WIMLIB_ERR_RESOURCE_ORDER;
2045 if (old_lookup_table_end > old_xml_begin) {
2046 ERROR("Didn't expect the lookup table to be after the XML data");
2047 return WIMLIB_ERR_RESOURCE_ORDER;
2050 /* Set @old_wim_end, which indicates the point beyond which we don't
2051 * allow any file and metadata resources to appear without returning
2052 * WIMLIB_ERR_RESOURCE_ORDER (due to the fact that we would otherwise
2053 * overwrite these resources). */
2054 if (!w->deletion_occurred && !any_images_modified(w)) {
2055 /* If no images have been modified and no images have been
2056 * deleted, a new lookup table does not need to be written. We
2057 * shall write the new XML data and optional integrity table
2058 * immediately after the lookup table. Note that this may
2059 * overwrite an existing integrity table. */
2060 DEBUG("Skipping writing lookup table "
2061 "(no images modified or deleted)");
2062 old_wim_end = old_lookup_table_end;
2063 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
2064 WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
2065 } else if (w->hdr.integrity.offset) {
2066 /* Old WIM has an integrity table; begin writing new streams
2068 old_wim_end = w->hdr.integrity.offset + w->hdr.integrity.size;
2070 /* No existing integrity table; begin writing new streams after
2071 * the old XML data. */
2072 old_wim_end = old_xml_end;
2075 ret = prepare_streams_for_overwrite(w, old_wim_end, &stream_list);
2080 if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
2081 open_flags |= O_RDWR;
2083 open_flags |= O_WRONLY;
2084 ret = open_wim_writable(w, w->filename, open_flags);
2088 ret = lock_wim(w, w->out_fd);
2090 close_wim_writable(w);
2094 if (lseek(w->out_fd, old_wim_end, SEEK_SET) == -1) {
2095 ERROR_WITH_ERRNO("Can't seek to end of WIM");
2096 close_wim_writable(w);
2098 return WIMLIB_ERR_WRITE;
2101 DEBUG("Writing newly added streams (offset = %"PRIu64")",
2103 ret = write_stream_list(&stream_list,
2106 wimlib_get_compression_type(w),
2113 for (int i = 0; i < w->hdr.image_count; i++) {
2114 if (w->image_metadata[i]->modified) {
2115 select_wim_image(w, i + 1);
2116 ret = write_metadata_resource(w);
2121 write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
2122 ret = finish_write(w, WIMLIB_ALL_IMAGES, write_flags,
2125 close_wim_writable(w);
2126 if (ret != 0 && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
2127 WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
2128 w->filename, old_wim_end);
2129 /* Return value of truncate() is ignored because this is already
2131 (void)ttruncate(w->filename, old_wim_end);
2138 overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
2139 unsigned num_threads,
2140 wimlib_progress_func_t progress_func)
2142 size_t wim_name_len;
2145 DEBUG("Overwriting `%"TS"' via a temporary file", w->filename);
2147 /* Write the WIM to a temporary file in the same directory as the
2149 wim_name_len = tstrlen(w->filename);
2150 tchar tmpfile[wim_name_len + 10];
2151 tmemcpy(tmpfile, w->filename, wim_name_len);
2152 randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
2153 tmpfile[wim_name_len + 9] = T('\0');
2155 ret = wimlib_write(w, tmpfile, WIMLIB_ALL_IMAGES,
2156 write_flags | WIMLIB_WRITE_FLAG_FSYNC,
2157 num_threads, progress_func);
2159 ERROR("Failed to write the WIM file `%"TS"'", tmpfile);
2165 DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, w->filename);
2166 /* Rename the new file to the old file .*/
2167 if (trename(tmpfile, w->filename) != 0) {
2168 ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
2169 tmpfile, w->filename);
2170 ret = WIMLIB_ERR_RENAME;
2174 if (progress_func) {
2175 union wimlib_progress_info progress;
2176 progress.rename.from = tmpfile;
2177 progress.rename.to = w->filename;
2178 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
2182 /* Remove temporary file. */
2183 if (tunlink(tmpfile) != 0)
2184 WARNING_WITH_ERRNO("Failed to remove `%"TS"'", tmpfile);
2190 * Writes a WIM file to the original file that it was read from, overwriting it.
2193 wimlib_overwrite(WIMStruct *w, int write_flags,
2194 unsigned num_threads,
2195 wimlib_progress_func_t progress_func)
2197 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2200 return WIMLIB_ERR_NO_FILENAME;
2202 if (w->hdr.total_parts != 1) {
2203 ERROR("Cannot modify a split WIM");
2204 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
2207 if ((!w->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
2208 && !(write_flags & WIMLIB_WRITE_FLAG_REBUILD))
2211 ret = overwrite_wim_inplace(w, write_flags, num_threads,
2213 if (ret == WIMLIB_ERR_RESOURCE_ORDER)
2214 WARNING("Falling back to re-building entire WIM");
2218 return overwrite_wim_via_tmpfile(w, write_flags, num_threads,