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/.
29 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
30 /* On BSD, this should be included before "list.h" so that "list.h" can
31 * overwrite the LIST_HEAD macro. */
32 # include <sys/file.h>
40 #include "wimlib_internal.h"
41 #include "buffer_io.h"
43 #include "lookup_table.h"
46 #ifdef ENABLE_MULTITHREADED_COMPRESSION
55 # include <ntfs-3g/attrib.h>
56 # include <ntfs-3g/inode.h>
57 # include <ntfs-3g/dir.h>
68 /* Chunk table that's located at the beginning of each compressed resource in
69 * the WIM. (This is not the on-disk format; the on-disk format just has an
70 * array of offsets.) */
74 u64 original_resource_size;
75 u64 bytes_per_chunk_entry;
83 * Allocates and initializes a chunk table, and reserves space for it in the
87 begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
90 struct chunk_table **chunk_tab_ret)
92 u64 size = wim_resource_size(lte);
93 u64 num_chunks = (size + WIM_CHUNK_SIZE - 1) / WIM_CHUNK_SIZE;
94 size_t alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
95 struct chunk_table *chunk_tab = CALLOC(1, alloc_size);
97 DEBUG("Begin chunk table for stream with size %"PRIu64, size);
100 ERROR("Failed to allocate chunk table for %"PRIu64" byte "
102 return WIMLIB_ERR_NOMEM;
104 chunk_tab->file_offset = file_offset;
105 chunk_tab->num_chunks = num_chunks;
106 chunk_tab->original_resource_size = size;
107 chunk_tab->bytes_per_chunk_entry = (size >= (1ULL << 32)) ? 8 : 4;
108 chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
110 chunk_tab->cur_offset = 0;
111 chunk_tab->cur_offset_p = chunk_tab->offsets;
113 if (fwrite(chunk_tab, 1, chunk_tab->table_disk_size, out_fp) !=
114 chunk_tab->table_disk_size) {
115 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
118 return WIMLIB_ERR_WRITE;
120 *chunk_tab_ret = chunk_tab;
125 * compress_func_t- Pointer to a function to compresses a chunk
126 * of a WIM resource. This may be either
127 * wimlib_xpress_compress() (xpress-compress.c) or
128 * wimlib_lzx_compress() (lzx-compress.c).
130 * @chunk: Uncompressed data of the chunk.
131 * @chunk_size: Size of the uncompressed chunk, in bytes.
132 * @out: Pointer to output buffer of size at least (@chunk_size - 1) bytes.
134 * Returns the size of the compressed data written to @out in bytes, or 0 if the
135 * data could not be compressed to (@chunk_size - 1) bytes or fewer.
137 * As a special requirement, the compression code is optimized for the WIM
138 * format and therefore requires (@chunk_size <= 32768).
140 * As another special requirement, the compression code will read up to 8 bytes
141 * off the end of the @chunk array for performance reasons. The values of these
142 * bytes will not affect the output of the compression, but the calling code
143 * must make sure that the buffer holding the uncompressed chunk is actually at
144 * least (@chunk_size + 8) bytes, or at least that these extra bytes are in
145 * mapped memory that will not cause a memory access violation if accessed.
147 typedef unsigned (*compress_func_t)(const void *chunk, unsigned chunk_size,
150 static compress_func_t
151 get_compress_func(int out_ctype)
153 if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX)
154 return wimlib_lzx_compress;
156 return wimlib_xpress_compress;
160 * Writes a chunk of a WIM resource to an output file.
162 * @chunk: Uncompressed data of the chunk.
163 * @chunk_size: Size of the chunk (<= WIM_CHUNK_SIZE)
164 * @out_fp: FILE * to write the chunk to.
165 * @compress: Compression function to use (NULL if writing uncompressed
167 * @chunk_tab: Pointer to chunk table being created. It is updated with the
168 * offset of the chunk we write.
170 * Returns 0 on success; nonzero on failure.
173 write_wim_resource_chunk(const void * restrict chunk,
175 FILE * restrict out_fp,
176 compress_func_t compress,
177 struct chunk_table * restrict chunk_tab)
179 const void *out_chunk;
180 unsigned out_chunk_size;
182 void *compressed_chunk = alloca(chunk_size);
184 out_chunk_size = (*compress)(chunk, chunk_size, compressed_chunk);
185 if (out_chunk_size) {
186 /* Write compressed */
187 out_chunk = compressed_chunk;
189 /* Write uncompressed */
191 out_chunk_size = chunk_size;
193 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
194 chunk_tab->cur_offset += out_chunk_size;
196 /* Write uncompressed */
198 out_chunk_size = chunk_size;
200 if (fwrite(out_chunk, 1, out_chunk_size, out_fp) != out_chunk_size) {
201 ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
202 return WIMLIB_ERR_WRITE;
208 * Finishes a WIM chunk table and writes it to the output file at the correct
211 * The final size of the full compressed resource is returned in the
212 * @compressed_size_p.
215 finish_wim_resource_chunk_tab(struct chunk_table * restrict chunk_tab,
216 FILE * restrict out_fp,
217 u64 * restrict compressed_size_p)
219 size_t bytes_written;
220 if (fseeko(out_fp, chunk_tab->file_offset, SEEK_SET) != 0) {
221 ERROR_WITH_ERRNO("Failed to seek to byte %"PRIu64" of output "
222 "WIM file", chunk_tab->file_offset);
223 return WIMLIB_ERR_WRITE;
226 if (chunk_tab->bytes_per_chunk_entry == 8) {
227 array_cpu_to_le64(chunk_tab->offsets, chunk_tab->num_chunks);
229 for (u64 i = 0; i < chunk_tab->num_chunks; i++)
230 ((u32*)chunk_tab->offsets)[i] =
231 cpu_to_le32(chunk_tab->offsets[i]);
233 bytes_written = fwrite((u8*)chunk_tab->offsets +
234 chunk_tab->bytes_per_chunk_entry,
235 1, chunk_tab->table_disk_size, out_fp);
236 if (bytes_written != chunk_tab->table_disk_size) {
237 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
239 return WIMLIB_ERR_WRITE;
241 if (fseeko(out_fp, 0, SEEK_END) != 0) {
242 ERROR_WITH_ERRNO("Failed to seek to end of output WIM file");
243 return WIMLIB_ERR_WRITE;
245 *compressed_size_p = chunk_tab->cur_offset + chunk_tab->table_disk_size;
250 fflush_and_ftruncate(FILE *out_fp, off_t offset)
252 if (fseeko(out_fp, offset, SEEK_SET) ||
254 ftruncate(fileno(out_fp), offset))
256 ERROR_WITH_ERRNO("Failed to flush and/or truncate "
258 return WIMLIB_ERR_WRITE;
265 finalize_and_check_sha1(SHA_CTX * restrict sha_ctx,
266 struct wim_lookup_table_entry * restrict lte)
268 u8 md[SHA1_HASH_SIZE];
269 sha1_final(md, sha_ctx);
271 copy_hash(lte->hash, md);
272 } else if (!hashes_equal(md, lte->hash)) {
273 ERROR("WIM resource has incorrect hash!");
274 if (lte_filename_valid(lte)) {
275 ERROR("We were reading it from \"%"TS"\"; maybe "
276 "it changed while we were reading it.",
279 return WIMLIB_ERR_INVALID_RESOURCE_HASH;
285 struct write_resource_ctx {
286 compress_func_t compress;
287 struct chunk_table *chunk_tab;
294 write_resource_cb(const void *restrict chunk, size_t chunk_size,
297 struct write_resource_ctx *ctx = _ctx;
300 sha1_update(&ctx->sha_ctx, chunk, chunk_size);
301 return write_wim_resource_chunk(chunk, chunk_size,
302 ctx->out_fp, ctx->compress,
307 * Write a resource to an output WIM.
309 * @lte: Lookup table entry for the resource, which could be in another WIM,
310 * in an external file, or in another location.
312 * @out_fp: FILE * opened to the output WIM.
314 * @out_ctype: One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate
315 * which compression algorithm to use.
317 * @out_res_entry: On success, this is filled in with the offset, flags,
318 * compressed size, and uncompressed size of the resource
321 * @flags: WIMLIB_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed
322 * even if it could otherwise be copied directly from the input.
324 * Additional notes: The SHA1 message digest of the uncompressed data is
325 * calculated (except when doing a raw copy --- see below). If the @unhashed
326 * flag is set on the lookup table entry, this message digest is simply copied
327 * to it; otherwise, the message digest is compared with the existing one, and
328 * the function will fail if they do not match.
331 write_wim_resource(struct wim_lookup_table_entry *lte,
332 FILE *out_fp, int out_ctype,
333 struct resource_entry *out_res_entry,
336 struct write_resource_ctx write_ctx;
342 flags &= ~WIMLIB_RESOURCE_FLAG_RECOMPRESS;
344 /* Get current position in output WIM */
345 offset = ftello(out_fp);
347 ERROR_WITH_ERRNO("Can't get position in output WIM");
348 return WIMLIB_ERR_WRITE;
351 /* If we are not forcing the data to be recompressed, and the input
352 * resource is located in a WIM with the same compression type as that
353 * desired other than no compression, we can simply copy the compressed
354 * data without recompressing it. This also means we must skip
355 * calculating the SHA1, as we never will see the uncompressed data. */
356 if (!(flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
357 lte->resource_location == RESOURCE_IN_WIM &&
358 out_ctype != WIMLIB_COMPRESSION_TYPE_NONE &&
359 wimlib_get_compression_type(lte->wim) == out_ctype)
361 flags |= WIMLIB_RESOURCE_FLAG_RAW;
362 write_ctx.doing_sha = false;
363 read_size = lte->resource_entry.size;
365 write_ctx.doing_sha = true;
366 sha1_init(&write_ctx.sha_ctx);
367 read_size = lte->resource_entry.original_size;
370 /* Initialize the chunk table and set the compression function if
371 * compressing the resource. */
372 if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
373 (flags & WIMLIB_RESOURCE_FLAG_RAW)) {
374 write_ctx.compress = NULL;
375 write_ctx.chunk_tab = NULL;
377 write_ctx.compress = get_compress_func(out_ctype);
378 ret = begin_wim_resource_chunk_tab(lte, out_fp,
380 &write_ctx.chunk_tab);
385 /* Write the entire resource by reading the entire resource and feeding
386 * the data through the write_resource_cb function. */
387 write_ctx.out_fp = out_fp;
389 ret = read_resource_prefix(lte, read_size,
390 write_resource_cb, &write_ctx, flags);
392 goto out_free_chunk_tab;
394 /* Verify SHA1 message digest of the resource, or set the hash for the
396 if (write_ctx.doing_sha) {
397 ret = finalize_and_check_sha1(&write_ctx.sha_ctx, lte);
399 goto out_free_chunk_tab;
402 out_res_entry->flags = lte->resource_entry.flags;
403 out_res_entry->original_size = wim_resource_size(lte);
404 out_res_entry->offset = offset;
405 if (flags & WIMLIB_RESOURCE_FLAG_RAW) {
406 /* Doing a raw write: The new compressed size is the same as
407 * the compressed size in the other WIM. */
408 new_size = lte->resource_entry.size;
409 } else if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE) {
410 /* Using WIMLIB_COMPRESSION_TYPE_NONE: The new compressed size
411 * is the original size. */
412 new_size = lte->resource_entry.original_size;
413 out_res_entry->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
415 /* Using a different compression type: Call
416 * finish_wim_resource_chunk_tab() and it will provide the new
417 * compressed size. */
418 ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab, out_fp,
421 goto out_free_chunk_tab;
422 if (new_size >= wim_resource_size(lte)) {
423 /* Oops! We compressed the resource to larger than the original
424 * size. Write the resource uncompressed instead. */
425 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
426 "writing uncompressed instead",
427 wim_resource_size(lte), new_size);
428 ret = fflush_and_ftruncate(out_fp, offset);
430 goto out_free_chunk_tab;
431 write_ctx.compress = NULL;
432 write_ctx.doing_sha = false;
433 out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
434 goto try_write_again;
436 out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
438 out_res_entry->size = new_size;
441 FREE(write_ctx.chunk_tab);
445 #ifdef ENABLE_MULTITHREADED_COMPRESSION
447 /* Blocking shared queue (solves the producer-consumer problem) */
448 struct shared_queue {
452 unsigned filled_slots;
454 pthread_mutex_t lock;
455 pthread_cond_t msg_avail_cond;
456 pthread_cond_t space_avail_cond;
460 shared_queue_init(struct shared_queue *q, unsigned size)
462 wimlib_assert(size != 0);
463 q->array = CALLOC(sizeof(q->array[0]), size);
470 if (pthread_mutex_init(&q->lock, NULL)) {
471 ERROR_WITH_ERRNO("Failed to initialize mutex");
474 if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
475 ERROR_WITH_ERRNO("Failed to initialize condition variable");
476 goto err_destroy_lock;
478 if (pthread_cond_init(&q->space_avail_cond, NULL)) {
479 ERROR_WITH_ERRNO("Failed to initialize condition variable");
480 goto err_destroy_msg_avail_cond;
483 err_destroy_msg_avail_cond:
484 pthread_cond_destroy(&q->msg_avail_cond);
486 pthread_mutex_destroy(&q->lock);
488 return WIMLIB_ERR_NOMEM;
492 shared_queue_destroy(struct shared_queue *q)
495 pthread_mutex_destroy(&q->lock);
496 pthread_cond_destroy(&q->msg_avail_cond);
497 pthread_cond_destroy(&q->space_avail_cond);
501 shared_queue_put(struct shared_queue *q, void *obj)
503 pthread_mutex_lock(&q->lock);
504 while (q->filled_slots == q->size)
505 pthread_cond_wait(&q->space_avail_cond, &q->lock);
507 q->back = (q->back + 1) % q->size;
508 q->array[q->back] = obj;
511 pthread_cond_broadcast(&q->msg_avail_cond);
512 pthread_mutex_unlock(&q->lock);
516 shared_queue_get(struct shared_queue *q)
520 pthread_mutex_lock(&q->lock);
521 while (q->filled_slots == 0)
522 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
524 obj = q->array[q->front];
525 q->array[q->front] = NULL;
526 q->front = (q->front + 1) % q->size;
529 pthread_cond_broadcast(&q->space_avail_cond);
530 pthread_mutex_unlock(&q->lock);
534 struct compressor_thread_params {
535 struct shared_queue *res_to_compress_queue;
536 struct shared_queue *compressed_res_queue;
537 compress_func_t compress;
540 #define MAX_CHUNKS_PER_MSG 2
543 struct wim_lookup_table_entry *lte;
544 u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
545 u8 *out_compressed_chunks[MAX_CHUNKS_PER_MSG];
546 u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
547 unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
548 unsigned compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
550 struct list_head list;
556 compress_chunks(struct message *msg, compress_func_t compress)
558 for (unsigned i = 0; i < msg->num_chunks; i++) {
559 unsigned len = compress(msg->uncompressed_chunks[i],
560 msg->uncompressed_chunk_sizes[i],
561 msg->compressed_chunks[i]);
563 /* To be written compressed */
564 msg->out_compressed_chunks[i] = msg->compressed_chunks[i];
565 msg->compressed_chunk_sizes[i] = len;
567 /* To be written uncompressed */
568 msg->out_compressed_chunks[i] = msg->uncompressed_chunks[i];
569 msg->compressed_chunk_sizes[i] = msg->uncompressed_chunk_sizes[i];
575 /* Compressor thread routine. This is a lot simpler than the main thread
576 * routine: just repeatedly get a group of chunks from the
577 * res_to_compress_queue, compress them, and put them in the
578 * compressed_res_queue. A NULL pointer indicates that the thread should stop.
581 compressor_thread_proc(void *arg)
583 struct compressor_thread_params *params = arg;
584 struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
585 struct shared_queue *compressed_res_queue = params->compressed_res_queue;
586 compress_func_t compress = params->compress;
589 DEBUG("Compressor thread ready");
590 while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
591 compress_chunks(msg, compress);
592 shared_queue_put(compressed_res_queue, msg);
594 DEBUG("Compressor thread terminating");
597 #endif /* ENABLE_MULTITHREADED_COMPRESSION */
600 do_write_streams_progress(union wimlib_progress_info *progress,
601 wimlib_progress_func_t progress_func,
604 progress->write_streams.completed_bytes += size_added;
605 progress->write_streams.completed_streams++;
607 progress->write_streams.completed_bytes >= progress->write_streams._private)
609 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
611 if (progress->write_streams._private == progress->write_streams.total_bytes) {
612 progress->write_streams._private = ~0;
614 progress->write_streams._private =
615 min(progress->write_streams.total_bytes,
616 progress->write_streams.completed_bytes +
617 progress->write_streams.total_bytes / 100);
622 struct serial_write_stream_ctx {
625 int write_resource_flags;
629 serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
631 struct serial_write_stream_ctx *ctx = _ctx;
632 return write_wim_resource(lte, ctx->out_fp,
633 ctx->out_ctype, <e->output_resource_entry,
634 ctx->write_resource_flags);
637 /* Write a list of streams, taking into account that some streams may be
638 * duplicates that are checksummed and discarded on the fly, and also delegating
639 * the actual writing of a stream to a function @write_stream_cb, which is
640 * passed the context @write_stream_ctx. */
642 do_write_stream_list(struct list_head *stream_list,
643 struct wim_lookup_table *lookup_table,
644 int (*write_stream_cb)(struct wim_lookup_table_entry *, void *),
645 void *write_stream_ctx,
646 wimlib_progress_func_t progress_func,
647 union wimlib_progress_info *progress)
650 struct wim_lookup_table_entry *lte;
652 /* For each stream in @stream_list ... */
653 while (!list_empty(stream_list)) {
654 lte = container_of(stream_list->next,
655 struct wim_lookup_table_entry,
657 list_del(<e->write_streams_list);
658 if (lte->unhashed && !lte->unique_size) {
659 /* Unhashed stream that shares a size with some other
660 * stream in the WIM we are writing. The stream must be
661 * checksummed to know if we need to write it or not. */
662 struct wim_lookup_table_entry *tmp;
663 u32 orig_refcnt = lte->out_refcnt;
665 ret = hash_unhashed_stream(lte, lookup_table, &tmp);
670 /* We found a duplicate stream. */
671 if (orig_refcnt != tmp->out_refcnt) {
672 /* We have already written, or are going
673 * to write, the duplicate stream. So
674 * just skip to the next stream. */
675 DEBUG("Discarding duplicate stream of length %"PRIu64,
676 wim_resource_size(lte));
677 lte->no_progress = 0;
678 goto skip_to_progress;
683 /* Here, @lte is either a hashed stream or an unhashed stream
684 * with a unique size. In either case we know that the stream
685 * has to be written. In either case the SHA1 message digest
686 * will be calculated over the stream while writing it; however,
687 * in the former case this is done merely to check the data,
688 * while in the latter case this is done because we do not have
689 * the SHA1 message digest yet. */
690 wimlib_assert(lte->out_refcnt != 0);
692 lte->no_progress = 0;
693 ret = (*write_stream_cb)(lte, write_stream_ctx);
696 /* In parallel mode, some streams are deferred for later,
697 * serialized processing; ignore them here. */
701 list_del(<e->unhashed_list);
702 lookup_table_insert(lookup_table, lte);
706 if (!lte->no_progress) {
707 do_write_streams_progress(progress,
709 wim_resource_size(lte));
716 do_write_stream_list_serial(struct list_head *stream_list,
717 struct wim_lookup_table *lookup_table,
720 int write_resource_flags,
721 wimlib_progress_func_t progress_func,
722 union wimlib_progress_info *progress)
724 struct serial_write_stream_ctx ctx = {
726 .out_ctype = out_ctype,
727 .write_resource_flags = write_resource_flags,
729 return do_write_stream_list(stream_list,
738 write_flags_to_resource_flags(int write_flags)
740 int resource_flags = 0;
742 if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
743 resource_flags |= WIMLIB_RESOURCE_FLAG_RECOMPRESS;
744 return resource_flags;
748 write_stream_list_serial(struct list_head *stream_list,
749 struct wim_lookup_table *lookup_table,
752 int write_resource_flags,
753 wimlib_progress_func_t progress_func,
754 union wimlib_progress_info *progress)
756 DEBUG("Writing stream list (serial version)");
757 progress->write_streams.num_threads = 1;
759 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
760 return do_write_stream_list_serial(stream_list,
764 write_resource_flags,
769 #ifdef ENABLE_MULTITHREADED_COMPRESSION
771 write_wim_chunks(struct message *msg, FILE *out_fp,
772 struct chunk_table *chunk_tab)
774 for (unsigned i = 0; i < msg->num_chunks; i++) {
775 unsigned chunk_csize = msg->compressed_chunk_sizes[i];
777 if (fwrite(msg->out_compressed_chunks[i], 1, chunk_csize, out_fp)
780 ERROR_WITH_ERRNO("Failed to write WIM chunk");
781 return WIMLIB_ERR_WRITE;
784 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
785 chunk_tab->cur_offset += chunk_csize;
790 struct main_writer_thread_ctx {
791 struct list_head *stream_list;
792 struct wim_lookup_table *lookup_table;
795 int write_resource_flags;
796 struct shared_queue *res_to_compress_queue;
797 struct shared_queue *compressed_res_queue;
799 wimlib_progress_func_t progress_func;
800 union wimlib_progress_info *progress;
802 struct list_head available_msgs;
803 struct list_head outstanding_streams;
804 struct list_head serial_streams;
805 size_t num_outstanding_messages;
807 SHA_CTX next_sha_ctx;
810 struct wim_lookup_table_entry *next_lte;
812 struct message *msgs;
813 struct message *next_msg;
814 struct chunk_table *cur_chunk_tab;
818 init_message(struct message *msg)
820 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
821 msg->compressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
822 msg->uncompressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
823 if (msg->compressed_chunks[i] == NULL ||
824 msg->uncompressed_chunks[i] == NULL)
825 return WIMLIB_ERR_NOMEM;
831 destroy_message(struct message *msg)
833 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
834 FREE(msg->compressed_chunks[i]);
835 FREE(msg->uncompressed_chunks[i]);
840 free_messages(struct message *msgs, size_t num_messages)
843 for (size_t i = 0; i < num_messages; i++)
844 destroy_message(&msgs[i]);
849 static struct message *
850 allocate_messages(size_t num_messages)
852 struct message *msgs;
854 msgs = CALLOC(num_messages, sizeof(struct message));
857 for (size_t i = 0; i < num_messages; i++) {
858 if (init_message(&msgs[i])) {
859 free_messages(msgs, num_messages);
867 main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
869 while (ctx->num_outstanding_messages--)
870 shared_queue_get(ctx->compressed_res_queue);
871 free_messages(ctx->msgs, ctx->num_messages);
872 FREE(ctx->cur_chunk_tab);
876 main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
878 /* Pre-allocate all the buffers that will be needed to do the chunk
880 ctx->msgs = allocate_messages(ctx->num_messages);
882 return WIMLIB_ERR_NOMEM;
884 /* Initially, all the messages are available to use. */
885 INIT_LIST_HEAD(&ctx->available_msgs);
886 for (size_t i = 0; i < ctx->num_messages; i++)
887 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
889 /* outstanding_streams is the list of streams that currently have had
890 * chunks sent off for compression.
892 * The first stream in outstanding_streams is the stream that is
893 * currently being written.
895 * The last stream in outstanding_streams is the stream that is
896 * currently being read and having chunks fed to the compressor threads.
898 INIT_LIST_HEAD(&ctx->outstanding_streams);
899 ctx->num_outstanding_messages = 0;
901 ctx->next_msg = NULL;
903 /* Resources that don't need any chunks compressed are added to this
904 * list and written directly by the main thread. */
905 INIT_LIST_HEAD(&ctx->serial_streams);
907 ctx->cur_chunk_tab = NULL;
913 receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
916 struct wim_lookup_table_entry *cur_lte;
919 wimlib_assert(!list_empty(&ctx->outstanding_streams));
920 wimlib_assert(ctx->num_outstanding_messages != 0);
922 cur_lte = container_of(ctx->outstanding_streams.next,
923 struct wim_lookup_table_entry,
924 being_compressed_list);
926 /* Get the next message from the queue and process it.
927 * The message will contain 1 or more data chunks that have been
929 msg = shared_queue_get(ctx->compressed_res_queue);
930 msg->complete = true;
931 --ctx->num_outstanding_messages;
933 /* Is this the next chunk in the current resource? If it's not
934 * (i.e., an earlier chunk in a same or different resource
935 * hasn't been compressed yet), do nothing, and keep this
936 * message around until all earlier chunks are received.
938 * Otherwise, write all the chunks we can. */
939 while (cur_lte != NULL &&
940 !list_empty(&cur_lte->msg_list)
941 && (msg = container_of(cur_lte->msg_list.next,
945 list_move(&msg->list, &ctx->available_msgs);
946 if (msg->begin_chunk == 0) {
947 /* This is the first set of chunks. Leave space
948 * for the chunk table in the output file. */
949 off_t cur_offset = ftello(ctx->out_fp);
950 if (cur_offset == -1)
951 return WIMLIB_ERR_WRITE;
952 ret = begin_wim_resource_chunk_tab(cur_lte,
955 &ctx->cur_chunk_tab);
960 /* Write the compressed chunks from the message. */
961 ret = write_wim_chunks(msg, ctx->out_fp, ctx->cur_chunk_tab);
965 /* Was this the last chunk of the stream? If so, finish
967 if (list_empty(&cur_lte->msg_list) &&
968 msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
973 ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
979 list_del(&cur_lte->being_compressed_list);
981 /* Grab the offset of this stream in the output file
982 * from the chunk table before we free it. */
983 offset = ctx->cur_chunk_tab->file_offset;
985 FREE(ctx->cur_chunk_tab);
986 ctx->cur_chunk_tab = NULL;
988 if (res_csize >= wim_resource_size(cur_lte)) {
989 /* Oops! We compressed the resource to
990 * larger than the original size. Write
991 * the resource uncompressed instead. */
992 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
993 "writing uncompressed instead",
994 wim_resource_size(cur_lte), res_csize);
995 ret = fflush_and_ftruncate(ctx->out_fp, offset);
998 ret = write_wim_resource(cur_lte,
1000 WIMLIB_COMPRESSION_TYPE_NONE,
1001 &cur_lte->output_resource_entry,
1002 ctx->write_resource_flags);
1006 cur_lte->output_resource_entry.size =
1009 cur_lte->output_resource_entry.original_size =
1010 cur_lte->resource_entry.original_size;
1012 cur_lte->output_resource_entry.offset =
1015 cur_lte->output_resource_entry.flags =
1016 cur_lte->resource_entry.flags |
1017 WIM_RESHDR_FLAG_COMPRESSED;
1020 do_write_streams_progress(ctx->progress,
1022 wim_resource_size(cur_lte));
1024 /* Since we just finished writing a stream, write any
1025 * streams that have been added to the serial_streams
1026 * list for direct writing by the main thread (e.g.
1027 * resources that don't need to be compressed because
1028 * the desired compression type is the same as the
1029 * previous compression type). */
1030 if (!list_empty(&ctx->serial_streams)) {
1031 ret = do_write_stream_list_serial(&ctx->serial_streams,
1035 ctx->write_resource_flags,
1042 /* Advance to the next stream to write. */
1043 if (list_empty(&ctx->outstanding_streams)) {
1046 cur_lte = container_of(ctx->outstanding_streams.next,
1047 struct wim_lookup_table_entry,
1048 being_compressed_list);
1055 /* Called when the main thread has read a new chunk of data. */
1057 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
1059 struct main_writer_thread_ctx *ctx = _ctx;
1061 struct message *next_msg;
1062 u64 next_chunk_in_msg;
1064 /* Update SHA1 message digest for the stream currently being read by the
1066 sha1_update(&ctx->next_sha_ctx, chunk, chunk_size);
1068 /* We send chunks of data to the compressor chunks in batches which we
1069 * refer to as "messages". @next_msg is the message that is currently
1070 * being prepared to send off. If it is NULL, that indicates that we
1071 * need to start a new message. */
1072 next_msg = ctx->next_msg;
1074 /* We need to start a new message. First check to see if there
1075 * is a message available in the list of available messages. If
1076 * so, we can just take one. If not, all the messages (there is
1077 * a fixed number of them, proportional to the number of
1078 * threads) have been sent off to the compressor threads, so we
1079 * receive messages from the compressor threads containing
1080 * compressed chunks of data.
1082 * We may need to receive multiple messages before one is
1083 * actually available to use because messages received that are
1084 * *not* for the very next set of chunks to compress must be
1085 * buffered until it's time to write those chunks. */
1086 while (list_empty(&ctx->available_msgs)) {
1087 ret = receive_compressed_chunks(ctx);
1092 next_msg = container_of(ctx->available_msgs.next,
1093 struct message, list);
1094 list_del(&next_msg->list);
1095 next_msg->complete = false;
1096 next_msg->begin_chunk = ctx->next_chunk;
1097 next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
1098 ctx->next_num_chunks - ctx->next_chunk);
1099 ctx->next_msg = next_msg;
1102 /* Fill in the next chunk to compress */
1103 next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;
1105 next_msg->uncompressed_chunk_sizes[next_chunk_in_msg] = chunk_size;
1106 memcpy(next_msg->uncompressed_chunks[next_chunk_in_msg],
1109 if (++next_chunk_in_msg == next_msg->num_chunks) {
1110 /* Send off an array of chunks to compress */
1111 list_add_tail(&next_msg->list, &ctx->next_lte->msg_list);
1112 shared_queue_put(ctx->res_to_compress_queue, next_msg);
1113 ++ctx->num_outstanding_messages;
1114 ctx->next_msg = NULL;
1120 main_writer_thread_finish(void *_ctx)
1122 struct main_writer_thread_ctx *ctx = _ctx;
1124 while (ctx->num_outstanding_messages != 0) {
1125 ret = receive_compressed_chunks(ctx);
1129 wimlib_assert(list_empty(&ctx->outstanding_streams));
1130 return do_write_stream_list_serial(&ctx->serial_streams,
1134 ctx->write_resource_flags,
1140 submit_stream_for_compression(struct wim_lookup_table_entry *lte,
1141 struct main_writer_thread_ctx *ctx)
1145 /* Read the entire stream @lte, feeding its data chunks to the
1146 * compressor threads. Also SHA1-sum the stream; this is required in
1147 * the case that @lte is unhashed, and a nice additional verification
1148 * when @lte is already hashed. */
1149 sha1_init(&ctx->next_sha_ctx);
1150 ctx->next_chunk = 0;
1151 ctx->next_num_chunks = wim_resource_chunks(lte);
1152 ctx->next_lte = lte;
1153 INIT_LIST_HEAD(<e->msg_list);
1154 list_add_tail(<e->being_compressed_list, &ctx->outstanding_streams);
1155 ret = read_resource_prefix(lte, wim_resource_size(lte),
1156 main_writer_thread_cb, ctx, 0);
1158 wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
1159 ret = finalize_and_check_sha1(&ctx->next_sha_ctx, lte);
1165 main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
1167 struct main_writer_thread_ctx *ctx = _ctx;
1170 if (wim_resource_size(lte) < 1000 ||
1171 ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
1172 (lte->resource_location == RESOURCE_IN_WIM &&
1173 !(ctx->write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
1174 wimlib_get_compression_type(lte->wim) == ctx->out_ctype))
1176 /* Stream is too small or isn't being compressed. Process it by
1177 * the main thread when we have a chance. We can't necessarily
1178 * process it right here, as the main thread could be in the
1179 * middle of writing a different stream. */
1180 list_add_tail(<e->write_streams_list, &ctx->serial_streams);
1184 ret = submit_stream_for_compression(lte, ctx);
1186 lte->no_progress = 1;
1191 get_default_num_threads()
1194 return win32_get_number_of_processors();
1196 return sysconf(_SC_NPROCESSORS_ONLN);
1200 /* Equivalent to write_stream_list_serial(), except this takes a @num_threads
1201 * parameter and will perform compression using that many threads. Falls
1202 * back to write_stream_list_serial() on certain errors, such as a failure to
1203 * create the number of threads requested.
1205 * High level description of the algorithm for writing compressed streams in
1206 * parallel: We perform compression on chunks of size WIM_CHUNK_SIZE bytes
1207 * rather than on full files. The currently executing thread becomes the main
1208 * thread and is entirely in charge of reading the data to compress (which may
1209 * be in any location understood by the resource code--- such as in an external
1210 * file being captured, or in another WIM file from which an image is being
1211 * exported) and actually writing the compressed data to the output file.
1212 * Additional threads are "compressor threads" and all execute the
1213 * compressor_thread_proc, where they repeatedly retrieve buffers of data from
1214 * the main thread, compress them, and hand them back to the main thread.
1216 * Certain streams, such as streams that do not need to be compressed (e.g.
1217 * input compression type same as output compression type) or streams of very
1218 * small size are placed in a list (main_writer_thread_ctx.serial_list) and
1219 * handled entirely by the main thread at an appropriate time.
1221 * At any given point in time, multiple streams may be having chunks compressed
1222 * concurrently. The stream that the main thread is currently *reading* may be
1223 * later in the list that the stream that the main thread is currently
1227 write_stream_list_parallel(struct list_head *stream_list,
1228 struct wim_lookup_table *lookup_table,
1231 int write_resource_flags,
1232 wimlib_progress_func_t progress_func,
1233 union wimlib_progress_info *progress,
1234 unsigned num_threads)
1237 struct shared_queue res_to_compress_queue;
1238 struct shared_queue compressed_res_queue;
1239 pthread_t *compressor_threads = NULL;
1241 if (num_threads == 0) {
1242 long nthreads = get_default_num_threads();
1243 if (nthreads < 1 || nthreads > UINT_MAX) {
1244 WARNING("Could not determine number of processors! Assuming 1");
1246 } else if (nthreads == 1) {
1247 goto out_serial_quiet;
1249 num_threads = nthreads;
1253 DEBUG("Writing stream list (parallel version, num_threads=%u)",
1256 progress->write_streams.num_threads = num_threads;
1258 static const size_t MESSAGES_PER_THREAD = 2;
1259 size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1261 DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1263 ret = shared_queue_init(&res_to_compress_queue, queue_size);
1267 ret = shared_queue_init(&compressed_res_queue, queue_size);
1269 goto out_destroy_res_to_compress_queue;
1271 struct compressor_thread_params params;
1272 params.res_to_compress_queue = &res_to_compress_queue;
1273 params.compressed_res_queue = &compressed_res_queue;
1274 params.compress = get_compress_func(out_ctype);
1276 compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1277 if (!compressor_threads) {
1278 ret = WIMLIB_ERR_NOMEM;
1279 goto out_destroy_compressed_res_queue;
1282 for (unsigned i = 0; i < num_threads; i++) {
1283 DEBUG("pthread_create thread %u of %u", i + 1, num_threads);
1284 ret = pthread_create(&compressor_threads[i], NULL,
1285 compressor_thread_proc, ¶ms);
1288 ERROR_WITH_ERRNO("Failed to create compressor "
1290 i + 1, num_threads);
1297 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
1299 struct main_writer_thread_ctx ctx;
1300 ctx.stream_list = stream_list;
1301 ctx.lookup_table = lookup_table;
1302 ctx.out_fp = out_fp;
1303 ctx.out_ctype = out_ctype;
1304 ctx.res_to_compress_queue = &res_to_compress_queue;
1305 ctx.compressed_res_queue = &compressed_res_queue;
1306 ctx.num_messages = queue_size;
1307 ctx.write_resource_flags = write_resource_flags | WIMLIB_RESOURCE_FLAG_THREADSAFE_READ;
1308 ctx.progress_func = progress_func;
1309 ctx.progress = progress;
1310 ret = main_writer_thread_init_ctx(&ctx);
1313 ret = do_write_stream_list(stream_list, lookup_table,
1314 main_thread_process_next_stream,
1315 &ctx, progress_func, progress);
1317 goto out_destroy_ctx;
1319 /* The main thread has finished reading all streams that are going to be
1320 * compressed in parallel, and it now needs to wait for all remaining
1321 * chunks to be compressed so that the remaining streams can actually be
1322 * written to the output file. Furthermore, any remaining streams that
1323 * had processing deferred to the main thread need to be handled. These
1324 * tasks are done by the main_writer_thread_finish() function. */
1325 ret = main_writer_thread_finish(&ctx);
1327 main_writer_thread_destroy_ctx(&ctx);
1329 for (unsigned i = 0; i < num_threads; i++)
1330 shared_queue_put(&res_to_compress_queue, NULL);
1332 for (unsigned i = 0; i < num_threads; i++) {
1333 if (pthread_join(compressor_threads[i], NULL)) {
1334 WARNING_WITH_ERRNO("Failed to join compressor "
1336 i + 1, num_threads);
1339 FREE(compressor_threads);
1340 out_destroy_compressed_res_queue:
1341 shared_queue_destroy(&compressed_res_queue);
1342 out_destroy_res_to_compress_queue:
1343 shared_queue_destroy(&res_to_compress_queue);
1344 if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
1347 WARNING("Falling back to single-threaded compression");
1349 return write_stream_list_serial(stream_list,
1353 write_resource_flags,
1361 * Write a list of streams to a WIM (@out_fp) using the compression type
1362 * @out_ctype and up to @num_threads compressor threads.
1365 write_stream_list(struct list_head *stream_list,
1366 struct wim_lookup_table *lookup_table,
1367 FILE *out_fp, int out_ctype, int write_flags,
1368 unsigned num_threads, wimlib_progress_func_t progress_func)
1370 struct wim_lookup_table_entry *lte;
1371 size_t num_streams = 0;
1372 u64 total_bytes = 0;
1373 u64 total_compression_bytes = 0;
1374 union wimlib_progress_info progress;
1376 int write_resource_flags;
1378 if (list_empty(stream_list))
1381 write_resource_flags = write_flags_to_resource_flags(write_flags);
1383 /* Calculate the total size of the streams to be written. Note: this
1384 * will be the uncompressed size, as we may not know the compressed size
1385 * yet, and also this will assume that every unhashed stream will be
1386 * written (which will not necessarily be the case). */
1387 list_for_each_entry(lte, stream_list, write_streams_list) {
1389 total_bytes += wim_resource_size(lte);
1390 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
1391 && (wim_resource_compression_type(lte) != out_ctype ||
1392 (write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS)))
1394 total_compression_bytes += wim_resource_size(lte);
1397 progress.write_streams.total_bytes = total_bytes;
1398 progress.write_streams.total_streams = num_streams;
1399 progress.write_streams.completed_bytes = 0;
1400 progress.write_streams.completed_streams = 0;
1401 progress.write_streams.num_threads = num_threads;
1402 progress.write_streams.compression_type = out_ctype;
1403 progress.write_streams._private = 0;
1405 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1406 if (total_compression_bytes >= 1000000 && num_threads != 1)
1407 ret = write_stream_list_parallel(stream_list,
1411 write_resource_flags,
1417 ret = write_stream_list_serial(stream_list,
1421 write_resource_flags,
1427 struct stream_size_table {
1428 struct hlist_head *array;
1434 init_stream_size_table(struct stream_size_table *tab, size_t capacity)
1436 tab->array = CALLOC(capacity, sizeof(tab->array[0]));
1438 return WIMLIB_ERR_NOMEM;
1439 tab->num_entries = 0;
1440 tab->capacity = capacity;
1445 destroy_stream_size_table(struct stream_size_table *tab)
1451 stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
1453 struct stream_size_table *tab = _tab;
1455 struct wim_lookup_table_entry *same_size_lte;
1456 struct hlist_node *tmp;
1458 pos = hash_u64(wim_resource_size(lte)) % tab->capacity;
1459 lte->unique_size = 1;
1460 hlist_for_each_entry(same_size_lte, tmp, &tab->array[pos], hash_list_2) {
1461 if (wim_resource_size(same_size_lte) == wim_resource_size(lte)) {
1462 lte->unique_size = 0;
1463 same_size_lte->unique_size = 0;
1468 hlist_add_head(<e->hash_list_2, &tab->array[pos]);
1474 struct lte_overwrite_prepare_args {
1477 struct list_head stream_list;
1478 struct stream_size_table stream_size_tab;
1481 /* First phase of preparing streams for an in-place overwrite. This is called
1482 * on all streams, both hashed and unhashed, except the metadata resources. */
1484 lte_overwrite_prepare(struct wim_lookup_table_entry *lte, void *_args)
1486 struct lte_overwrite_prepare_args *args = _args;
1488 wimlib_assert(!(lte->resource_entry.flags & WIM_RESHDR_FLAG_METADATA));
1489 if (lte->resource_location != RESOURCE_IN_WIM || lte->wim != args->wim)
1490 list_add_tail(<e->write_streams_list, &args->stream_list);
1491 lte->out_refcnt = lte->refcnt;
1492 stream_size_table_insert(lte, &args->stream_size_tab);
1496 /* Second phase of preparing streams for an in-place overwrite. This is called
1497 * on existing metadata resources and hashed streams, but not unhashed streams.
1499 * NOTE: lte->output_resource_entry is in union with lte->hash_list_2, so
1500 * lte_overwrite_prepare_2() must be called after lte_overwrite_prepare(), as
1501 * the latter uses lte->hash_list_2, while the former expects to set
1502 * lte->output_resource_entry. */
1504 lte_overwrite_prepare_2(struct wim_lookup_table_entry *lte, void *_args)
1506 struct lte_overwrite_prepare_args *args = _args;
1508 if (lte->resource_location == RESOURCE_IN_WIM && lte->wim == args->wim) {
1509 /* We can't do an in place overwrite on the WIM if there are
1510 * streams after the XML data. */
1511 if (lte->resource_entry.offset +
1512 lte->resource_entry.size > args->end_offset)
1514 #ifdef ENABLE_ERROR_MESSAGES
1515 ERROR("The following resource is after the XML data:");
1516 print_lookup_table_entry(lte, stderr);
1518 return WIMLIB_ERR_RESOURCE_ORDER;
1520 copy_resource_entry(<e->output_resource_entry,
1521 <e->resource_entry);
1526 /* Given a WIM that we are going to overwrite in place with zero or more
1527 * additional streams added, construct a list the list of new unique streams
1528 * ('struct wim_lookup_table_entry's) that must be written, plus any unhashed
1529 * streams that need to be added but may be identical to other hashed or
1530 * unhashed streams. These unhashed streams are checksummed while the streams
1531 * are being written. To aid this process, the member @unique_size is set to 1
1532 * on streams that have a unique size and therefore must be written.
1534 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1535 * indicate the number of times the stream is referenced in only the streams
1536 * that are being written; this may still be adjusted later when unhashed
1537 * streams are being resolved.
1540 prepare_streams_for_overwrite(WIMStruct *wim, off_t end_offset,
1541 struct list_head *stream_list)
1544 struct lte_overwrite_prepare_args args;
1548 args.end_offset = end_offset;
1549 ret = init_stream_size_table(&args.stream_size_tab,
1550 wim->lookup_table->capacity);
1554 INIT_LIST_HEAD(&args.stream_list);
1555 for (i = 0; i < wim->hdr.image_count; i++) {
1556 struct wim_image_metadata *imd;
1557 struct wim_lookup_table_entry *lte;
1559 imd = wim->image_metadata[i];
1560 image_for_each_unhashed_stream(lte, imd)
1561 lte_overwrite_prepare(lte, &args);
1563 for_lookup_table_entry(wim->lookup_table, lte_overwrite_prepare, &args);
1564 list_transfer(&args.stream_list, stream_list);
1566 for (i = 0; i < wim->hdr.image_count; i++) {
1567 ret = lte_overwrite_prepare_2(wim->image_metadata[i]->metadata_lte,
1570 goto out_destroy_stream_size_table;
1572 ret = for_lookup_table_entry(wim->lookup_table,
1573 lte_overwrite_prepare_2, &args);
1574 out_destroy_stream_size_table:
1575 destroy_stream_size_table(&args.stream_size_tab);
1580 struct find_streams_ctx {
1581 struct list_head stream_list;
1582 struct stream_size_table stream_size_tab;
1586 inode_find_streams_to_write(struct wim_inode *inode,
1587 struct wim_lookup_table *table,
1588 struct list_head *stream_list,
1589 struct stream_size_table *tab)
1591 struct wim_lookup_table_entry *lte;
1592 for (unsigned i = 0; i <= inode->i_num_ads; i++) {
1593 lte = inode_stream_lte(inode, i, table);
1595 if (lte->out_refcnt == 0) {
1597 stream_size_table_insert(lte, tab);
1598 list_add_tail(<e->write_streams_list, stream_list);
1600 lte->out_refcnt += inode->i_nlink;
1606 image_find_streams_to_write(WIMStruct *w)
1608 struct find_streams_ctx *ctx;
1609 struct wim_image_metadata *imd;
1610 struct wim_inode *inode;
1611 struct wim_lookup_table_entry *lte;
1614 imd = wim_get_current_image_metadata(w);
1616 image_for_each_unhashed_stream(lte, imd)
1617 lte->out_refcnt = 0;
1619 /* Go through this image's inodes to find any streams that have not been
1621 image_for_each_inode(inode, imd) {
1622 inode_find_streams_to_write(inode, w->lookup_table,
1624 &ctx->stream_size_tab);
1629 /* Given a WIM that from which one or all of the images is being written, build
1630 * the list of unique streams ('struct wim_lookup_table_entry's) that must be
1631 * written, plus any unhashed streams that need to be written but may be
1632 * identical to other hashed or unhashed streams being written. These unhashed
1633 * streams are checksummed while the streams are being written. To aid this
1634 * process, the member @unique_size is set to 1 on streams that have a unique
1635 * size and therefore must be written.
1637 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1638 * indicate the number of times the stream is referenced in only the streams
1639 * that are being written; this may still be adjusted later when unhashed
1640 * streams are being resolved.
1643 prepare_stream_list(WIMStruct *wim, int image, struct list_head *stream_list)
1646 struct find_streams_ctx ctx;
1648 for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
1649 ret = init_stream_size_table(&ctx.stream_size_tab,
1650 wim->lookup_table->capacity);
1653 for_lookup_table_entry(wim->lookup_table, stream_size_table_insert,
1654 &ctx.stream_size_tab);
1655 INIT_LIST_HEAD(&ctx.stream_list);
1656 wim->private = &ctx;
1657 ret = for_image(wim, image, image_find_streams_to_write);
1658 destroy_stream_size_table(&ctx.stream_size_tab);
1660 list_transfer(&ctx.stream_list, stream_list);
1664 /* Writes the streams for the specified @image in @wim to @wim->out_fp.
1667 write_wim_streams(WIMStruct *wim, int image, int write_flags,
1668 unsigned num_threads,
1669 wimlib_progress_func_t progress_func)
1672 struct list_head stream_list;
1674 ret = prepare_stream_list(wim, image, &stream_list);
1677 return write_stream_list(&stream_list,
1680 wimlib_get_compression_type(wim),
1687 * Finish writing a WIM file: write the lookup table, xml data, and integrity
1688 * table (optional), then overwrite the WIM header.
1690 * write_flags is a bitwise OR of the following:
1692 * (public) WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
1693 * Include an integrity table.
1695 * (public) WIMLIB_WRITE_FLAG_SHOW_PROGRESS:
1696 * Show progress information when (if) writing the integrity table.
1698 * (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
1699 * Don't write the lookup table.
1701 * (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
1702 * When (if) writing the integrity table, re-use entries from the
1703 * existing integrity table, if possible.
1705 * (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
1706 * After writing the XML data but before writing the integrity
1707 * table, write a temporary WIM header and flush the stream so that
1708 * the WIM is less likely to become corrupted upon abrupt program
1711 * (private) WIMLIB_WRITE_FLAG_FSYNC:
1712 * fsync() the output file before closing it.
1716 finish_write(WIMStruct *w, int image, int write_flags,
1717 wimlib_progress_func_t progress_func)
1720 struct wim_header hdr;
1721 FILE *out = w->out_fp;
1723 /* @hdr will be the header for the new WIM. First copy all the data
1724 * from the header in the WIMStruct; then set all the fields that may
1725 * have changed, including the resource entries, boot index, and image
1727 memcpy(&hdr, &w->hdr, sizeof(struct wim_header));
1729 /* Set image count and boot index correctly for single image writes */
1730 if (image != WIMLIB_ALL_IMAGES) {
1731 hdr.image_count = 1;
1732 if (hdr.boot_idx == image)
1738 /* In the WIM header, there is room for the resource entry for a
1739 * metadata resource labeled as the "boot metadata". This entry should
1740 * be zeroed out if there is no bootable image (boot_idx 0). Otherwise,
1741 * it should be a copy of the resource entry for the image that is
1742 * marked as bootable. This is not well documented... */
1743 if (hdr.boot_idx == 0) {
1744 zero_resource_entry(&hdr.boot_metadata_res_entry);
1746 copy_resource_entry(&hdr.boot_metadata_res_entry,
1747 &w->image_metadata[ hdr.boot_idx- 1
1748 ]->metadata_lte->output_resource_entry);
1751 if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
1752 ret = write_lookup_table(w, image, &hdr.lookup_table_res_entry);
1757 ret = write_xml_data(w->wim_info, image, out,
1758 (write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE) ?
1759 wim_info_get_total_bytes(w->wim_info) : 0,
1760 &hdr.xml_res_entry);
1764 if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
1765 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
1766 struct wim_header checkpoint_hdr;
1767 memcpy(&checkpoint_hdr, &hdr, sizeof(struct wim_header));
1768 zero_resource_entry(&checkpoint_hdr.integrity);
1769 if (fseeko(out, 0, SEEK_SET)) {
1770 ERROR_WITH_ERRNO("Failed to seek to beginning "
1771 "of WIM being written");
1772 ret = WIMLIB_ERR_WRITE;
1775 ret = write_header(&checkpoint_hdr, out);
1779 if (fflush(out) != 0) {
1780 ERROR_WITH_ERRNO("Can't write data to WIM");
1781 ret = WIMLIB_ERR_WRITE;
1785 if (fseeko(out, 0, SEEK_END) != 0) {
1786 ERROR_WITH_ERRNO("Failed to seek to end "
1787 "of WIM being written");
1788 ret = WIMLIB_ERR_WRITE;
1793 off_t old_lookup_table_end;
1794 off_t new_lookup_table_end;
1795 if (write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE) {
1796 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
1797 w->hdr.lookup_table_res_entry.size;
1799 old_lookup_table_end = 0;
1801 new_lookup_table_end = hdr.lookup_table_res_entry.offset +
1802 hdr.lookup_table_res_entry.size;
1804 ret = write_integrity_table(out,
1806 new_lookup_table_end,
1807 old_lookup_table_end,
1812 zero_resource_entry(&hdr.integrity);
1815 if (fseeko(out, 0, SEEK_SET) != 0) {
1816 ERROR_WITH_ERRNO("Failed to seek to beginning of WIM "
1818 ret = WIMLIB_ERR_WRITE;
1822 ret = write_header(&hdr, out);
1826 if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
1827 if (fflush(out) != 0
1828 || fsync(fileno(out)) != 0)
1830 ERROR_WITH_ERRNO("Error flushing data to WIM file");
1831 ret = WIMLIB_ERR_WRITE;
1835 if (fclose(out) != 0) {
1836 ERROR_WITH_ERRNO("Failed to close the output WIM file");
1838 ret = WIMLIB_ERR_WRITE;
1844 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
1846 lock_wim(WIMStruct *w, FILE *fp)
1849 if (fp && !w->wim_locked) {
1850 ret = flock(fileno(fp), LOCK_EX | LOCK_NB);
1852 if (errno == EWOULDBLOCK) {
1853 ERROR("`%"TS"' is already being modified or has been "
1854 "mounted read-write\n"
1855 " by another process!", w->filename);
1856 ret = WIMLIB_ERR_ALREADY_LOCKED;
1858 WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
1871 open_wim_writable(WIMStruct *w, const tchar *path,
1872 bool trunc, bool also_readable)
1883 wimlib_assert(w->out_fp == NULL);
1884 w->out_fp = tfopen(path, mode);
1888 ERROR_WITH_ERRNO("Failed to open `%"TS"' for writing", path);
1889 return WIMLIB_ERR_OPEN;
1895 close_wim_writable(WIMStruct *w)
1898 if (fclose(w->out_fp) != 0) {
1899 WARNING_WITH_ERRNO("Failed to close output WIM");
1905 /* Open file stream and write dummy header for WIM. */
1907 begin_write(WIMStruct *w, const tchar *path, int write_flags)
1910 ret = open_wim_writable(w, path, true,
1911 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
1914 /* Write dummy header. It will be overwritten later. */
1915 return write_header(&w->hdr, w->out_fp);
1918 /* Writes a stand-alone WIM to a file. */
1920 wimlib_write(WIMStruct *w, const tchar *path,
1921 int image, int write_flags, unsigned num_threads,
1922 wimlib_progress_func_t progress_func)
1927 return WIMLIB_ERR_INVALID_PARAM;
1929 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1931 if (image != WIMLIB_ALL_IMAGES &&
1932 (image < 1 || image > w->hdr.image_count))
1933 return WIMLIB_ERR_INVALID_IMAGE;
1935 if (w->hdr.total_parts != 1) {
1936 ERROR("Cannot call wimlib_write() on part of a split WIM");
1937 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1940 ret = begin_write(w, path, write_flags);
1944 ret = write_wim_streams(w, image, write_flags, num_threads,
1950 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
1952 ret = for_image(w, image, write_metadata_resource);
1957 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
1959 ret = finish_write(w, image, write_flags, progress_func);
1960 /* finish_write() closed the WIM for us */
1963 close_wim_writable(w);
1965 DEBUG("wimlib_write(path=%"TS") = %d", path, ret);
1970 any_images_modified(WIMStruct *w)
1972 for (int i = 0; i < w->hdr.image_count; i++)
1973 if (w->image_metadata[i]->modified)
1979 * Overwrite a WIM, possibly appending streams to it.
1981 * A WIM looks like (or is supposed to look like) the following:
1983 * Header (212 bytes)
1984 * Streams and metadata resources (variable size)
1985 * Lookup table (variable size)
1986 * XML data (variable size)
1987 * Integrity table (optional) (variable size)
1989 * If we are not adding any streams or metadata resources, the lookup table is
1990 * unchanged--- so we only need to overwrite the XML data, integrity table, and
1991 * header. This operation is potentially unsafe if the program is abruptly
1992 * terminated while the XML data or integrity table are being overwritten, but
1993 * before the new header has been written. To partially alleviate this problem,
1994 * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
1995 * finish_write() to cause a temporary WIM header to be written after the XML
1996 * data has been written. This may prevent the WIM from becoming corrupted if
1997 * the program is terminated while the integrity table is being calculated (but
1998 * no guarantees, due to write re-ordering...).
2000 * If we are adding new streams or images (metadata resources), the lookup table
2001 * needs to be changed, and those streams need to be written. In this case, we
2002 * try to perform a safe update of the WIM file by writing the streams *after*
2003 * the end of the previous WIM, then writing the new lookup table, XML data, and
2004 * (optionally) integrity table following the new streams. This will produce a
2005 * layout like the following:
2007 * Header (212 bytes)
2008 * (OLD) Streams and metadata resources (variable size)
2009 * (OLD) Lookup table (variable size)
2010 * (OLD) XML data (variable size)
2011 * (OLD) Integrity table (optional) (variable size)
2012 * (NEW) Streams and metadata resources (variable size)
2013 * (NEW) Lookup table (variable size)
2014 * (NEW) XML data (variable size)
2015 * (NEW) Integrity table (optional) (variable size)
2017 * At all points, the WIM is valid as nothing points to the new data yet. Then,
2018 * the header is overwritten to point to the new lookup table, XML data, and
2019 * integrity table, to produce the following layout:
2021 * Header (212 bytes)
2022 * Streams and metadata resources (variable size)
2023 * Nothing (variable size)
2024 * More Streams and metadata resources (variable size)
2025 * Lookup table (variable size)
2026 * XML data (variable size)
2027 * Integrity table (optional) (variable size)
2029 * This method allows an image to be appended to a large WIM very quickly, and
2030 * is is crash-safe except in the case of write re-ordering, but the
2031 * disadvantage is that a small hole is left in the WIM where the old lookup
2032 * table, xml data, and integrity table were. (These usually only take up a
2033 * small amount of space compared to the streams, however.)
2036 overwrite_wim_inplace(WIMStruct *w, int write_flags,
2037 unsigned num_threads,
2038 wimlib_progress_func_t progress_func)
2041 struct list_head stream_list;
2043 u64 old_lookup_table_end, old_xml_begin, old_xml_end;
2045 DEBUG("Overwriting `%"TS"' in-place", w->filename);
2047 /* Make sure that the integrity table (if present) is after the XML
2048 * data, and that there are no stream resources, metadata resources, or
2049 * lookup tables after the XML data. Otherwise, these data would be
2051 old_xml_begin = w->hdr.xml_res_entry.offset;
2052 old_xml_end = old_xml_begin + w->hdr.xml_res_entry.size;
2053 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
2054 w->hdr.lookup_table_res_entry.size;
2055 if (w->hdr.integrity.offset != 0 && w->hdr.integrity.offset < old_xml_end) {
2056 ERROR("Didn't expect the integrity table to be before the XML data");
2057 return WIMLIB_ERR_RESOURCE_ORDER;
2060 if (old_lookup_table_end > old_xml_begin) {
2061 ERROR("Didn't expect the lookup table to be after the XML data");
2062 return WIMLIB_ERR_RESOURCE_ORDER;
2065 /* Set @old_wim_end, which indicates the point beyond which we don't
2066 * allow any file and metadata resources to appear without returning
2067 * WIMLIB_ERR_RESOURCE_ORDER (due to the fact that we would otherwise
2068 * overwrite these resources). */
2069 if (!w->deletion_occurred && !any_images_modified(w)) {
2070 /* If no images have been modified and no images have been
2071 * deleted, a new lookup table does not need to be written. We
2072 * shall write the new XML data and optional integrity table
2073 * immediately after the lookup table. Note that this may
2074 * overwrite an existing integrity table. */
2075 DEBUG("Skipping writing lookup table "
2076 "(no images modified or deleted)");
2077 old_wim_end = old_lookup_table_end;
2078 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
2079 WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
2080 } else if (w->hdr.integrity.offset) {
2081 /* Old WIM has an integrity table; begin writing new streams
2083 old_wim_end = w->hdr.integrity.offset + w->hdr.integrity.size;
2085 /* No existing integrity table; begin writing new streams after
2086 * the old XML data. */
2087 old_wim_end = old_xml_end;
2090 ret = prepare_streams_for_overwrite(w, old_wim_end, &stream_list);
2094 ret = open_wim_writable(w, w->filename, false,
2095 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
2099 ret = lock_wim(w, w->out_fp);
2101 close_wim_writable(w);
2105 if (fseeko(w->out_fp, old_wim_end, SEEK_SET) != 0) {
2106 ERROR_WITH_ERRNO("Can't seek to end of WIM");
2107 close_wim_writable(w);
2109 return WIMLIB_ERR_WRITE;
2112 DEBUG("Writing newly added streams (offset = %"PRIu64")",
2114 ret = write_stream_list(&stream_list,
2117 wimlib_get_compression_type(w),
2124 for (int i = 0; i < w->hdr.image_count; i++) {
2125 if (w->image_metadata[i]->modified) {
2126 select_wim_image(w, i + 1);
2127 ret = write_metadata_resource(w);
2132 write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
2133 ret = finish_write(w, WIMLIB_ALL_IMAGES, write_flags,
2136 close_wim_writable(w);
2137 if (ret != 0 && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
2138 WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
2139 w->filename, old_wim_end);
2140 /* Return value of truncate() is ignored because this is already
2142 (void)ttruncate(w->filename, old_wim_end);
2149 overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
2150 unsigned num_threads,
2151 wimlib_progress_func_t progress_func)
2153 size_t wim_name_len;
2156 DEBUG("Overwriting `%"TS"' via a temporary file", w->filename);
2158 /* Write the WIM to a temporary file in the same directory as the
2160 wim_name_len = tstrlen(w->filename);
2161 tchar tmpfile[wim_name_len + 10];
2162 tmemcpy(tmpfile, w->filename, wim_name_len);
2163 randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
2164 tmpfile[wim_name_len + 9] = T('\0');
2166 ret = wimlib_write(w, tmpfile, WIMLIB_ALL_IMAGES,
2167 write_flags | WIMLIB_WRITE_FLAG_FSYNC,
2168 num_threads, progress_func);
2170 ERROR("Failed to write the WIM file `%"TS"'", tmpfile);
2174 DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, w->filename);
2177 /* Windows won't let you delete open files unless FILE_SHARE_DELETE was
2178 * specified to CreateFile(). The WIM was opened with fopen(), which
2179 * didn't provided this flag to CreateFile, so the handle must be closed
2180 * before executing the rename(). */
2181 if (w->fp != NULL) {
2187 /* Rename the new file to the old file .*/
2188 if (trename(tmpfile, w->filename) != 0) {
2189 ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
2190 tmpfile, w->filename);
2191 ret = WIMLIB_ERR_RENAME;
2195 if (progress_func) {
2196 union wimlib_progress_info progress;
2197 progress.rename.from = tmpfile;
2198 progress.rename.to = w->filename;
2199 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
2202 /* Close the original WIM file that was opened for reading. */
2203 if (w->fp != NULL) {
2208 /* Re-open the WIM read-only. */
2209 w->fp = tfopen(w->filename, T("rb"));
2210 if (w->fp == NULL) {
2211 ret = WIMLIB_ERR_REOPEN;
2212 WARNING_WITH_ERRNO("Failed to re-open `%"TS"' read-only",
2219 /* Remove temporary file. */
2220 if (tunlink(tmpfile) != 0)
2221 WARNING_WITH_ERRNO("Failed to remove `%"TS"'", tmpfile);
2227 * Writes a WIM file to the original file that it was read from, overwriting it.
2230 wimlib_overwrite(WIMStruct *w, int write_flags,
2231 unsigned num_threads,
2232 wimlib_progress_func_t progress_func)
2234 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2237 return WIMLIB_ERR_NO_FILENAME;
2239 if (w->hdr.total_parts != 1) {
2240 ERROR("Cannot modify a split WIM");
2241 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
2244 if ((!w->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
2245 && !(write_flags & WIMLIB_WRITE_FLAG_REBUILD))
2248 ret = overwrite_wim_inplace(w, write_flags, num_threads,
2250 if (ret == WIMLIB_ERR_RESOURCE_ORDER)
2251 WARNING("Falling back to re-building entire WIM");
2255 return overwrite_wim_via_tmpfile(w, write_flags, num_threads,