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 #if defined(__WIN32__) && !defined(INVALID_HANDLE_VALUE)
69 # define INVALID_HANDLE_VALUE ((HANDLE)(-1))
72 /* Chunk table that's located at the beginning of each compressed resource in
73 * the WIM. (This is not the on-disk format; the on-disk format just has an
74 * array of offsets.) */
78 u64 original_resource_size;
79 u64 bytes_per_chunk_entry;
87 * Allocates and initializes a chunk table, and reserves space for it in the
91 begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
94 struct chunk_table **chunk_tab_ret)
96 u64 size = wim_resource_size(lte);
97 u64 num_chunks = (size + WIM_CHUNK_SIZE - 1) / WIM_CHUNK_SIZE;
98 size_t alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
99 struct chunk_table *chunk_tab = CALLOC(1, alloc_size);
102 ERROR("Failed to allocate chunk table for %"PRIu64" byte "
104 return WIMLIB_ERR_NOMEM;
106 chunk_tab->file_offset = file_offset;
107 chunk_tab->num_chunks = num_chunks;
108 chunk_tab->original_resource_size = size;
109 chunk_tab->bytes_per_chunk_entry = (size >= (1ULL << 32)) ? 8 : 4;
110 chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
112 chunk_tab->cur_offset = 0;
113 chunk_tab->cur_offset_p = chunk_tab->offsets;
115 if (fwrite(chunk_tab, 1, chunk_tab->table_disk_size, out_fp) !=
116 chunk_tab->table_disk_size) {
117 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
120 return WIMLIB_ERR_WRITE;
122 *chunk_tab_ret = chunk_tab;
127 * compress_func_t- Pointer to a function to compresses a chunk
128 * of a WIM resource. This may be either
129 * wimlib_xpress_compress() (xpress-compress.c) or
130 * wimlib_lzx_compress() (lzx-compress.c).
132 * @chunk: Uncompressed data of the chunk.
133 * @chunk_size: Size of the uncompressed chunk, in bytes.
134 * @out: Pointer to output buffer of size at least (@chunk_size - 1) bytes.
136 * Returns the size of the compressed data written to @out in bytes, or 0 if the
137 * data could not be compressed to (@chunk_size - 1) bytes or fewer.
139 * As a special requirement, the compression code is optimized for the WIM
140 * format and therefore requires (@chunk_size <= 32768).
142 * As another special requirement, the compression code will read up to 8 bytes
143 * off the end of the @chunk array for performance reasons. The values of these
144 * bytes will not affect the output of the compression, but the calling code
145 * must make sure that the buffer holding the uncompressed chunk is actually at
146 * least (@chunk_size + 8) bytes, or at least that these extra bytes are in
147 * mapped memory that will not cause a memory access violation if accessed.
149 typedef unsigned (*compress_func_t)(const void *chunk, unsigned chunk_size,
152 static compress_func_t
153 get_compress_func(int out_ctype)
155 if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX)
156 return wimlib_lzx_compress;
158 return wimlib_xpress_compress;
162 * Writes a chunk of a WIM resource to an output file.
164 * @chunk: Uncompressed data of the chunk.
165 * @chunk_size: Size of the chunk (<= WIM_CHUNK_SIZE)
166 * @out_fp: FILE * to write the chunk to.
167 * @compress: Compression function to use (NULL if writing uncompressed
169 * @chunk_tab: Pointer to chunk table being created. It is updated with the
170 * offset of the chunk we write.
172 * Returns 0 on success; nonzero on failure.
175 write_wim_resource_chunk(const void *chunk, unsigned chunk_size,
176 FILE *out_fp, compress_func_t compress,
177 struct chunk_table *chunk_tab)
179 const void *out_chunk;
180 unsigned out_chunk_size;
182 u8 *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 *chunk_tab,
216 FILE *out_fp, u64 *compressed_size_p)
218 size_t bytes_written;
219 if (fseeko(out_fp, chunk_tab->file_offset, SEEK_SET) != 0) {
220 ERROR_WITH_ERRNO("Failed to seek to byte %"PRIu64" of output "
221 "WIM file", chunk_tab->file_offset);
222 return WIMLIB_ERR_WRITE;
225 if (chunk_tab->bytes_per_chunk_entry == 8) {
226 array_cpu_to_le64(chunk_tab->offsets, chunk_tab->num_chunks);
228 for (u64 i = 0; i < chunk_tab->num_chunks; i++)
229 ((u32*)chunk_tab->offsets)[i] =
230 cpu_to_le32(chunk_tab->offsets[i]);
232 bytes_written = fwrite((u8*)chunk_tab->offsets +
233 chunk_tab->bytes_per_chunk_entry,
234 1, chunk_tab->table_disk_size, out_fp);
235 if (bytes_written != chunk_tab->table_disk_size) {
236 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
238 return WIMLIB_ERR_WRITE;
240 if (fseeko(out_fp, 0, SEEK_END) != 0) {
241 ERROR_WITH_ERRNO("Failed to seek to end of output WIM file");
242 return WIMLIB_ERR_WRITE;
244 *compressed_size_p = chunk_tab->cur_offset + chunk_tab->table_disk_size;
249 fflush_and_ftruncate(FILE *out_fp, off_t offset)
251 if (fseeko(out_fp, offset, SEEK_SET) ||
253 ftruncate(fileno(out_fp), offset))
255 ERROR_WITH_ERRNO("Failed to flush and/or truncate "
257 return WIMLIB_ERR_WRITE;
264 finalize_and_check_sha1(SHA_CTX *sha_ctx, struct wim_lookup_table_entry *lte)
266 u8 md[SHA1_HASH_SIZE];
267 sha1_final(md, sha_ctx);
269 copy_hash(lte->hash, md);
270 } else if (!hashes_equal(md, lte->hash)) {
271 ERROR("WIM resource has incorrect hash!");
272 if (lte_filename_valid(lte)) {
273 ERROR("We were reading it from \"%"TS"\"; maybe "
274 "it changed while we were reading it.",
277 return WIMLIB_ERR_INVALID_RESOURCE_HASH;
283 struct write_resource_ctx {
284 compress_func_t compress;
285 struct chunk_table *chunk_tab;
292 write_resource_cb(const void *chunk, size_t chunk_size, void *_ctx)
294 struct write_resource_ctx *ctx = _ctx;
297 sha1_update(&ctx->sha_ctx, chunk, chunk_size);
298 return write_wim_resource_chunk(chunk, chunk_size,
299 ctx->out_fp, ctx->compress,
304 * Write a resource to an output WIM.
306 * @lte: Lookup table entry for the resource, which could be in another WIM,
307 * in an external file, or in another location.
309 * @out_fp: FILE * opened to the output WIM.
311 * @out_ctype: One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate
312 * which compression algorithm to use.
314 * @out_res_entry: On success, this is filled in with the offset, flags,
315 * compressed size, and uncompressed size of the resource
318 * @flags: WIMLIB_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed
319 * even if it could otherwise be copied directly from the input.
321 * Additional notes: The SHA1 message digest of the uncompressed data is
322 * calculated (except when doing a raw copy --- see below). If the @unhashed
323 * flag is set on the lookup table entry, this message digest is simply copied
324 * to it; otherwise, the message digest is compared with the existing one, and
325 * the function will fail if they do not match.
328 write_wim_resource(struct wim_lookup_table_entry *lte,
329 FILE *out_fp, int out_ctype,
330 struct resource_entry *out_res_entry,
333 struct write_resource_ctx write_ctx;
339 flags &= ~WIMLIB_RESOURCE_FLAG_RECOMPRESS;
341 /* Get current position in output WIM */
342 offset = ftello(out_fp);
344 ERROR_WITH_ERRNO("Can't get position in output WIM");
345 return WIMLIB_ERR_WRITE;
348 /* If we are not forcing the data to be recompressed, and the input
349 * resource is located in a WIM with the same compression type as that
350 * desired other than no compression, we can simply copy the compressed
351 * data without recompressing it. This also means we must skip
352 * calculating the SHA1, as we never will see the uncompressed data. */
353 if (!(flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
354 lte->resource_location == RESOURCE_IN_WIM &&
355 out_ctype != WIMLIB_COMPRESSION_TYPE_NONE &&
356 wimlib_get_compression_type(lte->wim) == out_ctype)
358 flags |= WIMLIB_RESOURCE_FLAG_RAW;
359 write_ctx.doing_sha = false;
360 read_size = lte->resource_entry.size;
362 write_ctx.doing_sha = true;
363 sha1_init(&write_ctx.sha_ctx);
364 read_size = lte->resource_entry.original_size;
367 /* Initialize the chunk table and set the compression function if
368 * compressing the resource. */
369 if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
370 (flags & WIMLIB_RESOURCE_FLAG_RAW)) {
371 write_ctx.compress = NULL;
372 write_ctx.chunk_tab = NULL;
374 write_ctx.compress = get_compress_func(out_ctype);
375 ret = begin_wim_resource_chunk_tab(lte, out_fp,
377 &write_ctx.chunk_tab);
382 /* Write the entire resource by reading the entire resource and feeding
383 * the data through the write_resource_cb function. */
384 write_ctx.out_fp = out_fp;
386 ret = read_resource_prefix(lte, read_size,
387 write_resource_cb, &write_ctx, flags);
389 goto out_free_chunk_tab;
391 /* Verify SHA1 message digest of the resource, or set the hash for the
393 if (write_ctx.doing_sha) {
394 ret = finalize_and_check_sha1(&write_ctx.sha_ctx, lte);
396 goto out_free_chunk_tab;
399 out_res_entry->flags = lte->resource_entry.flags;
400 out_res_entry->original_size = wim_resource_size(lte);
401 out_res_entry->offset = offset;
402 if (flags & WIMLIB_RESOURCE_FLAG_RAW) {
403 /* Doing a raw write: The new compressed size is the same as
404 * the compressed size in the other WIM. */
405 new_size = lte->resource_entry.size;
406 } else if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE) {
407 /* Using WIMLIB_COMPRESSION_TYPE_NONE: The new compressed size
408 * is the original size. */
409 new_size = lte->resource_entry.original_size;
410 out_res_entry->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
412 /* Using a different compression type: Call
413 * finish_wim_resource_chunk_tab() and it will provide the new
414 * compressed size. */
415 ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab, out_fp,
418 goto out_free_chunk_tab;
419 if (new_size >= wim_resource_size(lte)) {
420 /* Oops! We compressed the resource to larger than the original
421 * size. Write the resource uncompressed instead. */
422 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
423 "writing uncompressed instead",
424 wim_resource_size(lte), new_size);
425 ret = fflush_and_ftruncate(out_fp, offset);
427 goto out_free_chunk_tab;
428 write_ctx.compress = NULL;
429 write_ctx.doing_sha = false;
430 out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
431 goto try_write_again;
433 out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
435 out_res_entry->size = new_size;
438 FREE(write_ctx.chunk_tab);
442 #ifdef ENABLE_MULTITHREADED_COMPRESSION
444 /* Blocking shared queue (solves the producer-consumer problem) */
445 struct shared_queue {
449 unsigned filled_slots;
451 pthread_mutex_t lock;
452 pthread_cond_t msg_avail_cond;
453 pthread_cond_t space_avail_cond;
457 shared_queue_init(struct shared_queue *q, unsigned size)
459 wimlib_assert(size != 0);
460 q->array = CALLOC(sizeof(q->array[0]), size);
467 if (pthread_mutex_init(&q->lock, NULL)) {
468 ERROR_WITH_ERRNO("Failed to initialize mutex");
471 if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
472 ERROR_WITH_ERRNO("Failed to initialize condition variable");
473 goto err_destroy_lock;
475 if (pthread_cond_init(&q->space_avail_cond, NULL)) {
476 ERROR_WITH_ERRNO("Failed to initialize condition variable");
477 goto err_destroy_msg_avail_cond;
480 err_destroy_msg_avail_cond:
481 pthread_cond_destroy(&q->msg_avail_cond);
483 pthread_mutex_destroy(&q->lock);
485 return WIMLIB_ERR_NOMEM;
489 shared_queue_destroy(struct shared_queue *q)
492 pthread_mutex_destroy(&q->lock);
493 pthread_cond_destroy(&q->msg_avail_cond);
494 pthread_cond_destroy(&q->space_avail_cond);
498 shared_queue_put(struct shared_queue *q, void *obj)
500 pthread_mutex_lock(&q->lock);
501 while (q->filled_slots == q->size)
502 pthread_cond_wait(&q->space_avail_cond, &q->lock);
504 q->back = (q->back + 1) % q->size;
505 q->array[q->back] = obj;
508 pthread_cond_broadcast(&q->msg_avail_cond);
509 pthread_mutex_unlock(&q->lock);
513 shared_queue_get(struct shared_queue *q)
517 pthread_mutex_lock(&q->lock);
518 while (q->filled_slots == 0)
519 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
521 obj = q->array[q->front];
522 q->array[q->front] = NULL;
523 q->front = (q->front + 1) % q->size;
526 pthread_cond_broadcast(&q->space_avail_cond);
527 pthread_mutex_unlock(&q->lock);
531 struct compressor_thread_params {
532 struct shared_queue *res_to_compress_queue;
533 struct shared_queue *compressed_res_queue;
534 compress_func_t compress;
537 #define MAX_CHUNKS_PER_MSG 2
540 struct wim_lookup_table_entry *lte;
541 u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
542 u8 *out_compressed_chunks[MAX_CHUNKS_PER_MSG];
543 u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
544 unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
545 unsigned compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
547 struct list_head list;
553 compress_chunks(struct message *msg, compress_func_t compress)
555 for (unsigned i = 0; i < msg->num_chunks; i++) {
556 unsigned len = compress(msg->uncompressed_chunks[i],
557 msg->uncompressed_chunk_sizes[i],
558 msg->compressed_chunks[i]);
560 /* To be written compressed */
561 msg->out_compressed_chunks[i] = msg->compressed_chunks[i];
562 msg->compressed_chunk_sizes[i] = len;
564 /* To be written uncompressed */
565 msg->out_compressed_chunks[i] = msg->uncompressed_chunks[i];
566 msg->compressed_chunk_sizes[i] = msg->uncompressed_chunk_sizes[i];
572 /* Compressor thread routine. This is a lot simpler than the main thread
573 * routine: just repeatedly get a group of chunks from the
574 * res_to_compress_queue, compress them, and put them in the
575 * compressed_res_queue. A NULL pointer indicates that the thread should stop.
578 compressor_thread_proc(void *arg)
580 struct compressor_thread_params *params = arg;
581 struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
582 struct shared_queue *compressed_res_queue = params->compressed_res_queue;
583 compress_func_t compress = params->compress;
586 DEBUG("Compressor thread ready");
587 while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
588 compress_chunks(msg, compress);
589 shared_queue_put(compressed_res_queue, msg);
591 DEBUG("Compressor thread terminating");
594 #endif /* ENABLE_MULTITHREADED_COMPRESSION */
597 do_write_streams_progress(union wimlib_progress_info *progress,
598 wimlib_progress_func_t progress_func,
601 progress->write_streams.completed_bytes += size_added;
602 progress->write_streams.completed_streams++;
604 progress->write_streams.completed_bytes >= progress->write_streams._private)
606 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
608 if (progress->write_streams._private == progress->write_streams.total_bytes) {
609 progress->write_streams._private = ~0;
611 progress->write_streams._private =
612 min(progress->write_streams.total_bytes,
613 progress->write_streams.completed_bytes +
614 progress->write_streams.total_bytes / 100);
619 struct serial_write_stream_ctx {
622 int write_resource_flags;
626 serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
628 struct serial_write_stream_ctx *ctx = _ctx;
629 return write_wim_resource(lte, ctx->out_fp,
630 ctx->out_ctype, <e->output_resource_entry,
631 ctx->write_resource_flags);
634 /* Write a list of streams, taking into account that some streams may be
635 * duplicates that are checksummed and discarded on the fly, and also delegating
636 * the actual writing of a stream to a function @write_stream_cb, which is
637 * passed the context @write_stream_ctx. */
639 do_write_stream_list(struct list_head *stream_list,
640 struct wim_lookup_table *lookup_table,
641 int (*write_stream_cb)(struct wim_lookup_table_entry *, void *),
642 void *write_stream_ctx,
643 wimlib_progress_func_t progress_func,
644 union wimlib_progress_info *progress)
647 struct wim_lookup_table_entry *lte;
649 /* For each stream in @stream_list ... */
650 while (!list_empty(stream_list)) {
651 lte = container_of(stream_list->next,
652 struct wim_lookup_table_entry,
654 list_del(<e->write_streams_list);
655 if (lte->unhashed && !lte->unique_size) {
656 /* Unhashed stream that shares a size with some other
657 * stream in the WIM we are writing. The stream must be
658 * checksummed to know if we need to write it or not. */
659 struct wim_lookup_table_entry *tmp;
660 u32 orig_refcnt = lte->out_refcnt;
662 ret = hash_unhashed_stream(lte, lookup_table, &tmp);
667 /* We found a duplicate stream. */
668 if (orig_refcnt != tmp->out_refcnt) {
669 /* We have already written, or are going
670 * to write, the duplicate stream. So
671 * just skip to the next stream. */
672 DEBUG("Discarding duplicate stream of length %"PRIu64,
673 wim_resource_size(lte));
674 lte->no_progress = 0;
675 goto skip_to_progress;
680 /* Here, @lte is either a hashed stream or an unhashed stream
681 * with a unique size. In either case we know that the stream
682 * has to be written. In either case the SHA1 message digest
683 * will be calculated over the stream while writing it; however,
684 * in the former case this is done merely to check the data,
685 * while in the latter case this is done because we do not have
686 * the SHA1 message digest yet. */
687 wimlib_assert(lte->out_refcnt != 0);
689 lte->no_progress = 0;
690 ret = (*write_stream_cb)(lte, write_stream_ctx);
693 /* In parallel mode, some streams are deferred for later,
694 * serialized processing; ignore them here. */
698 list_del(<e->unhashed_list);
699 lookup_table_insert(lookup_table, lte);
703 if (!lte->no_progress) {
704 do_write_streams_progress(progress,
706 wim_resource_size(lte));
713 do_write_stream_list_serial(struct list_head *stream_list,
714 struct wim_lookup_table *lookup_table,
717 int write_resource_flags,
718 wimlib_progress_func_t progress_func,
719 union wimlib_progress_info *progress)
721 struct serial_write_stream_ctx ctx = {
723 .out_ctype = out_ctype,
724 .write_resource_flags = write_resource_flags,
726 return do_write_stream_list(stream_list,
735 write_flags_to_resource_flags(int write_flags)
737 return (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS) ?
738 WIMLIB_RESOURCE_FLAG_RECOMPRESS : 0;
742 write_stream_list_serial(struct list_head *stream_list,
743 struct wim_lookup_table *lookup_table,
746 int write_resource_flags,
747 wimlib_progress_func_t progress_func,
748 union wimlib_progress_info *progress)
750 DEBUG("Writing stream list (serial version)");
751 progress->write_streams.num_threads = 1;
753 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
754 return do_write_stream_list_serial(stream_list,
758 write_resource_flags,
763 #ifdef ENABLE_MULTITHREADED_COMPRESSION
765 write_wim_chunks(struct message *msg, FILE *out_fp,
766 struct chunk_table *chunk_tab)
768 for (unsigned i = 0; i < msg->num_chunks; i++) {
769 unsigned chunk_csize = msg->compressed_chunk_sizes[i];
771 if (fwrite(msg->out_compressed_chunks[i], 1, chunk_csize, out_fp)
774 ERROR_WITH_ERRNO("Failed to write WIM chunk");
775 return WIMLIB_ERR_WRITE;
778 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
779 chunk_tab->cur_offset += chunk_csize;
784 struct main_writer_thread_ctx {
785 struct list_head *stream_list;
786 struct wim_lookup_table *lookup_table;
789 int write_resource_flags;
790 struct shared_queue *res_to_compress_queue;
791 struct shared_queue *compressed_res_queue;
793 wimlib_progress_func_t progress_func;
794 union wimlib_progress_info *progress;
796 struct list_head available_msgs;
797 struct list_head outstanding_streams;
798 struct list_head serial_streams;
799 size_t num_outstanding_messages;
801 SHA_CTX next_sha_ctx;
804 struct wim_lookup_table_entry *next_lte;
806 struct message *msgs;
807 struct message *next_msg;
808 struct chunk_table *cur_chunk_tab;
812 init_message(struct message *msg)
814 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
815 msg->compressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
816 msg->uncompressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
817 if (msg->compressed_chunks[i] == NULL ||
818 msg->uncompressed_chunks[i] == NULL)
819 return WIMLIB_ERR_NOMEM;
825 destroy_message(struct message *msg)
827 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
828 FREE(msg->compressed_chunks[i]);
829 FREE(msg->uncompressed_chunks[i]);
834 free_messages(struct message *msgs, size_t num_messages)
837 for (size_t i = 0; i < num_messages; i++)
838 destroy_message(&msgs[i]);
843 static struct message *
844 allocate_messages(size_t num_messages)
846 struct message *msgs;
848 msgs = CALLOC(num_messages, sizeof(struct message));
851 for (size_t i = 0; i < num_messages; i++) {
852 if (init_message(&msgs[i])) {
853 free_messages(msgs, num_messages);
861 main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
863 while (ctx->num_outstanding_messages--)
864 shared_queue_get(ctx->compressed_res_queue);
865 free_messages(ctx->msgs, ctx->num_messages);
866 FREE(ctx->cur_chunk_tab);
870 main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
872 /* Pre-allocate all the buffers that will be needed to do the chunk
874 ctx->msgs = allocate_messages(ctx->num_messages);
876 return WIMLIB_ERR_NOMEM;
878 /* Initially, all the messages are available to use. */
879 INIT_LIST_HEAD(&ctx->available_msgs);
880 for (size_t i = 0; i < ctx->num_messages; i++)
881 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
883 /* outstanding_streams is the list of streams that currently have had
884 * chunks sent off for compression.
886 * The first stream in outstanding_streams is the stream that is
887 * currently being written.
889 * The last stream in outstanding_streams is the stream that is
890 * currently being read and having chunks fed to the compressor threads.
892 INIT_LIST_HEAD(&ctx->outstanding_streams);
893 ctx->num_outstanding_messages = 0;
895 ctx->next_msg = NULL;
897 /* Resources that don't need any chunks compressed are added to this
898 * list and written directly by the main thread. */
899 INIT_LIST_HEAD(&ctx->serial_streams);
901 ctx->cur_chunk_tab = NULL;
907 receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
910 struct wim_lookup_table_entry *cur_lte;
913 wimlib_assert(!list_empty(&ctx->outstanding_streams));
914 wimlib_assert(ctx->num_outstanding_messages != 0);
916 cur_lte = container_of(ctx->outstanding_streams.next,
917 struct wim_lookup_table_entry,
918 being_compressed_list);
920 /* Get the next message from the queue and process it.
921 * The message will contain 1 or more data chunks that have been
923 msg = shared_queue_get(ctx->compressed_res_queue);
924 msg->complete = true;
925 --ctx->num_outstanding_messages;
927 /* Is this the next chunk in the current resource? If it's not
928 * (i.e., an earlier chunk in a same or different resource
929 * hasn't been compressed yet), do nothing, and keep this
930 * message around until all earlier chunks are received.
932 * Otherwise, write all the chunks we can. */
933 while (cur_lte != NULL &&
934 !list_empty(&cur_lte->msg_list)
935 && (msg = container_of(cur_lte->msg_list.next,
939 list_move(&msg->list, &ctx->available_msgs);
940 if (msg->begin_chunk == 0) {
941 /* This is the first set of chunks. Leave space
942 * for the chunk table in the output file. */
943 off_t cur_offset = ftello(ctx->out_fp);
944 if (cur_offset == -1)
945 return WIMLIB_ERR_WRITE;
946 ret = begin_wim_resource_chunk_tab(cur_lte,
949 &ctx->cur_chunk_tab);
954 /* Write the compressed chunks from the message. */
955 ret = write_wim_chunks(msg, ctx->out_fp, ctx->cur_chunk_tab);
959 /* Was this the last chunk of the stream? If so, finish
961 if (list_empty(&cur_lte->msg_list) &&
962 msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
967 ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
973 list_del(&cur_lte->being_compressed_list);
975 /* Grab the offset of this stream in the output file
976 * from the chunk table before we free it. */
977 offset = ctx->cur_chunk_tab->file_offset;
979 FREE(ctx->cur_chunk_tab);
980 ctx->cur_chunk_tab = NULL;
982 if (res_csize >= wim_resource_size(cur_lte)) {
983 /* Oops! We compressed the resource to
984 * larger than the original size. Write
985 * the resource uncompressed instead. */
986 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
987 "writing uncompressed instead",
988 wim_resource_size(cur_lte), res_csize);
989 ret = fflush_and_ftruncate(ctx->out_fp, offset);
992 ret = write_wim_resource(cur_lte,
994 WIMLIB_COMPRESSION_TYPE_NONE,
995 &cur_lte->output_resource_entry,
996 ctx->write_resource_flags);
1000 cur_lte->output_resource_entry.size =
1003 cur_lte->output_resource_entry.original_size =
1004 cur_lte->resource_entry.original_size;
1006 cur_lte->output_resource_entry.offset =
1009 cur_lte->output_resource_entry.flags =
1010 cur_lte->resource_entry.flags |
1011 WIM_RESHDR_FLAG_COMPRESSED;
1014 do_write_streams_progress(ctx->progress,
1016 wim_resource_size(cur_lte));
1018 /* Since we just finished writing a stream, write any
1019 * streams that have been added to the serial_streams
1020 * list for direct writing by the main thread (e.g.
1021 * resources that don't need to be compressed because
1022 * the desired compression type is the same as the
1023 * previous compression type). */
1024 if (!list_empty(&ctx->serial_streams)) {
1025 ret = do_write_stream_list_serial(&ctx->serial_streams,
1029 ctx->write_resource_flags,
1036 /* Advance to the next stream to write. */
1037 if (list_empty(&ctx->outstanding_streams)) {
1040 cur_lte = container_of(ctx->outstanding_streams.next,
1041 struct wim_lookup_table_entry,
1042 being_compressed_list);
1049 /* Called when the main thread has read a new chunk of data. */
1051 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
1053 struct main_writer_thread_ctx *ctx = _ctx;
1055 struct message *next_msg;
1056 u64 next_chunk_in_msg;
1058 /* Update SHA1 message digest for the stream currently being read by the
1060 sha1_update(&ctx->next_sha_ctx, chunk, chunk_size);
1062 /* We send chunks of data to the compressor chunks in batches which we
1063 * refer to as "messages". @next_msg is the message that is currently
1064 * being prepared to send off. If it is NULL, that indicates that we
1065 * need to start a new message. */
1066 next_msg = ctx->next_msg;
1068 /* We need to start a new message. First check to see if there
1069 * is a message available in the list of available messages. If
1070 * so, we can just take one. If not, all the messages (there is
1071 * a fixed number of them, proportional to the number of
1072 * threads) have been sent off to the compressor threads, so we
1073 * receive messages from the compressor threads containing
1074 * compressed chunks of data.
1076 * We may need to receive multiple messages before one is
1077 * actually available to use because messages received that are
1078 * *not* for the very next set of chunks to compress must be
1079 * buffered until it's time to write those chunks. */
1080 while (list_empty(&ctx->available_msgs)) {
1081 ret = receive_compressed_chunks(ctx);
1086 next_msg = container_of(ctx->available_msgs.next,
1087 struct message, list);
1088 list_del(&next_msg->list);
1089 next_msg->complete = false;
1090 next_msg->begin_chunk = ctx->next_chunk;
1091 next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
1092 ctx->next_num_chunks - ctx->next_chunk);
1093 ctx->next_msg = next_msg;
1096 /* Fill in the next chunk to compress */
1097 next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;
1099 next_msg->uncompressed_chunk_sizes[next_chunk_in_msg] = chunk_size;
1100 memcpy(next_msg->uncompressed_chunks[next_chunk_in_msg],
1103 if (++next_chunk_in_msg == next_msg->num_chunks) {
1104 /* Send off an array of chunks to compress */
1105 list_add_tail(&next_msg->list, &ctx->next_lte->msg_list);
1106 shared_queue_put(ctx->res_to_compress_queue, next_msg);
1107 ++ctx->num_outstanding_messages;
1108 ctx->next_msg = NULL;
1114 main_writer_thread_finish(void *_ctx)
1116 struct main_writer_thread_ctx *ctx = _ctx;
1118 while (ctx->num_outstanding_messages != 0) {
1119 ret = receive_compressed_chunks(ctx);
1123 wimlib_assert(list_empty(&ctx->outstanding_streams));
1124 return do_write_stream_list_serial(&ctx->serial_streams,
1128 ctx->write_resource_flags,
1134 submit_stream_for_compression(struct wim_lookup_table_entry *lte,
1135 struct main_writer_thread_ctx *ctx)
1139 /* Read the entire stream @lte, feeding its data chunks to the
1140 * compressor threads. Also SHA1-sum the stream; this is required in
1141 * the case that @lte is unhashed, and a nice additional verification
1142 * when @lte is already hashed. */
1143 sha1_init(&ctx->next_sha_ctx);
1144 ctx->next_chunk = 0;
1145 ctx->next_num_chunks = wim_resource_chunks(lte);
1146 ctx->next_lte = lte;
1147 INIT_LIST_HEAD(<e->msg_list);
1148 list_add_tail(<e->being_compressed_list, &ctx->outstanding_streams);
1149 ret = read_resource_prefix(lte, wim_resource_size(lte),
1150 main_writer_thread_cb, ctx, 0);
1152 wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
1153 ret = finalize_and_check_sha1(&ctx->next_sha_ctx, lte);
1159 main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
1161 struct main_writer_thread_ctx *ctx = _ctx;
1164 if (wim_resource_size(lte) < 1000 ||
1165 ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
1166 (lte->resource_location == RESOURCE_IN_WIM &&
1167 !(ctx->write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
1168 wimlib_get_compression_type(lte->wim) == ctx->out_ctype))
1170 /* Stream is too small or isn't being compressed. Process it by
1171 * the main thread when we have a chance. We can't necessarily
1172 * process it right here, as the main thread could be in the
1173 * middle of writing a different stream. */
1174 list_add_tail(<e->write_streams_list, &ctx->serial_streams);
1178 ret = submit_stream_for_compression(lte, ctx);
1180 lte->no_progress = 1;
1185 get_default_num_threads()
1188 return win32_get_number_of_processors();
1190 return sysconf(_SC_NPROCESSORS_ONLN);
1194 /* Equivalent to write_stream_list_serial(), except this takes a @num_threads
1195 * parameter and will perform compression using that many threads. Falls
1196 * back to write_stream_list_serial() on certain errors, such as a failure to
1197 * create the number of threads requested.
1199 * High level description of the algorithm for writing compressed streams in
1200 * parallel: We perform compression on chunks of size WIM_CHUNK_SIZE bytes
1201 * rather than on full files. The currently executing thread becomes the main
1202 * thread and is entirely in charge of reading the data to compress (which may
1203 * be in any location understood by the resource code--- such as in an external
1204 * file being captured, or in another WIM file from which an image is being
1205 * exported) and actually writing the compressed data to the output file.
1206 * Additional threads are "compressor threads" and all execute the
1207 * compressor_thread_proc, where they repeatedly retrieve buffers of data from
1208 * the main thread, compress them, and hand them back to the main thread.
1210 * Certain streams, such as streams that do not need to be compressed (e.g.
1211 * input compression type same as output compression type) or streams of very
1212 * small size are placed in a list (main_writer_thread_ctx.serial_list) and
1213 * handled entirely by the main thread at an appropriate time.
1215 * At any given point in time, multiple streams may be having chunks compressed
1216 * concurrently. The stream that the main thread is currently *reading* may be
1217 * later in the list that the stream that the main thread is currently
1221 write_stream_list_parallel(struct list_head *stream_list,
1222 struct wim_lookup_table *lookup_table,
1225 int write_resource_flags,
1226 wimlib_progress_func_t progress_func,
1227 union wimlib_progress_info *progress,
1228 unsigned num_threads)
1231 struct shared_queue res_to_compress_queue;
1232 struct shared_queue compressed_res_queue;
1233 pthread_t *compressor_threads = NULL;
1235 if (num_threads == 0) {
1236 long nthreads = get_default_num_threads();
1237 if (nthreads < 1 || nthreads > UINT_MAX) {
1238 WARNING("Could not determine number of processors! Assuming 1");
1240 } else if (nthreads == 1) {
1241 goto out_serial_quiet;
1243 num_threads = nthreads;
1247 DEBUG("Writing stream list (parallel version, num_threads=%u)",
1250 progress->write_streams.num_threads = num_threads;
1252 static const size_t MESSAGES_PER_THREAD = 2;
1253 size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1255 DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1257 ret = shared_queue_init(&res_to_compress_queue, queue_size);
1261 ret = shared_queue_init(&compressed_res_queue, queue_size);
1263 goto out_destroy_res_to_compress_queue;
1265 struct compressor_thread_params params;
1266 params.res_to_compress_queue = &res_to_compress_queue;
1267 params.compressed_res_queue = &compressed_res_queue;
1268 params.compress = get_compress_func(out_ctype);
1270 compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1271 if (!compressor_threads) {
1272 ret = WIMLIB_ERR_NOMEM;
1273 goto out_destroy_compressed_res_queue;
1276 for (unsigned i = 0; i < num_threads; i++) {
1277 DEBUG("pthread_create thread %u of %u", i + 1, num_threads);
1278 ret = pthread_create(&compressor_threads[i], NULL,
1279 compressor_thread_proc, ¶ms);
1282 ERROR_WITH_ERRNO("Failed to create compressor "
1284 i + 1, num_threads);
1291 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
1293 struct main_writer_thread_ctx ctx;
1294 ctx.stream_list = stream_list;
1295 ctx.lookup_table = lookup_table;
1296 ctx.out_fp = out_fp;
1297 ctx.out_ctype = out_ctype;
1298 ctx.res_to_compress_queue = &res_to_compress_queue;
1299 ctx.compressed_res_queue = &compressed_res_queue;
1300 ctx.num_messages = queue_size;
1301 ctx.write_resource_flags = write_resource_flags | WIMLIB_RESOURCE_FLAG_THREADSAFE_READ;
1302 ctx.progress_func = progress_func;
1303 ctx.progress = progress;
1304 ret = main_writer_thread_init_ctx(&ctx);
1307 ret = do_write_stream_list(stream_list, lookup_table,
1308 main_thread_process_next_stream,
1309 &ctx, progress_func, progress);
1311 goto out_destroy_ctx;
1313 /* The main thread has finished reading all streams that are going to be
1314 * compressed in parallel, and it now needs to wait for all remaining
1315 * chunks to be compressed so that the remaining streams can actually be
1316 * written to the output file. Furthermore, any remaining streams that
1317 * had processing deferred to the main thread need to be handled. These
1318 * tasks are done by the main_writer_thread_finish() function. */
1319 ret = main_writer_thread_finish(&ctx);
1321 main_writer_thread_destroy_ctx(&ctx);
1323 for (unsigned i = 0; i < num_threads; i++)
1324 shared_queue_put(&res_to_compress_queue, NULL);
1326 for (unsigned i = 0; i < num_threads; i++) {
1327 if (pthread_join(compressor_threads[i], NULL)) {
1328 WARNING_WITH_ERRNO("Failed to join compressor "
1330 i + 1, num_threads);
1333 FREE(compressor_threads);
1334 out_destroy_compressed_res_queue:
1335 shared_queue_destroy(&compressed_res_queue);
1336 out_destroy_res_to_compress_queue:
1337 shared_queue_destroy(&res_to_compress_queue);
1338 if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
1341 WARNING("Falling back to single-threaded compression");
1343 return write_stream_list_serial(stream_list,
1347 write_resource_flags,
1355 * Write a list of streams to a WIM (@out_fp) using the compression type
1356 * @out_ctype and up to @num_threads compressor threads.
1359 write_stream_list(struct list_head *stream_list,
1360 struct wim_lookup_table *lookup_table,
1361 FILE *out_fp, int out_ctype, int write_flags,
1362 unsigned num_threads, wimlib_progress_func_t progress_func)
1364 struct wim_lookup_table_entry *lte;
1365 size_t num_streams = 0;
1366 u64 total_bytes = 0;
1367 u64 total_compression_bytes = 0;
1368 union wimlib_progress_info progress;
1370 int write_resource_flags;
1372 if (list_empty(stream_list))
1375 write_resource_flags = write_flags_to_resource_flags(write_flags);
1377 /* Calculate the total size of the streams to be written. Note: this
1378 * will be the uncompressed size, as we may not know the compressed size
1379 * yet, and also this will assume that every unhashed stream will be
1380 * written (which will not necessarily be the case). */
1381 list_for_each_entry(lte, stream_list, write_streams_list) {
1383 total_bytes += wim_resource_size(lte);
1384 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
1385 && (wim_resource_compression_type(lte) != out_ctype ||
1386 (write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS)))
1388 total_compression_bytes += wim_resource_size(lte);
1391 progress.write_streams.total_bytes = total_bytes;
1392 progress.write_streams.total_streams = num_streams;
1393 progress.write_streams.completed_bytes = 0;
1394 progress.write_streams.completed_streams = 0;
1395 progress.write_streams.num_threads = num_threads;
1396 progress.write_streams.compression_type = out_ctype;
1397 progress.write_streams._private = 0;
1399 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1400 if (total_compression_bytes >= 1000000 && num_threads != 1)
1401 ret = write_stream_list_parallel(stream_list,
1405 write_resource_flags,
1411 ret = write_stream_list_serial(stream_list,
1415 write_resource_flags,
1421 struct stream_size_table {
1422 struct hlist_head *array;
1428 init_stream_size_table(struct stream_size_table *tab, size_t capacity)
1430 tab->array = CALLOC(capacity, sizeof(tab->array[0]));
1432 return WIMLIB_ERR_NOMEM;
1433 tab->num_entries = 0;
1434 tab->capacity = capacity;
1439 destroy_stream_size_table(struct stream_size_table *tab)
1445 stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
1447 struct stream_size_table *tab = _tab;
1449 struct wim_lookup_table_entry *same_size_lte;
1450 struct hlist_node *tmp;
1452 pos = hash_u64(wim_resource_size(lte)) % tab->capacity;
1453 lte->unique_size = 1;
1454 hlist_for_each_entry(same_size_lte, tmp, &tab->array[pos], hash_list_2) {
1455 if (wim_resource_size(same_size_lte) == wim_resource_size(lte)) {
1456 lte->unique_size = 0;
1457 same_size_lte->unique_size = 0;
1462 hlist_add_head(<e->hash_list_2, &tab->array[pos]);
1468 struct lte_overwrite_prepare_args {
1471 struct list_head stream_list;
1472 struct stream_size_table stream_size_tab;
1475 /* First phase of preparing streams for an in-place overwrite. This is called
1476 * on all streams, both hashed and unhashed, except the metadata resources. */
1478 lte_overwrite_prepare(struct wim_lookup_table_entry *lte, void *_args)
1480 struct lte_overwrite_prepare_args *args = _args;
1482 wimlib_assert(!(lte->resource_entry.flags & WIM_RESHDR_FLAG_METADATA));
1483 if (lte->resource_location != RESOURCE_IN_WIM || lte->wim != args->wim)
1484 list_add_tail(<e->write_streams_list, &args->stream_list);
1485 lte->out_refcnt = lte->refcnt;
1486 stream_size_table_insert(lte, &args->stream_size_tab);
1490 /* Second phase of preparing streams for an in-place overwrite. This is called
1491 * on existing metadata resources and hashed streams, but not unhashed streams.
1493 * NOTE: lte->output_resource_entry is in union with lte->hash_list_2, so
1494 * lte_overwrite_prepare_2() must be called after lte_overwrite_prepare(), as
1495 * the latter uses lte->hash_list_2, while the former expects to set
1496 * lte->output_resource_entry. */
1498 lte_overwrite_prepare_2(struct wim_lookup_table_entry *lte, void *_args)
1500 struct lte_overwrite_prepare_args *args = _args;
1502 if (lte->resource_location == RESOURCE_IN_WIM && lte->wim == args->wim) {
1503 /* We can't do an in place overwrite on the WIM if there are
1504 * streams after the XML data. */
1505 if (lte->resource_entry.offset +
1506 lte->resource_entry.size > args->end_offset)
1508 #ifdef ENABLE_ERROR_MESSAGES
1509 ERROR("The following resource is after the XML data:");
1510 print_lookup_table_entry(lte, stderr);
1512 return WIMLIB_ERR_RESOURCE_ORDER;
1514 copy_resource_entry(<e->output_resource_entry,
1515 <e->resource_entry);
1520 /* Given a WIM that we are going to overwrite in place with zero or more
1521 * additional streams added, construct a list the list of new unique streams
1522 * ('struct wim_lookup_table_entry's) that must be written, plus any unhashed
1523 * streams that need to be added but may be identical to other hashed or
1524 * unhashed streams. These unhashed streams are checksummed while the streams
1525 * are being written. To aid this process, the member @unique_size is set to 1
1526 * on streams that have a unique size and therefore must be written.
1528 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1529 * indicate the number of times the stream is referenced in only the streams
1530 * that are being written; this may still be adjusted later when unhashed
1531 * streams are being resolved.
1534 prepare_streams_for_overwrite(WIMStruct *wim, off_t end_offset,
1535 struct list_head *stream_list)
1538 struct lte_overwrite_prepare_args args;
1542 args.end_offset = end_offset;
1543 ret = init_stream_size_table(&args.stream_size_tab,
1544 wim->lookup_table->capacity);
1548 INIT_LIST_HEAD(&args.stream_list);
1549 for (i = 0; i < wim->hdr.image_count; i++) {
1550 struct wim_image_metadata *imd;
1551 struct wim_lookup_table_entry *lte;
1553 imd = wim->image_metadata[i];
1554 image_for_each_unhashed_stream(lte, imd)
1555 lte_overwrite_prepare(lte, &args);
1557 for_lookup_table_entry(wim->lookup_table, lte_overwrite_prepare, &args);
1558 list_transfer(&args.stream_list, stream_list);
1560 for (i = 0; i < wim->hdr.image_count; i++) {
1561 ret = lte_overwrite_prepare_2(wim->image_metadata[i]->metadata_lte,
1564 goto out_destroy_stream_size_table;
1566 ret = for_lookup_table_entry(wim->lookup_table,
1567 lte_overwrite_prepare_2, &args);
1568 out_destroy_stream_size_table:
1569 destroy_stream_size_table(&args.stream_size_tab);
1574 struct find_streams_ctx {
1575 struct list_head stream_list;
1576 struct stream_size_table stream_size_tab;
1580 inode_find_streams_to_write(struct wim_inode *inode,
1581 struct wim_lookup_table *table,
1582 struct list_head *stream_list,
1583 struct stream_size_table *tab)
1585 struct wim_lookup_table_entry *lte;
1586 for (unsigned i = 0; i <= inode->i_num_ads; i++) {
1587 lte = inode_stream_lte(inode, i, table);
1589 if (lte->out_refcnt == 0) {
1591 stream_size_table_insert(lte, tab);
1592 list_add_tail(<e->write_streams_list, stream_list);
1594 lte->out_refcnt += inode->i_nlink;
1600 image_find_streams_to_write(WIMStruct *w)
1602 struct find_streams_ctx *ctx;
1603 struct wim_image_metadata *imd;
1604 struct wim_inode *inode;
1605 struct wim_lookup_table_entry *lte;
1608 imd = wim_get_current_image_metadata(w);
1610 image_for_each_unhashed_stream(lte, imd)
1611 lte->out_refcnt = 0;
1613 /* Go through this image's inodes to find any streams that have not been
1615 image_for_each_inode(inode, imd) {
1616 inode_find_streams_to_write(inode, w->lookup_table,
1618 &ctx->stream_size_tab);
1623 /* Given a WIM that from which one or all of the images is being written, build
1624 * the list of unique streams ('struct wim_lookup_table_entry's) that must be
1625 * written, plus any unhashed streams that need to be written but may be
1626 * identical to other hashed or unhashed streams being written. These unhashed
1627 * streams are checksummed while the streams are being written. To aid this
1628 * process, the member @unique_size is set to 1 on streams that have a unique
1629 * size and therefore must be written.
1631 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1632 * indicate the number of times the stream is referenced in only the streams
1633 * that are being written; this may still be adjusted later when unhashed
1634 * streams are being resolved.
1637 prepare_stream_list(WIMStruct *wim, int image, struct list_head *stream_list)
1640 struct find_streams_ctx ctx;
1642 for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
1643 ret = init_stream_size_table(&ctx.stream_size_tab,
1644 wim->lookup_table->capacity);
1647 for_lookup_table_entry(wim->lookup_table, stream_size_table_insert,
1648 &ctx.stream_size_tab);
1649 INIT_LIST_HEAD(&ctx.stream_list);
1650 wim->private = &ctx;
1651 ret = for_image(wim, image, image_find_streams_to_write);
1652 destroy_stream_size_table(&ctx.stream_size_tab);
1654 list_transfer(&ctx.stream_list, stream_list);
1658 /* Writes the streams for the specified @image in @wim to @wim->out_fp.
1661 write_wim_streams(WIMStruct *wim, int image, int write_flags,
1662 unsigned num_threads,
1663 wimlib_progress_func_t progress_func)
1666 struct list_head stream_list;
1668 ret = prepare_stream_list(wim, image, &stream_list);
1671 return write_stream_list(&stream_list,
1674 wimlib_get_compression_type(wim),
1681 * Finish writing a WIM file: write the lookup table, xml data, and integrity
1682 * table (optional), then overwrite the WIM header.
1684 * write_flags is a bitwise OR of the following:
1686 * (public) WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
1687 * Include an integrity table.
1689 * (public) WIMLIB_WRITE_FLAG_SHOW_PROGRESS:
1690 * Show progress information when (if) writing the integrity table.
1692 * (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
1693 * Don't write the lookup table.
1695 * (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
1696 * When (if) writing the integrity table, re-use entries from the
1697 * existing integrity table, if possible.
1699 * (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
1700 * After writing the XML data but before writing the integrity
1701 * table, write a temporary WIM header and flush the stream so that
1702 * the WIM is less likely to become corrupted upon abrupt program
1705 * (private) WIMLIB_WRITE_FLAG_FSYNC:
1706 * fsync() the output file before closing it.
1710 finish_write(WIMStruct *w, int image, int write_flags,
1711 wimlib_progress_func_t progress_func)
1714 struct wim_header hdr;
1715 FILE *out = w->out_fp;
1717 /* @hdr will be the header for the new WIM. First copy all the data
1718 * from the header in the WIMStruct; then set all the fields that may
1719 * have changed, including the resource entries, boot index, and image
1721 memcpy(&hdr, &w->hdr, sizeof(struct wim_header));
1723 /* Set image count and boot index correctly for single image writes */
1724 if (image != WIMLIB_ALL_IMAGES) {
1725 hdr.image_count = 1;
1726 if (hdr.boot_idx == image)
1732 /* In the WIM header, there is room for the resource entry for a
1733 * metadata resource labeled as the "boot metadata". This entry should
1734 * be zeroed out if there is no bootable image (boot_idx 0). Otherwise,
1735 * it should be a copy of the resource entry for the image that is
1736 * marked as bootable. This is not well documented... */
1737 if (hdr.boot_idx == 0) {
1738 zero_resource_entry(&hdr.boot_metadata_res_entry);
1740 copy_resource_entry(&hdr.boot_metadata_res_entry,
1741 &w->image_metadata[ hdr.boot_idx- 1
1742 ]->metadata_lte->output_resource_entry);
1745 if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
1746 ret = write_lookup_table(w, image, &hdr.lookup_table_res_entry);
1751 ret = write_xml_data(w->wim_info, image, out,
1752 (write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE) ?
1753 wim_info_get_total_bytes(w->wim_info) : 0,
1754 &hdr.xml_res_entry);
1758 if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
1759 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
1760 struct wim_header checkpoint_hdr;
1761 memcpy(&checkpoint_hdr, &hdr, sizeof(struct wim_header));
1762 zero_resource_entry(&checkpoint_hdr.integrity);
1763 if (fseeko(out, 0, SEEK_SET)) {
1764 ERROR_WITH_ERRNO("Failed to seek to beginning "
1765 "of WIM being written");
1766 ret = WIMLIB_ERR_WRITE;
1769 ret = write_header(&checkpoint_hdr, out);
1773 if (fflush(out) != 0) {
1774 ERROR_WITH_ERRNO("Can't write data to WIM");
1775 ret = WIMLIB_ERR_WRITE;
1779 if (fseeko(out, 0, SEEK_END) != 0) {
1780 ERROR_WITH_ERRNO("Failed to seek to end "
1781 "of WIM being written");
1782 ret = WIMLIB_ERR_WRITE;
1787 off_t old_lookup_table_end;
1788 off_t new_lookup_table_end;
1789 if (write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE) {
1790 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
1791 w->hdr.lookup_table_res_entry.size;
1793 old_lookup_table_end = 0;
1795 new_lookup_table_end = hdr.lookup_table_res_entry.offset +
1796 hdr.lookup_table_res_entry.size;
1798 ret = write_integrity_table(out,
1800 new_lookup_table_end,
1801 old_lookup_table_end,
1806 zero_resource_entry(&hdr.integrity);
1809 if (fseeko(out, 0, SEEK_SET) != 0) {
1810 ERROR_WITH_ERRNO("Failed to seek to beginning of WIM "
1812 ret = WIMLIB_ERR_WRITE;
1816 ret = write_header(&hdr, out);
1820 if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
1821 if (fflush(out) != 0
1822 || fsync(fileno(out)) != 0)
1824 ERROR_WITH_ERRNO("Error flushing data to WIM file");
1825 ret = WIMLIB_ERR_WRITE;
1829 if (fclose(out) != 0) {
1830 ERROR_WITH_ERRNO("Failed to close the output WIM file");
1832 ret = WIMLIB_ERR_WRITE;
1838 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
1840 lock_wim(WIMStruct *w, FILE *fp)
1843 if (fp && !w->wim_locked) {
1844 ret = flock(fileno(fp), LOCK_EX | LOCK_NB);
1846 if (errno == EWOULDBLOCK) {
1847 ERROR("`%"TS"' is already being modified or has been "
1848 "mounted read-write\n"
1849 " by another process!", w->filename);
1850 ret = WIMLIB_ERR_ALREADY_LOCKED;
1852 WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
1865 open_wim_writable(WIMStruct *w, const tchar *path,
1866 bool trunc, bool also_readable)
1877 wimlib_assert(w->out_fp == NULL);
1878 w->out_fp = tfopen(path, mode);
1882 ERROR_WITH_ERRNO("Failed to open `%"TS"' for writing", path);
1883 return WIMLIB_ERR_OPEN;
1889 close_wim_writable(WIMStruct *w)
1892 if (fclose(w->out_fp) != 0) {
1893 WARNING_WITH_ERRNO("Failed to close output WIM");
1899 /* Open file stream and write dummy header for WIM. */
1901 begin_write(WIMStruct *w, const tchar *path, int write_flags)
1904 ret = open_wim_writable(w, path, true,
1905 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
1908 /* Write dummy header. It will be overwritten later. */
1909 return write_header(&w->hdr, w->out_fp);
1912 /* Writes a stand-alone WIM to a file. */
1914 wimlib_write(WIMStruct *w, const tchar *path,
1915 int image, int write_flags, unsigned num_threads,
1916 wimlib_progress_func_t progress_func)
1921 return WIMLIB_ERR_INVALID_PARAM;
1923 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1925 if (image != WIMLIB_ALL_IMAGES &&
1926 (image < 1 || image > w->hdr.image_count))
1927 return WIMLIB_ERR_INVALID_IMAGE;
1929 if (w->hdr.total_parts != 1) {
1930 ERROR("Cannot call wimlib_write() on part of a split WIM");
1931 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1934 ret = begin_write(w, path, write_flags);
1938 ret = write_wim_streams(w, image, write_flags, num_threads,
1944 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
1946 ret = for_image(w, image, write_metadata_resource);
1951 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
1953 ret = finish_write(w, image, write_flags, progress_func);
1954 /* finish_write() closed the WIM for us */
1957 close_wim_writable(w);
1959 DEBUG("wimlib_write(path=%"TS") = %d", path, ret);
1964 any_images_modified(WIMStruct *w)
1966 for (int i = 0; i < w->hdr.image_count; i++)
1967 if (w->image_metadata[i]->modified)
1973 * Overwrite a WIM, possibly appending streams to it.
1975 * A WIM looks like (or is supposed to look like) the following:
1977 * Header (212 bytes)
1978 * Streams and metadata resources (variable size)
1979 * Lookup table (variable size)
1980 * XML data (variable size)
1981 * Integrity table (optional) (variable size)
1983 * If we are not adding any streams or metadata resources, the lookup table is
1984 * unchanged--- so we only need to overwrite the XML data, integrity table, and
1985 * header. This operation is potentially unsafe if the program is abruptly
1986 * terminated while the XML data or integrity table are being overwritten, but
1987 * before the new header has been written. To partially alleviate this problem,
1988 * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
1989 * finish_write() to cause a temporary WIM header to be written after the XML
1990 * data has been written. This may prevent the WIM from becoming corrupted if
1991 * the program is terminated while the integrity table is being calculated (but
1992 * no guarantees, due to write re-ordering...).
1994 * If we are adding new streams or images (metadata resources), the lookup table
1995 * needs to be changed, and those streams need to be written. In this case, we
1996 * try to perform a safe update of the WIM file by writing the streams *after*
1997 * the end of the previous WIM, then writing the new lookup table, XML data, and
1998 * (optionally) integrity table following the new streams. This will produce a
1999 * layout like the following:
2001 * Header (212 bytes)
2002 * (OLD) Streams and metadata resources (variable size)
2003 * (OLD) Lookup table (variable size)
2004 * (OLD) XML data (variable size)
2005 * (OLD) Integrity table (optional) (variable size)
2006 * (NEW) Streams and metadata resources (variable size)
2007 * (NEW) Lookup table (variable size)
2008 * (NEW) XML data (variable size)
2009 * (NEW) Integrity table (optional) (variable size)
2011 * At all points, the WIM is valid as nothing points to the new data yet. Then,
2012 * the header is overwritten to point to the new lookup table, XML data, and
2013 * integrity table, to produce the following layout:
2015 * Header (212 bytes)
2016 * Streams and metadata resources (variable size)
2017 * Nothing (variable size)
2018 * More Streams and metadata resources (variable size)
2019 * Lookup table (variable size)
2020 * XML data (variable size)
2021 * Integrity table (optional) (variable size)
2023 * This method allows an image to be appended to a large WIM very quickly, and
2024 * is is crash-safe except in the case of write re-ordering, but the
2025 * disadvantage is that a small hole is left in the WIM where the old lookup
2026 * table, xml data, and integrity table were. (These usually only take up a
2027 * small amount of space compared to the streams, however.)
2030 overwrite_wim_inplace(WIMStruct *w, int write_flags,
2031 unsigned num_threads,
2032 wimlib_progress_func_t progress_func)
2035 struct list_head stream_list;
2037 u64 old_lookup_table_end, old_xml_begin, old_xml_end;
2039 DEBUG("Overwriting `%"TS"' in-place", w->filename);
2041 /* Make sure that the integrity table (if present) is after the XML
2042 * data, and that there are no stream resources, metadata resources, or
2043 * lookup tables after the XML data. Otherwise, these data would be
2045 old_xml_begin = w->hdr.xml_res_entry.offset;
2046 old_xml_end = old_xml_begin + w->hdr.xml_res_entry.size;
2047 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
2048 w->hdr.lookup_table_res_entry.size;
2049 if (w->hdr.integrity.offset != 0 && w->hdr.integrity.offset < old_xml_end) {
2050 ERROR("Didn't expect the integrity table to be before the XML data");
2051 return WIMLIB_ERR_RESOURCE_ORDER;
2054 if (old_lookup_table_end > old_xml_begin) {
2055 ERROR("Didn't expect the lookup table to be after the XML data");
2056 return WIMLIB_ERR_RESOURCE_ORDER;
2059 /* Set @old_wim_end, which indicates the point beyond which we don't
2060 * allow any file and metadata resources to appear without returning
2061 * WIMLIB_ERR_RESOURCE_ORDER (due to the fact that we would otherwise
2062 * overwrite these resources). */
2063 if (!w->deletion_occurred && !any_images_modified(w)) {
2064 /* If no images have been modified and no images have been
2065 * deleted, a new lookup table does not need to be written. We
2066 * shall write the new XML data and optional integrity table
2067 * immediately after the lookup table. Note that this may
2068 * overwrite an existing integrity table. */
2069 DEBUG("Skipping writing lookup table "
2070 "(no images modified or deleted)");
2071 old_wim_end = old_lookup_table_end;
2072 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
2073 WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
2074 } else if (w->hdr.integrity.offset) {
2075 /* Old WIM has an integrity table; begin writing new streams
2077 old_wim_end = w->hdr.integrity.offset + w->hdr.integrity.size;
2079 /* No existing integrity table; begin writing new streams after
2080 * the old XML data. */
2081 old_wim_end = old_xml_end;
2084 ret = prepare_streams_for_overwrite(w, old_wim_end, &stream_list);
2088 ret = open_wim_writable(w, w->filename, false,
2089 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
2093 ret = lock_wim(w, w->out_fp);
2095 close_wim_writable(w);
2099 if (fseeko(w->out_fp, old_wim_end, SEEK_SET) != 0) {
2100 ERROR_WITH_ERRNO("Can't seek to end of WIM");
2101 close_wim_writable(w);
2103 return WIMLIB_ERR_WRITE;
2106 DEBUG("Writing newly added streams (offset = %"PRIu64")",
2108 ret = write_stream_list(&stream_list,
2111 wimlib_get_compression_type(w),
2118 for (int i = 0; i < w->hdr.image_count; i++) {
2119 if (w->image_metadata[i]->modified) {
2120 select_wim_image(w, i + 1);
2121 ret = write_metadata_resource(w);
2126 write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
2127 ret = finish_write(w, WIMLIB_ALL_IMAGES, write_flags,
2130 close_wim_writable(w);
2131 if (ret != 0 && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
2132 WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
2133 w->filename, old_wim_end);
2134 /* Return value of truncate() is ignored because this is already
2136 (void)ttruncate(w->filename, old_wim_end);
2143 overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
2144 unsigned num_threads,
2145 wimlib_progress_func_t progress_func)
2147 size_t wim_name_len;
2150 DEBUG("Overwriting `%"TS"' via a temporary file", w->filename);
2152 /* Write the WIM to a temporary file in the same directory as the
2154 wim_name_len = tstrlen(w->filename);
2155 tchar tmpfile[wim_name_len + 10];
2156 tmemcpy(tmpfile, w->filename, wim_name_len);
2157 randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
2158 tmpfile[wim_name_len + 9] = T('\0');
2160 ret = wimlib_write(w, tmpfile, WIMLIB_ALL_IMAGES,
2161 write_flags | WIMLIB_WRITE_FLAG_FSYNC,
2162 num_threads, progress_func);
2164 ERROR("Failed to write the WIM file `%"TS"'", tmpfile);
2168 DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, w->filename);
2171 /* Windows won't let you delete open files unless FILE_SHARE_DELETE was
2172 * specified to CreateFile(). The WIM was opened with fopen(), which
2173 * didn't provided this flag to CreateFile, so the handle must be closed
2174 * before executing the rename(). */
2175 if (w->fp != NULL) {
2181 /* Rename the new file to the old file .*/
2182 if (trename(tmpfile, w->filename) != 0) {
2183 ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
2184 tmpfile, w->filename);
2185 ret = WIMLIB_ERR_RENAME;
2189 if (progress_func) {
2190 union wimlib_progress_info progress;
2191 progress.rename.from = tmpfile;
2192 progress.rename.to = w->filename;
2193 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
2196 /* Close the original WIM file that was opened for reading. */
2197 if (w->fp != NULL) {
2202 /* Re-open the WIM read-only. */
2203 w->fp = tfopen(w->filename, T("rb"));
2204 if (w->fp == NULL) {
2205 ret = WIMLIB_ERR_REOPEN;
2206 WARNING_WITH_ERRNO("Failed to re-open `%"TS"' read-only",
2213 /* Remove temporary file. */
2214 if (tunlink(tmpfile) != 0)
2215 WARNING_WITH_ERRNO("Failed to remove `%"TS"'", tmpfile);
2221 * Writes a WIM file to the original file that it was read from, overwriting it.
2224 wimlib_overwrite(WIMStruct *w, int write_flags,
2225 unsigned num_threads,
2226 wimlib_progress_func_t progress_func)
2228 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2231 return WIMLIB_ERR_NO_FILENAME;
2233 if (w->hdr.total_parts != 1) {
2234 ERROR("Cannot modify a split WIM");
2235 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
2238 if ((!w->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
2239 && !(write_flags & WIMLIB_WRITE_FLAG_REBUILD))
2242 ret = overwrite_wim_inplace(w, write_flags, num_threads,
2244 if (ret == WIMLIB_ERR_RESOURCE_ORDER)
2245 WARNING("Falling back to re-building entire WIM");
2249 return overwrite_wim_via_tmpfile(w, write_flags, num_threads,