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)
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 finalize_and_check_sha1(SHA_CTX *sha_ctx, struct wim_lookup_table_entry *lte)
251 u8 md[SHA1_HASH_SIZE];
252 sha1_final(md, sha_ctx);
254 copy_hash(lte->hash, md);
255 } else if (!hashes_equal(md, lte->hash)) {
256 ERROR("WIM resource has incorrect hash!");
257 if (lte_filename_valid(lte)) {
258 ERROR("We were reading it from \"%"TS"\"; maybe "
259 "it changed while we were reading it.",
262 return WIMLIB_ERR_INVALID_RESOURCE_HASH;
268 struct write_resource_ctx {
269 compress_func_t compress;
270 struct chunk_table *chunk_tab;
277 write_resource_cb(const void *chunk, size_t chunk_size, void *_ctx)
279 struct write_resource_ctx *ctx = _ctx;
282 sha1_update(&ctx->sha_ctx, chunk, chunk_size);
285 return write_wim_resource_chunk(chunk, chunk_size,
286 ctx->out_fp, ctx->compress,
289 if (fwrite(chunk, 1, chunk_size, ctx->out_fp) != chunk_size) {
290 ERROR_WITH_ERRNO("Error writing to output WIM");
291 return WIMLIB_ERR_WRITE;
299 * Write a resource to an output WIM.
301 * @lte: Lookup table entry for the resource, which could be in another WIM,
302 * in an external file, or in another location.
304 * @out_fp: FILE * opened to the output WIM.
306 * @out_ctype: One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate
307 * which compression algorithm to use.
309 * @out_res_entry: On success, this is filled in with the offset, flags,
310 * compressed size, and uncompressed size of the resource
313 * @flags: WIMLIB_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed
314 * even if it could otherwise be copied directly from the input.
316 * Additional notes: The SHA1 message digest of the uncompressed data is
317 * calculated (except when doing a raw copy --- see below). If the @unhashed
318 * flag is set on the lookup table entry, this message digest is simply copied
319 * to it; otherwise, the message digest is compared with the existing one, and
320 * the function will fail if they do not match.
323 write_wim_resource(struct wim_lookup_table_entry *lte,
324 FILE *out_fp, int out_ctype,
325 struct resource_entry *out_res_entry,
328 struct write_resource_ctx write_ctx;
334 flags &= ~WIMLIB_RESOURCE_FLAG_RECOMPRESS;
336 /* Get current position in output WIM */
337 offset = ftello(out_fp);
339 ERROR_WITH_ERRNO("Can't get position in output WIM");
340 return WIMLIB_ERR_WRITE;
343 /* If we are not forcing the data to be recompressed, and the input
344 * resource is located in a WIM with the same compression type as that
345 * desired other than no compression, we can simply copy the compressed
346 * data without recompressing it. This also means we must skip
347 * calculating the SHA1, as we never will see the uncompressed data. */
348 if (!(flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
349 lte->resource_location == RESOURCE_IN_WIM &&
350 out_ctype != WIMLIB_COMPRESSION_TYPE_NONE &&
351 wimlib_get_compression_type(lte->wim) == out_ctype)
353 flags |= WIMLIB_RESOURCE_FLAG_RAW;
354 write_ctx.doing_sha = false;
355 read_size = lte->resource_entry.size;
357 write_ctx.doing_sha = true;
358 sha1_init(&write_ctx.sha_ctx);
359 read_size = lte->resource_entry.original_size;
362 /* Initialize the chunk table and set the compression function if
363 * compressing the resource. */
364 if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
365 (flags & WIMLIB_RESOURCE_FLAG_RAW)) {
366 write_ctx.compress = NULL;
367 write_ctx.chunk_tab = NULL;
369 write_ctx.compress = get_compress_func(out_ctype);
370 ret = begin_wim_resource_chunk_tab(lte, out_fp,
372 &write_ctx.chunk_tab);
377 /* Write the entire resource by reading the entire resource and feeding
378 * the data through the write_resource_cb function. */
379 write_ctx.out_fp = out_fp;
381 ret = read_resource_prefix(lte, read_size,
382 write_resource_cb, &write_ctx, flags);
384 goto out_free_chunk_tab;
386 /* Verify SHA1 message digest of the resource, or set the hash for the
388 if (write_ctx.doing_sha) {
389 ret = finalize_and_check_sha1(&write_ctx.sha_ctx, lte);
391 goto out_free_chunk_tab;
394 out_res_entry->flags = lte->resource_entry.flags;
395 out_res_entry->original_size = wim_resource_size(lte);
396 out_res_entry->offset = offset;
397 if (flags & WIMLIB_RESOURCE_FLAG_RAW) {
398 /* Doing a raw write: The new compressed size is the same as
399 * the compressed size in the other WIM. */
400 new_size = lte->resource_entry.size;
401 } else if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE) {
402 /* Using WIMLIB_COMPRESSION_TYPE_NONE: The new compressed size
403 * is the original size. */
404 new_size = lte->resource_entry.original_size;
405 out_res_entry->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
407 /* Using a different compression type: Call
408 * finish_wim_resource_chunk_tab() and it will provide the new
409 * compressed size. */
410 ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab, out_fp,
413 goto out_free_chunk_tab;
414 if (new_size >= wim_resource_size(lte)) {
415 /* Oops! We compressed the resource to larger than the original
416 * size. Write the resource uncompressed instead. */
417 if (fseeko(out_fp, offset, SEEK_SET) ||
419 ftruncate(fileno(out_fp),
420 offset + wim_resource_size(lte)))
422 ERROR_WITH_ERRNO("Failed to flush and/or truncate "
424 ret = WIMLIB_ERR_WRITE;
425 goto out_free_chunk_tab;
427 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
428 "writing uncompressed instead",
429 wim_resource_size(lte), new_size);
430 write_ctx.compress = NULL;
431 write_ctx.doing_sha = false;
432 out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
433 goto try_write_again;
435 out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
437 out_res_entry->size = new_size;
440 FREE(write_ctx.chunk_tab);
444 #ifdef ENABLE_MULTITHREADED_COMPRESSION
446 /* Blocking shared queue (solves the producer-consumer problem) */
447 struct shared_queue {
451 unsigned filled_slots;
453 pthread_mutex_t lock;
454 pthread_cond_t msg_avail_cond;
455 pthread_cond_t space_avail_cond;
459 shared_queue_init(struct shared_queue *q, unsigned size)
461 wimlib_assert(size != 0);
462 q->array = CALLOC(sizeof(q->array[0]), size);
469 if (pthread_mutex_init(&q->lock, NULL)) {
470 ERROR_WITH_ERRNO("Failed to initialize mutex");
473 if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
474 ERROR_WITH_ERRNO("Failed to initialize condition variable");
475 goto err_destroy_lock;
477 if (pthread_cond_init(&q->space_avail_cond, NULL)) {
478 ERROR_WITH_ERRNO("Failed to initialize condition variable");
479 goto err_destroy_msg_avail_cond;
482 err_destroy_msg_avail_cond:
483 pthread_cond_destroy(&q->msg_avail_cond);
485 pthread_mutex_destroy(&q->lock);
487 return WIMLIB_ERR_NOMEM;
491 shared_queue_destroy(struct shared_queue *q)
494 pthread_mutex_destroy(&q->lock);
495 pthread_cond_destroy(&q->msg_avail_cond);
496 pthread_cond_destroy(&q->space_avail_cond);
500 shared_queue_put(struct shared_queue *q, void *obj)
502 pthread_mutex_lock(&q->lock);
503 while (q->filled_slots == q->size)
504 pthread_cond_wait(&q->space_avail_cond, &q->lock);
506 q->back = (q->back + 1) % q->size;
507 q->array[q->back] = obj;
510 pthread_cond_broadcast(&q->msg_avail_cond);
511 pthread_mutex_unlock(&q->lock);
515 shared_queue_get(struct shared_queue *q)
519 pthread_mutex_lock(&q->lock);
520 while (q->filled_slots == 0)
521 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
523 obj = q->array[q->front];
524 q->array[q->front] = NULL;
525 q->front = (q->front + 1) % q->size;
528 pthread_cond_broadcast(&q->space_avail_cond);
529 pthread_mutex_unlock(&q->lock);
533 struct compressor_thread_params {
534 struct shared_queue *res_to_compress_queue;
535 struct shared_queue *compressed_res_queue;
536 compress_func_t compress;
539 #define MAX_CHUNKS_PER_MSG 2
542 struct wim_lookup_table_entry *lte;
543 u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
544 u8 *out_compressed_chunks[MAX_CHUNKS_PER_MSG];
545 u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
546 unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
547 unsigned compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
549 struct list_head list;
555 compress_chunks(struct message *msg, compress_func_t compress)
557 for (unsigned i = 0; i < msg->num_chunks; i++) {
558 unsigned len = compress(msg->uncompressed_chunks[i],
559 msg->uncompressed_chunk_sizes[i],
560 msg->compressed_chunks[i]);
562 /* To be written compressed */
563 msg->out_compressed_chunks[i] = msg->compressed_chunks[i];
564 msg->compressed_chunk_sizes[i] = len;
566 /* To be written uncompressed */
567 msg->out_compressed_chunks[i] = msg->uncompressed_chunks[i];
568 msg->compressed_chunk_sizes[i] = msg->uncompressed_chunk_sizes[i];
574 /* Compressor thread routine. This is a lot simpler than the main thread
575 * routine: just repeatedly get a group of chunks from the
576 * res_to_compress_queue, compress them, and put them in the
577 * compressed_res_queue. A NULL pointer indicates that the thread should stop.
580 compressor_thread_proc(void *arg)
582 struct compressor_thread_params *params = arg;
583 struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
584 struct shared_queue *compressed_res_queue = params->compressed_res_queue;
585 compress_func_t compress = params->compress;
588 DEBUG("Compressor thread ready");
589 while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
590 compress_chunks(msg, compress);
591 shared_queue_put(compressed_res_queue, msg);
593 DEBUG("Compressor thread terminating");
596 #endif /* ENABLE_MULTITHREADED_COMPRESSION */
599 do_write_streams_progress(union wimlib_progress_info *progress,
600 wimlib_progress_func_t progress_func,
603 progress->write_streams.completed_bytes += size_added;
604 progress->write_streams.completed_streams++;
606 progress->write_streams.completed_bytes >= progress->write_streams._private)
608 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
610 if (progress->write_streams._private == progress->write_streams.total_bytes) {
611 progress->write_streams._private = ~0;
613 progress->write_streams._private =
614 min(progress->write_streams.total_bytes,
615 progress->write_streams.completed_bytes +
616 progress->write_streams.total_bytes / 100);
621 struct serial_write_stream_ctx {
624 int write_resource_flags;
628 serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
630 struct serial_write_stream_ctx *ctx = _ctx;
631 return write_wim_resource(lte, ctx->out_fp,
632 ctx->out_ctype, <e->output_resource_entry,
633 ctx->write_resource_flags);
636 /* Write a list of streams, taking into account that some streams may be
637 * duplicates that are checksummed and discarded on the fly, and also delegating
638 * the actual writing of a stream to a function @write_stream_cb, which is
639 * passed the context @write_stream_ctx. */
641 do_write_stream_list(struct list_head *stream_list,
642 struct wim_lookup_table *lookup_table,
643 int (*write_stream_cb)(struct wim_lookup_table_entry *, void *),
644 void *write_stream_ctx,
645 wimlib_progress_func_t progress_func,
646 union wimlib_progress_info *progress)
649 struct wim_lookup_table_entry *lte;
651 /* For each stream in @stream_list ... */
652 while (!list_empty(stream_list)) {
653 lte = container_of(stream_list->next,
654 struct wim_lookup_table_entry,
656 list_del(<e->write_streams_list);
657 if (lte->unhashed && !lte->unique_size) {
658 /* Unhashed stream that shares a size with some other
659 * stream in the WIM we are writing. The stream must be
660 * checksummed to know if we need to write it or not. */
661 struct wim_lookup_table_entry *tmp;
662 u32 orig_refcnt = lte->out_refcnt;
664 ret = hash_unhashed_stream(lte, lookup_table, &tmp);
669 /* We found a duplicate stream. */
670 if (orig_refcnt != tmp->out_refcnt) {
671 /* We have already written, or are going
672 * to write, the duplicate stream. So
673 * just skip to the next stream. */
674 DEBUG("Discarding duplicate stream of length %"PRIu64,
675 wim_resource_size(lte));
676 goto skip_to_progress;
681 /* Here, @lte is either a hashed stream or an unhashed stream
682 * with a unique size. In either case we know that the stream
683 * has to be written. In either case the SHA1 message digest
684 * will be calculated over the stream while writing it; however,
685 * in the former case this is done merely to check the data,
686 * while in the latter case this is done because we do not have
687 * the SHA1 message digest yet. */
688 wimlib_assert(lte->out_refcnt != 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);
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 progress->write_streams.num_threads = 1;
752 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
753 return do_write_stream_list_serial(stream_list,
757 write_resource_flags,
762 #ifdef ENABLE_MULTITHREADED_COMPRESSION
764 write_wim_chunks(struct message *msg, FILE *out_fp,
765 struct chunk_table *chunk_tab)
767 for (unsigned i = 0; i < msg->num_chunks; i++) {
768 unsigned chunk_csize = msg->compressed_chunk_sizes[i];
770 if (fwrite(msg->out_compressed_chunks[i], 1, chunk_csize, out_fp)
773 ERROR_WITH_ERRNO("Failed to write WIM chunk");
774 return WIMLIB_ERR_WRITE;
777 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
778 chunk_tab->cur_offset += chunk_csize;
783 struct main_writer_thread_ctx {
784 struct list_head *stream_list;
785 struct wim_lookup_table *lookup_table;
788 int write_resource_flags;
789 struct shared_queue *res_to_compress_queue;
790 struct shared_queue *compressed_res_queue;
792 wimlib_progress_func_t progress_func;
793 union wimlib_progress_info *progress;
795 struct list_head available_msgs;
796 struct list_head outstanding_streams;
797 struct list_head serial_streams;
798 size_t num_outstanding_messages;
800 SHA_CTX next_sha_ctx;
803 struct wim_lookup_table_entry *next_lte;
805 struct message *msgs;
806 struct message *next_msg;
807 struct chunk_table *cur_chunk_tab;
811 init_message(struct message *msg)
813 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
814 msg->compressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
815 msg->uncompressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
816 if (msg->compressed_chunks[i] == NULL ||
817 msg->uncompressed_chunks[i] == NULL)
818 return WIMLIB_ERR_NOMEM;
824 destroy_message(struct message *msg)
826 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
827 FREE(msg->compressed_chunks[i]);
828 FREE(msg->uncompressed_chunks[i]);
833 free_messages(struct message *msgs, size_t num_messages)
836 for (size_t i = 0; i < num_messages; i++)
837 destroy_message(&msgs[i]);
842 static struct message *
843 allocate_messages(size_t num_messages)
845 struct message *msgs;
847 msgs = CALLOC(num_messages, sizeof(struct message));
850 for (size_t i = 0; i < num_messages; i++) {
851 if (init_message(&msgs[i])) {
852 free_messages(msgs, num_messages);
860 main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
862 while (ctx->num_outstanding_messages--)
863 shared_queue_get(ctx->compressed_res_queue);
864 free_messages(ctx->msgs, ctx->num_messages);
865 FREE(ctx->cur_chunk_tab);
869 main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
871 /* Pre-allocate all the buffers that will be needed to do the chunk
873 ctx->msgs = allocate_messages(ctx->num_messages);
875 return WIMLIB_ERR_NOMEM;
877 /* Initially, all the messages are available to use. */
878 INIT_LIST_HEAD(&ctx->available_msgs);
879 for (size_t i = 0; i < ctx->num_messages; i++)
880 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
882 /* outstanding_streams is the list of streams that currently have had
883 * chunks sent off for compression.
885 * The first stream in outstanding_streams is the stream that is
886 * currently being written.
888 * The last stream in outstanding_streams is the stream that is
889 * currently being read and having chunks fed to the compressor threads.
891 INIT_LIST_HEAD(&ctx->outstanding_streams);
892 ctx->num_outstanding_messages = 0;
894 ctx->next_msg = NULL;
896 /* Resources that don't need any chunks compressed are added to this
897 * list and written directly by the main thread. */
898 INIT_LIST_HEAD(&ctx->serial_streams);
900 ctx->cur_chunk_tab = NULL;
906 receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
909 struct wim_lookup_table_entry *cur_lte;
912 wimlib_assert(!list_empty(&ctx->outstanding_streams));
913 wimlib_assert(ctx->num_outstanding_messages != 0);
915 cur_lte = container_of(ctx->outstanding_streams.next,
916 struct wim_lookup_table_entry,
917 being_compressed_list);
919 /* Get the next message from the queue and process it.
920 * The message will contain 1 or more data chunks that have been
922 msg = shared_queue_get(ctx->compressed_res_queue);
923 msg->complete = true;
924 --ctx->num_outstanding_messages;
926 /* Is this the next chunk in the current resource? If it's not
927 * (i.e., an earlier chunk in a same or different resource
928 * hasn't been compressed yet), do nothing, and keep this
929 * message around until all earlier chunks are received.
931 * Otherwise, write all the chunks we can. */
932 while (cur_lte != NULL &&
933 !list_empty(&cur_lte->msg_list)
934 && (msg = container_of(cur_lte->msg_list.next,
938 list_move(&msg->list, &ctx->available_msgs);
939 if (msg->begin_chunk == 0) {
940 /* This is the first set of chunks. Leave space
941 * for the chunk table in the output file. */
942 off_t cur_offset = ftello(ctx->out_fp);
943 if (cur_offset == -1)
944 return WIMLIB_ERR_WRITE;
945 ret = begin_wim_resource_chunk_tab(cur_lte,
948 &ctx->cur_chunk_tab);
953 /* Write the compressed chunks from the message. */
954 ret = write_wim_chunks(msg, ctx->out_fp, ctx->cur_chunk_tab);
958 /* Was this the last chunk of the stream? If so, finish
960 if (list_empty(&cur_lte->msg_list) &&
961 msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
964 ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
970 list_del(&cur_lte->being_compressed_list);
972 if (res_csize >= wim_resource_size(cur_lte)) {
973 /* Oops! We compressed the resource to
974 * larger than the original size. Write
975 * the resource uncompressed instead. */
976 ret = write_uncompressed_resource_and_truncate(
979 ctx->cur_chunk_tab->file_offset,
980 &cur_lte->output_resource_entry);
986 cur_lte->output_resource_entry.size =
989 cur_lte->output_resource_entry.original_size =
990 cur_lte->resource_entry.original_size;
992 cur_lte->output_resource_entry.offset =
993 ctx->cur_chunk_tab->file_offset;
995 cur_lte->output_resource_entry.flags =
996 cur_lte->resource_entry.flags |
997 WIM_RESHDR_FLAG_COMPRESSED;
999 do_write_streams_progress(ctx->progress, ctx->progress_func,
1000 wim_resource_size(cur_lte));
1001 FREE(ctx->cur_chunk_tab);
1002 ctx->cur_chunk_tab = NULL;
1004 /* Since we just finished writing a stream, write any
1005 * streams that have been added to the serial_streams
1006 * list for direct writing by the main thread (e.g.
1007 * resources that don't need to be compressed because
1008 * the desired compression type is the same as the
1009 * previous compression type). */
1010 ret = do_write_stream_list_serial(&ctx->serial_streams,
1014 ctx->write_resource_flags,
1020 /* Advance to the next stream to write. */
1021 if (list_empty(&ctx->outstanding_streams)) {
1024 cur_lte = container_of(ctx->outstanding_streams.next,
1025 struct wim_lookup_table_entry,
1026 being_compressed_list);
1033 /* Called when the main thread has read a new chunk of data. */
1035 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
1037 struct main_writer_thread_ctx *ctx = _ctx;
1039 struct message *next_msg;
1040 u64 next_chunk_in_msg;
1042 /* Update SHA1 message digest for the stream currently being read by the
1044 sha1_update(&ctx->next_sha_ctx, chunk, chunk_size);
1046 /* We send chunks of data to the compressor chunks in batches which we
1047 * refer to as "messages". @next_msg is the message that is currently
1048 * being prepared to send off. If it is NULL, that indicates that we
1049 * need to start a new message. */
1050 next_msg = ctx->next_msg;
1052 /* We need to start a new message. First check to see if there
1053 * is a message available in the list of available messages. If
1054 * so, we can just take one. If not, all the messages (there is
1055 * a fixed number of them, proportional to the number of
1056 * threads) have been sent off to the compressor threads, so we
1057 * receive messages from the compressor threads containing
1058 * compressed chunks of data.
1060 * We may need to receive multiple messages before one is
1061 * actually available to use because messages received that are
1062 * *not* for the very next set of chunks to compress must be
1063 * buffered until it's time to write those chunks. */
1064 while (list_empty(&ctx->available_msgs)) {
1065 ret = receive_compressed_chunks(ctx);
1070 next_msg = container_of(ctx->available_msgs.next,
1071 struct message, list);
1072 list_del(&next_msg->list);
1073 next_msg->complete = false;
1074 next_msg->begin_chunk = ctx->next_chunk;
1075 next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
1076 ctx->next_num_chunks - ctx->next_chunk);
1077 ctx->next_msg = next_msg;
1080 /* Fill in the next chunk to compress */
1081 next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;
1083 next_msg->uncompressed_chunk_sizes[next_chunk_in_msg] = chunk_size;
1084 memcpy(next_msg->uncompressed_chunks[next_chunk_in_msg],
1087 if (++next_chunk_in_msg == next_msg->num_chunks) {
1088 /* Send off an array of chunks to compress */
1089 list_add_tail(&next_msg->list, &ctx->next_lte->msg_list);
1090 shared_queue_put(ctx->res_to_compress_queue, next_msg);
1091 ++ctx->num_outstanding_messages;
1092 ctx->next_msg = NULL;
1098 main_writer_thread_finish(void *_ctx)
1100 struct main_writer_thread_ctx *ctx = _ctx;
1102 while (ctx->num_outstanding_messages != 0) {
1103 ret = receive_compressed_chunks(ctx);
1107 wimlib_assert(list_empty(&ctx->outstanding_streams));
1108 return do_write_stream_list_serial(&ctx->serial_streams,
1112 ctx->write_resource_flags,
1118 submit_stream_for_compression(struct wim_lookup_table_entry *lte,
1119 struct main_writer_thread_ctx *ctx)
1123 /* Read the entire stream @lte, feeding its data chunks to the
1124 * compressor threads. Also SHA1-sum the stream; this is required in
1125 * the case that @lte is unhashed, and a nice additional verification
1126 * when @lte is already hashed. */
1127 sha1_init(&ctx->next_sha_ctx);
1128 ctx->next_chunk = 0;
1129 ctx->next_num_chunks = wim_resource_chunks(lte);
1130 ctx->next_lte = lte;
1131 INIT_LIST_HEAD(<e->msg_list);
1132 list_add_tail(<e->being_compressed_list, &ctx->outstanding_streams);
1133 ret = read_resource_prefix(lte, wim_resource_size(lte),
1134 main_writer_thread_cb, ctx, 0);
1136 wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
1137 ret = finalize_and_check_sha1(&ctx->next_sha_ctx, lte);
1143 main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
1145 struct main_writer_thread_ctx *ctx = _ctx;
1148 if (wim_resource_size(lte) < 1000 ||
1149 ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
1150 (lte->resource_location == RESOURCE_IN_WIM &&
1151 !(ctx->write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
1152 wimlib_get_compression_type(lte->wim) == ctx->out_ctype))
1154 /* Stream is too small or isn't being compressed. Process it by
1155 * the main thread when we have a chance. We can't necessarily
1156 * process it right here, as the main thread could be in the
1157 * middle of writing a different stream. */
1158 list_add_tail(<e->write_streams_list, &ctx->serial_streams);
1162 ret = submit_stream_for_compression(lte, ctx);
1168 get_default_num_threads()
1171 return win32_get_number_of_processors();
1173 return sysconf(_SC_NPROCESSORS_ONLN);
1177 /* Equivalent to write_stream_list_serial(), except this takes a @num_threads
1178 * parameter and will perform compression using that many threads. Falls
1179 * back to write_stream_list_serial() on certain errors, such as a failure to
1180 * create the number of threads requested.
1182 * High level description of the algorithm for writing compressed streams in
1183 * parallel: We perform compression on chunks of size WIM_CHUNK_SIZE bytes
1184 * rather than on full files. The currently executing thread becomes the main
1185 * thread and is entirely in charge of reading the data to compress (which may
1186 * be in any location understood by the resource code--- such as in an external
1187 * file being captured, or in another WIM file from which an image is being
1188 * exported) and actually writing the compressed data to the output file.
1189 * Additional threads are "compressor threads" and all execute the
1190 * compressor_thread_proc, where they repeatedly retrieve buffers of data from
1191 * the main thread, compress them, and hand them back to the main thread.
1193 * Certain streams, such as streams that do not need to be compressed (e.g.
1194 * input compression type same as output compression type) or streams of very
1195 * small size are placed in a list (main_writer_thread_ctx.serial_list) and
1196 * handled entirely by the main thread at an appropriate time.
1198 * At any given point in time, multiple streams may be having chunks compressed
1199 * concurrently. The stream that the main thread is currently *reading* may be
1200 * later in the list that the stream that the main thread is currently
1204 write_stream_list_parallel(struct list_head *stream_list,
1205 struct wim_lookup_table *lookup_table,
1208 int write_resource_flags,
1209 wimlib_progress_func_t progress_func,
1210 union wimlib_progress_info *progress,
1211 unsigned num_threads)
1214 struct shared_queue res_to_compress_queue;
1215 struct shared_queue compressed_res_queue;
1216 pthread_t *compressor_threads = NULL;
1218 if (num_threads == 0) {
1219 long nthreads = get_default_num_threads();
1220 if (nthreads < 1 || nthreads > UINT_MAX) {
1221 WARNING("Could not determine number of processors! Assuming 1");
1224 num_threads = nthreads;
1228 progress->write_streams.num_threads = num_threads;
1230 static const size_t MESSAGES_PER_THREAD = 2;
1231 size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1233 DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1235 ret = shared_queue_init(&res_to_compress_queue, queue_size);
1239 ret = shared_queue_init(&compressed_res_queue, queue_size);
1241 goto out_destroy_res_to_compress_queue;
1243 struct compressor_thread_params params;
1244 params.res_to_compress_queue = &res_to_compress_queue;
1245 params.compressed_res_queue = &compressed_res_queue;
1246 params.compress = get_compress_func(out_ctype);
1248 compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1249 if (!compressor_threads) {
1250 ret = WIMLIB_ERR_NOMEM;
1251 goto out_destroy_compressed_res_queue;
1254 for (unsigned i = 0; i < num_threads; i++) {
1255 DEBUG("pthread_create thread %u of %u", i + 1, num_threads);
1256 ret = pthread_create(&compressor_threads[i], NULL,
1257 compressor_thread_proc, ¶ms);
1260 ERROR_WITH_ERRNO("Failed to create compressor "
1262 i + 1, num_threads);
1269 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
1271 struct main_writer_thread_ctx ctx;
1272 ctx.stream_list = stream_list;
1273 ctx.lookup_table = lookup_table;
1274 ctx.out_fp = out_fp;
1275 ctx.out_ctype = out_ctype;
1276 ctx.res_to_compress_queue = &res_to_compress_queue;
1277 ctx.compressed_res_queue = &compressed_res_queue;
1278 ctx.num_messages = queue_size;
1279 ctx.write_resource_flags = write_resource_flags | WIMLIB_RESOURCE_FLAG_THREADSAFE_READ;
1280 ctx.progress_func = progress_func;
1281 ctx.progress = progress;
1282 ret = main_writer_thread_init_ctx(&ctx);
1285 ret = do_write_stream_list(stream_list, lookup_table,
1286 main_thread_process_next_stream,
1289 goto out_destroy_ctx;
1291 /* The main thread has finished reading all streams that are going to be
1292 * compressed in parallel, and it now needs to wait for all remaining
1293 * chunks to be compressed so that the remaining streams can actually be
1294 * written to the output file. Furthermore, any remaining streams that
1295 * had processing deferred to the main thread need to be handled. These
1296 * tasks are done by the main_writer_thread_finish() function. */
1297 ret = main_writer_thread_finish(&ctx);
1299 main_writer_thread_destroy_ctx(&ctx);
1301 for (unsigned i = 0; i < num_threads; i++)
1302 shared_queue_put(&res_to_compress_queue, NULL);
1304 for (unsigned i = 0; i < num_threads; i++) {
1305 if (pthread_join(compressor_threads[i], NULL)) {
1306 WARNING_WITH_ERRNO("Failed to join compressor "
1308 i + 1, num_threads);
1311 FREE(compressor_threads);
1312 out_destroy_compressed_res_queue:
1313 shared_queue_destroy(&compressed_res_queue);
1314 out_destroy_res_to_compress_queue:
1315 shared_queue_destroy(&res_to_compress_queue);
1316 if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
1319 WARNING("Falling back to single-threaded compression");
1320 return write_stream_list_serial(stream_list,
1324 write_resource_flags,
1332 * Write a list of streams to a WIM (@out_fp) using the compression type
1333 * @out_ctype and up to @num_threads compressor threads.
1336 write_stream_list(struct list_head *stream_list,
1337 struct wim_lookup_table *lookup_table,
1338 FILE *out_fp, int out_ctype, int write_flags,
1339 unsigned num_threads, wimlib_progress_func_t progress_func)
1341 struct wim_lookup_table_entry *lte;
1342 size_t num_streams = 0;
1343 u64 total_bytes = 0;
1344 u64 total_compression_bytes = 0;
1345 union wimlib_progress_info progress;
1347 int write_resource_flags;
1349 if (list_empty(stream_list))
1352 write_resource_flags = write_flags_to_resource_flags(write_flags);
1354 /* Calculate the total size of the streams to be written. Note: this
1355 * will be the uncompressed size, as we may not know the compressed size
1356 * yet, and also this will assume that every unhashed stream will be
1357 * written (which will not necessarily be the case). */
1358 list_for_each_entry(lte, stream_list, write_streams_list) {
1360 total_bytes += wim_resource_size(lte);
1361 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
1362 && (wim_resource_compression_type(lte) != out_ctype ||
1363 (write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS)))
1365 total_compression_bytes += wim_resource_size(lte);
1368 progress.write_streams.total_bytes = total_bytes;
1369 progress.write_streams.total_streams = num_streams;
1370 progress.write_streams.completed_bytes = 0;
1371 progress.write_streams.completed_streams = 0;
1372 progress.write_streams.num_threads = num_threads;
1373 progress.write_streams.compression_type = out_ctype;
1374 progress.write_streams._private = 0;
1376 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1377 if (total_compression_bytes >= 1000000 && num_threads != 1)
1378 ret = write_stream_list_parallel(stream_list,
1382 write_resource_flags,
1388 ret = write_stream_list_serial(stream_list,
1392 write_resource_flags,
1398 struct stream_size_table {
1399 struct hlist_head *array;
1405 init_stream_size_table(struct stream_size_table *tab, size_t capacity)
1407 tab->array = CALLOC(capacity, sizeof(tab->array[0]));
1409 return WIMLIB_ERR_NOMEM;
1410 tab->num_entries = 0;
1411 tab->capacity = capacity;
1416 destroy_stream_size_table(struct stream_size_table *tab)
1422 stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
1424 struct stream_size_table *tab = _tab;
1426 struct wim_lookup_table_entry *same_size_lte;
1427 struct hlist_node *tmp;
1429 pos = hash_u64(wim_resource_size(lte)) % tab->capacity;
1430 lte->unique_size = 1;
1431 hlist_for_each_entry(same_size_lte, tmp, &tab->array[pos], hash_list_2) {
1432 if (wim_resource_size(same_size_lte) == wim_resource_size(lte)) {
1433 lte->unique_size = 0;
1434 same_size_lte->unique_size = 0;
1439 hlist_add_head(<e->hash_list_2, &tab->array[pos]);
1445 struct lte_overwrite_prepare_args {
1448 struct list_head stream_list;
1449 struct stream_size_table stream_size_tab;
1452 /* First phase of preparing streams for an in-place overwrite. This is called
1453 * on all streams, both hashed and unhashed, except the metadata resources. */
1455 lte_overwrite_prepare(struct wim_lookup_table_entry *lte, void *_args)
1457 struct lte_overwrite_prepare_args *args = _args;
1459 wimlib_assert(!(lte->resource_entry.flags & WIM_RESHDR_FLAG_METADATA));
1460 if (lte->resource_location != RESOURCE_IN_WIM || lte->wim != args->wim)
1461 list_add_tail(<e->write_streams_list, &args->stream_list);
1462 lte->out_refcnt = lte->refcnt;
1463 stream_size_table_insert(lte, &args->stream_size_tab);
1467 /* Second phase of preparing streams for an in-place overwrite. This is called
1468 * on existing metadata resources and hashed streams, but not unhashed streams.
1470 * NOTE: lte->output_resource_entry is in union with lte->hash_list_2, so
1471 * lte_overwrite_prepare_2() must be called after lte_overwrite_prepare(), as
1472 * the latter uses lte->hash_list_2, while the former expects to set
1473 * lte->output_resource_entry. */
1475 lte_overwrite_prepare_2(struct wim_lookup_table_entry *lte, void *_args)
1477 struct lte_overwrite_prepare_args *args = _args;
1479 if (lte->resource_location == RESOURCE_IN_WIM && lte->wim == args->wim) {
1480 /* We can't do an in place overwrite on the WIM if there are
1481 * streams after the XML data. */
1482 if (lte->resource_entry.offset +
1483 lte->resource_entry.size > args->end_offset)
1485 #ifdef ENABLE_ERROR_MESSAGES
1486 ERROR("The following resource is after the XML data:");
1487 print_lookup_table_entry(lte, stderr);
1489 return WIMLIB_ERR_RESOURCE_ORDER;
1491 copy_resource_entry(<e->output_resource_entry,
1492 <e->resource_entry);
1497 /* Given a WIM that we are going to overwrite in place with zero or more
1498 * additional streams added, construct a list the list of new unique streams
1499 * ('struct wim_lookup_table_entry's) that must be written, plus any unhashed
1500 * streams that need to be added but may be identical to other hashed or
1501 * unhashed streams. These unhashed streams are checksummed while the streams
1502 * are being written. To aid this process, the member @unique_size is set to 1
1503 * on streams that have a unique size and therefore must be written.
1505 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1506 * indicate the number of times the stream is referenced in only the streams
1507 * that are being written; this may still be adjusted later when unhashed
1508 * streams are being resolved.
1511 prepare_streams_for_overwrite(WIMStruct *wim, off_t end_offset,
1512 struct list_head *stream_list)
1515 struct lte_overwrite_prepare_args args;
1519 args.end_offset = end_offset;
1520 ret = init_stream_size_table(&args.stream_size_tab,
1521 wim->lookup_table->capacity);
1525 INIT_LIST_HEAD(&args.stream_list);
1526 for (i = 0; i < wim->hdr.image_count; i++) {
1527 struct wim_image_metadata *imd;
1528 struct wim_lookup_table_entry *lte;
1530 imd = wim->image_metadata[i];
1531 image_for_each_unhashed_stream(lte, imd)
1532 lte_overwrite_prepare(lte, &args);
1534 for_lookup_table_entry(wim->lookup_table, lte_overwrite_prepare, &args);
1535 list_transfer(&args.stream_list, stream_list);
1537 for (i = 0; i < wim->hdr.image_count; i++) {
1538 ret = lte_overwrite_prepare_2(wim->image_metadata[i]->metadata_lte,
1541 goto out_destroy_stream_size_table;
1543 ret = for_lookup_table_entry(wim->lookup_table,
1544 lte_overwrite_prepare_2, &args);
1545 out_destroy_stream_size_table:
1546 destroy_stream_size_table(&args.stream_size_tab);
1551 struct find_streams_ctx {
1552 struct list_head stream_list;
1553 struct stream_size_table stream_size_tab;
1557 inode_find_streams_to_write(struct wim_inode *inode,
1558 struct wim_lookup_table *table,
1559 struct list_head *stream_list,
1560 struct stream_size_table *tab)
1562 struct wim_lookup_table_entry *lte;
1563 for (unsigned i = 0; i <= inode->i_num_ads; i++) {
1564 lte = inode_stream_lte(inode, i, table);
1566 if (lte->out_refcnt == 0) {
1568 stream_size_table_insert(lte, tab);
1569 list_add_tail(<e->write_streams_list, stream_list);
1571 lte->out_refcnt += inode->i_nlink;
1577 image_find_streams_to_write(WIMStruct *w)
1579 struct find_streams_ctx *ctx;
1580 struct wim_image_metadata *imd;
1581 struct wim_inode *inode;
1582 struct wim_lookup_table_entry *lte;
1585 imd = wim_get_current_image_metadata(w);
1587 image_for_each_unhashed_stream(lte, imd)
1588 lte->out_refcnt = 0;
1590 /* Go through this image's inodes to find any streams that have not been
1592 image_for_each_inode(inode, imd) {
1593 inode_find_streams_to_write(inode, w->lookup_table,
1595 &ctx->stream_size_tab);
1600 /* Given a WIM that from which one or all of the images is being written, build
1601 * the list of unique streams ('struct wim_lookup_table_entry's) that must be
1602 * written, plus any unhashed streams that need to be written but may be
1603 * identical to other hashed or unhashed streams being written. These unhashed
1604 * streams are checksummed while the streams are being written. To aid this
1605 * process, the member @unique_size is set to 1 on streams that have a unique
1606 * size and therefore must be written.
1608 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1609 * indicate the number of times the stream is referenced in only the streams
1610 * that are being written; this may still be adjusted later when unhashed
1611 * streams are being resolved.
1614 prepare_stream_list(WIMStruct *wim, int image, struct list_head *stream_list)
1617 struct find_streams_ctx ctx;
1619 for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
1620 ret = init_stream_size_table(&ctx.stream_size_tab,
1621 wim->lookup_table->capacity);
1624 for_lookup_table_entry(wim->lookup_table, stream_size_table_insert,
1625 &ctx.stream_size_tab);
1626 INIT_LIST_HEAD(&ctx.stream_list);
1627 wim->private = &ctx;
1628 ret = for_image(wim, image, image_find_streams_to_write);
1629 destroy_stream_size_table(&ctx.stream_size_tab);
1631 list_transfer(&ctx.stream_list, stream_list);
1635 /* Writes the streams for the specified @image in @wim to @wim->out_fp.
1638 write_wim_streams(WIMStruct *wim, int image, int write_flags,
1639 unsigned num_threads,
1640 wimlib_progress_func_t progress_func)
1643 struct list_head stream_list;
1645 ret = prepare_stream_list(wim, image, &stream_list);
1648 return write_stream_list(&stream_list,
1651 wimlib_get_compression_type(wim),
1658 * Finish writing a WIM file: write the lookup table, xml data, and integrity
1659 * table (optional), then overwrite the WIM header.
1661 * write_flags is a bitwise OR of the following:
1663 * (public) WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
1664 * Include an integrity table.
1666 * (public) WIMLIB_WRITE_FLAG_SHOW_PROGRESS:
1667 * Show progress information when (if) writing the integrity table.
1669 * (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
1670 * Don't write the lookup table.
1672 * (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
1673 * When (if) writing the integrity table, re-use entries from the
1674 * existing integrity table, if possible.
1676 * (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
1677 * After writing the XML data but before writing the integrity
1678 * table, write a temporary WIM header and flush the stream so that
1679 * the WIM is less likely to become corrupted upon abrupt program
1682 * (private) WIMLIB_WRITE_FLAG_FSYNC:
1683 * fsync() the output file before closing it.
1687 finish_write(WIMStruct *w, int image, int write_flags,
1688 wimlib_progress_func_t progress_func)
1691 struct wim_header hdr;
1692 FILE *out = w->out_fp;
1694 /* @hdr will be the header for the new WIM. First copy all the data
1695 * from the header in the WIMStruct; then set all the fields that may
1696 * have changed, including the resource entries, boot index, and image
1698 memcpy(&hdr, &w->hdr, sizeof(struct wim_header));
1700 /* Set image count and boot index correctly for single image writes */
1701 if (image != WIMLIB_ALL_IMAGES) {
1702 hdr.image_count = 1;
1703 if (hdr.boot_idx == image)
1709 /* In the WIM header, there is room for the resource entry for a
1710 * metadata resource labeled as the "boot metadata". This entry should
1711 * be zeroed out if there is no bootable image (boot_idx 0). Otherwise,
1712 * it should be a copy of the resource entry for the image that is
1713 * marked as bootable. This is not well documented... */
1714 if (hdr.boot_idx == 0) {
1715 zero_resource_entry(&hdr.boot_metadata_res_entry);
1717 copy_resource_entry(&hdr.boot_metadata_res_entry,
1718 &w->image_metadata[ hdr.boot_idx- 1
1719 ]->metadata_lte->output_resource_entry);
1722 if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
1723 ret = write_lookup_table(w, image, &hdr.lookup_table_res_entry);
1728 ret = write_xml_data(w->wim_info, image, out,
1729 (write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE) ?
1730 wim_info_get_total_bytes(w->wim_info) : 0,
1731 &hdr.xml_res_entry);
1735 if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
1736 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
1737 struct wim_header checkpoint_hdr;
1738 memcpy(&checkpoint_hdr, &hdr, sizeof(struct wim_header));
1739 zero_resource_entry(&checkpoint_hdr.integrity);
1740 if (fseeko(out, 0, SEEK_SET)) {
1741 ERROR_WITH_ERRNO("Failed to seek to beginning "
1742 "of WIM being written");
1743 ret = WIMLIB_ERR_WRITE;
1746 ret = write_header(&checkpoint_hdr, out);
1750 if (fflush(out) != 0) {
1751 ERROR_WITH_ERRNO("Can't write data to WIM");
1752 ret = WIMLIB_ERR_WRITE;
1756 if (fseeko(out, 0, SEEK_END) != 0) {
1757 ERROR_WITH_ERRNO("Failed to seek to end "
1758 "of WIM being written");
1759 ret = WIMLIB_ERR_WRITE;
1764 off_t old_lookup_table_end;
1765 off_t new_lookup_table_end;
1766 if (write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE) {
1767 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
1768 w->hdr.lookup_table_res_entry.size;
1770 old_lookup_table_end = 0;
1772 new_lookup_table_end = hdr.lookup_table_res_entry.offset +
1773 hdr.lookup_table_res_entry.size;
1775 ret = write_integrity_table(out,
1777 new_lookup_table_end,
1778 old_lookup_table_end,
1783 zero_resource_entry(&hdr.integrity);
1786 if (fseeko(out, 0, SEEK_SET) != 0) {
1787 ERROR_WITH_ERRNO("Failed to seek to beginning of WIM "
1789 ret = WIMLIB_ERR_WRITE;
1793 ret = write_header(&hdr, out);
1797 if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
1798 if (fflush(out) != 0
1799 || fsync(fileno(out)) != 0)
1801 ERROR_WITH_ERRNO("Error flushing data to WIM file");
1802 ret = WIMLIB_ERR_WRITE;
1806 if (fclose(out) != 0) {
1807 ERROR_WITH_ERRNO("Failed to close the output WIM file");
1809 ret = WIMLIB_ERR_WRITE;
1815 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
1817 lock_wim(WIMStruct *w, FILE *fp)
1820 if (fp && !w->wim_locked) {
1821 ret = flock(fileno(fp), LOCK_EX | LOCK_NB);
1823 if (errno == EWOULDBLOCK) {
1824 ERROR("`%"TS"' is already being modified or has been "
1825 "mounted read-write\n"
1826 " by another process!", w->filename);
1827 ret = WIMLIB_ERR_ALREADY_LOCKED;
1829 WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
1842 open_wim_writable(WIMStruct *w, const tchar *path,
1843 bool trunc, bool also_readable)
1854 wimlib_assert(w->out_fp == NULL);
1855 w->out_fp = tfopen(path, mode);
1859 ERROR_WITH_ERRNO("Failed to open `%"TS"' for writing", path);
1860 return WIMLIB_ERR_OPEN;
1866 close_wim_writable(WIMStruct *w)
1869 if (fclose(w->out_fp) != 0) {
1870 WARNING_WITH_ERRNO("Failed to close output WIM");
1876 /* Open file stream and write dummy header for WIM. */
1878 begin_write(WIMStruct *w, const tchar *path, int write_flags)
1881 ret = open_wim_writable(w, path, true,
1882 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
1885 /* Write dummy header. It will be overwritten later. */
1886 return write_header(&w->hdr, w->out_fp);
1889 /* Writes a stand-alone WIM to a file. */
1891 wimlib_write(WIMStruct *w, const tchar *path,
1892 int image, int write_flags, unsigned num_threads,
1893 wimlib_progress_func_t progress_func)
1898 return WIMLIB_ERR_INVALID_PARAM;
1900 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1902 if (image != WIMLIB_ALL_IMAGES &&
1903 (image < 1 || image > w->hdr.image_count))
1904 return WIMLIB_ERR_INVALID_IMAGE;
1906 if (w->hdr.total_parts != 1) {
1907 ERROR("Cannot call wimlib_write() on part of a split WIM");
1908 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1911 ret = begin_write(w, path, write_flags);
1915 ret = write_wim_streams(w, image, write_flags, num_threads,
1921 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
1923 ret = for_image(w, image, write_metadata_resource);
1928 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
1930 ret = finish_write(w, image, write_flags, progress_func);
1931 /* finish_write() closed the WIM for us */
1934 close_wim_writable(w);
1936 DEBUG("wimlib_write(path=%"TS") = %d", path, ret);
1941 any_images_modified(WIMStruct *w)
1943 for (int i = 0; i < w->hdr.image_count; i++)
1944 if (w->image_metadata[i]->modified)
1950 * Overwrite a WIM, possibly appending streams to it.
1952 * A WIM looks like (or is supposed to look like) the following:
1954 * Header (212 bytes)
1955 * Streams and metadata resources (variable size)
1956 * Lookup table (variable size)
1957 * XML data (variable size)
1958 * Integrity table (optional) (variable size)
1960 * If we are not adding any streams or metadata resources, the lookup table is
1961 * unchanged--- so we only need to overwrite the XML data, integrity table, and
1962 * header. This operation is potentially unsafe if the program is abruptly
1963 * terminated while the XML data or integrity table are being overwritten, but
1964 * before the new header has been written. To partially alleviate this problem,
1965 * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
1966 * finish_write() to cause a temporary WIM header to be written after the XML
1967 * data has been written. This may prevent the WIM from becoming corrupted if
1968 * the program is terminated while the integrity table is being calculated (but
1969 * no guarantees, due to write re-ordering...).
1971 * If we are adding new streams or images (metadata resources), the lookup table
1972 * needs to be changed, and those streams need to be written. In this case, we
1973 * try to perform a safe update of the WIM file by writing the streams *after*
1974 * the end of the previous WIM, then writing the new lookup table, XML data, and
1975 * (optionally) integrity table following the new streams. This will produce a
1976 * layout like the following:
1978 * Header (212 bytes)
1979 * (OLD) Streams and metadata resources (variable size)
1980 * (OLD) Lookup table (variable size)
1981 * (OLD) XML data (variable size)
1982 * (OLD) Integrity table (optional) (variable size)
1983 * (NEW) Streams and metadata resources (variable size)
1984 * (NEW) Lookup table (variable size)
1985 * (NEW) XML data (variable size)
1986 * (NEW) Integrity table (optional) (variable size)
1988 * At all points, the WIM is valid as nothing points to the new data yet. Then,
1989 * the header is overwritten to point to the new lookup table, XML data, and
1990 * integrity table, to produce the following layout:
1992 * Header (212 bytes)
1993 * Streams and metadata resources (variable size)
1994 * Nothing (variable size)
1995 * More Streams and metadata resources (variable size)
1996 * Lookup table (variable size)
1997 * XML data (variable size)
1998 * Integrity table (optional) (variable size)
2000 * This method allows an image to be appended to a large WIM very quickly, and
2001 * is is crash-safe except in the case of write re-ordering, but the
2002 * disadvantage is that a small hole is left in the WIM where the old lookup
2003 * table, xml data, and integrity table were. (These usually only take up a
2004 * small amount of space compared to the streams, however.)
2007 overwrite_wim_inplace(WIMStruct *w, int write_flags,
2008 unsigned num_threads,
2009 wimlib_progress_func_t progress_func)
2012 struct list_head stream_list;
2014 u64 old_lookup_table_end, old_xml_begin, old_xml_end;
2016 DEBUG("Overwriting `%"TS"' in-place", w->filename);
2018 /* Make sure that the integrity table (if present) is after the XML
2019 * data, and that there are no stream resources, metadata resources, or
2020 * lookup tables after the XML data. Otherwise, these data would be
2022 old_xml_begin = w->hdr.xml_res_entry.offset;
2023 old_xml_end = old_xml_begin + w->hdr.xml_res_entry.size;
2024 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
2025 w->hdr.lookup_table_res_entry.size;
2026 if (w->hdr.integrity.offset != 0 && w->hdr.integrity.offset < old_xml_end) {
2027 ERROR("Didn't expect the integrity table to be before the XML data");
2028 return WIMLIB_ERR_RESOURCE_ORDER;
2031 if (old_lookup_table_end > old_xml_begin) {
2032 ERROR("Didn't expect the lookup table to be after the XML data");
2033 return WIMLIB_ERR_RESOURCE_ORDER;
2036 /* Set @old_wim_end, which indicates the point beyond which we don't
2037 * allow any file and metadata resources to appear without returning
2038 * WIMLIB_ERR_RESOURCE_ORDER (due to the fact that we would otherwise
2039 * overwrite these resources). */
2040 if (!w->deletion_occurred && !any_images_modified(w)) {
2041 /* If no images have been modified and no images have been
2042 * deleted, a new lookup table does not need to be written. We
2043 * shall write the new XML data and optional integrity table
2044 * immediately after the lookup table. Note that this may
2045 * overwrite an existing integrity table. */
2046 DEBUG("Skipping writing lookup table "
2047 "(no images modified or deleted)");
2048 old_wim_end = old_lookup_table_end;
2049 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
2050 WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
2051 } else if (w->hdr.integrity.offset) {
2052 /* Old WIM has an integrity table; begin writing new streams
2054 old_wim_end = w->hdr.integrity.offset + w->hdr.integrity.size;
2056 /* No existing integrity table; begin writing new streams after
2057 * the old XML data. */
2058 old_wim_end = old_xml_end;
2061 ret = prepare_streams_for_overwrite(w, old_wim_end, &stream_list);
2065 ret = open_wim_writable(w, w->filename, false,
2066 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
2070 ret = lock_wim(w, w->out_fp);
2072 close_wim_writable(w);
2076 if (fseeko(w->out_fp, old_wim_end, SEEK_SET) != 0) {
2077 ERROR_WITH_ERRNO("Can't seek to end of WIM");
2078 close_wim_writable(w);
2080 return WIMLIB_ERR_WRITE;
2083 DEBUG("Writing newly added streams (offset = %"PRIu64")",
2085 ret = write_stream_list(&stream_list,
2088 wimlib_get_compression_type(w),
2095 for (int i = 0; i < w->hdr.image_count; i++) {
2096 if (w->image_metadata[i]->modified) {
2097 select_wim_image(w, i + 1);
2098 ret = write_metadata_resource(w);
2103 write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
2104 ret = finish_write(w, WIMLIB_ALL_IMAGES, write_flags,
2107 close_wim_writable(w);
2108 if (ret != 0 && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
2109 WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
2110 w->filename, old_wim_end);
2111 /* Return value of truncate() is ignored because this is already
2113 (void)ttruncate(w->filename, old_wim_end);
2120 overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
2121 unsigned num_threads,
2122 wimlib_progress_func_t progress_func)
2124 size_t wim_name_len;
2127 DEBUG("Overwriting `%"TS"' via a temporary file", w->filename);
2129 /* Write the WIM to a temporary file in the same directory as the
2131 wim_name_len = tstrlen(w->filename);
2132 tchar tmpfile[wim_name_len + 10];
2133 tmemcpy(tmpfile, w->filename, wim_name_len);
2134 randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
2135 tmpfile[wim_name_len + 9] = T('\0');
2137 ret = wimlib_write(w, tmpfile, WIMLIB_ALL_IMAGES,
2138 write_flags | WIMLIB_WRITE_FLAG_FSYNC,
2139 num_threads, progress_func);
2141 ERROR("Failed to write the WIM file `%"TS"'", tmpfile);
2145 DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, w->filename);
2148 /* Windows won't let you delete open files unless FILE_SHARE_DELETE was
2149 * specified to CreateFile(). The WIM was opened with fopen(), which
2150 * didn't provided this flag to CreateFile, so the handle must be closed
2151 * before executing the rename(). */
2152 if (w->fp != NULL) {
2158 /* Rename the new file to the old file .*/
2159 if (trename(tmpfile, w->filename) != 0) {
2160 ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
2161 tmpfile, w->filename);
2162 ret = WIMLIB_ERR_RENAME;
2166 if (progress_func) {
2167 union wimlib_progress_info progress;
2168 progress.rename.from = tmpfile;
2169 progress.rename.to = w->filename;
2170 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
2173 /* Close the original WIM file that was opened for reading. */
2174 if (w->fp != NULL) {
2179 /* Re-open the WIM read-only. */
2180 w->fp = tfopen(w->filename, T("rb"));
2181 if (w->fp == NULL) {
2182 ret = WIMLIB_ERR_REOPEN;
2183 WARNING_WITH_ERRNO("Failed to re-open `%"TS"' read-only",
2190 /* Remove temporary file. */
2191 if (tunlink(tmpfile) != 0)
2192 WARNING_WITH_ERRNO("Failed to remove `%"TS"'", tmpfile);
2198 * Writes a WIM file to the original file that it was read from, overwriting it.
2201 wimlib_overwrite(WIMStruct *w, int write_flags,
2202 unsigned num_threads,
2203 wimlib_progress_func_t progress_func)
2205 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2208 return WIMLIB_ERR_NO_FILENAME;
2210 if (w->hdr.total_parts != 1) {
2211 ERROR("Cannot modify a split WIM");
2212 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
2215 if ((!w->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
2216 && !(write_flags & WIMLIB_WRITE_FLAG_REBUILD))
2219 ret = overwrite_wim_inplace(w, write_flags, num_threads,
2221 if (ret == WIMLIB_ERR_RESOURCE_ORDER)
2222 WARNING("Falling back to re-building entire WIM");
2226 return overwrite_wim_via_tmpfile(w, write_flags, num_threads,