4 * Support for writing WIM files; write a WIM file, overwrite a WIM file, write
5 * compressed file resources, etc.
9 * Copyright (C) 2012, 2013 Eric Biggers
11 * This file is part of wimlib, a library for working with WIM files.
13 * wimlib is free software; you can redistribute it and/or modify it under the
14 * terms of the GNU General Public License as published by the Free
15 * Software Foundation; either version 3 of the License, or (at your option)
18 * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
19 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
20 * A PARTICULAR PURPOSE. See the GNU General Public License for more
23 * You should have received a copy of the GNU General Public License
24 * along with wimlib; if not, see http://www.gnu.org/licenses/.
31 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
32 /* On BSD, this should be included before "wimlib/list.h" so that "wimlib/list.h" can
33 * overwrite the LIST_HEAD macro. */
34 # include <sys/file.h>
37 #include "wimlib/endianness.h"
38 #include "wimlib/error.h"
39 #include "wimlib/file_io.h"
40 #include "wimlib/header.h"
41 #include "wimlib/integrity.h"
42 #include "wimlib/lookup_table.h"
43 #include "wimlib/metadata.h"
44 #include "wimlib/resource.h"
45 #include "wimlib/write.h"
46 #include "wimlib/xml.h"
49 # include "wimlib/win32.h" /* win32_get_number_of_processors() */
52 #ifdef ENABLE_MULTITHREADED_COMPRESSION
62 # include <ntfs-3g/attrib.h>
63 # include <ntfs-3g/inode.h>
64 # include <ntfs-3g/dir.h>
76 # include <sys/uio.h> /* for `struct iovec' */
79 /* Chunk table that's located at the beginning of each compressed resource in
80 * the WIM. (This is not the on-disk format; the on-disk format just has an
81 * array of offsets.) */
85 u64 original_resource_size;
86 u64 bytes_per_chunk_entry;
97 * Allocates and initializes a chunk table, and reserves space for it in the
101 begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
104 struct chunk_table **chunk_tab_ret)
106 u64 size = wim_resource_size(lte);
107 u64 num_chunks = (size + WIM_CHUNK_SIZE - 1) / WIM_CHUNK_SIZE;
108 size_t alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
109 struct chunk_table *chunk_tab = CALLOC(1, alloc_size);
111 DEBUG("Begin chunk table for stream with size %"PRIu64, size);
114 ERROR("Failed to allocate chunk table for %"PRIu64" byte "
116 return WIMLIB_ERR_NOMEM;
118 chunk_tab->file_offset = file_offset;
119 chunk_tab->num_chunks = num_chunks;
120 chunk_tab->original_resource_size = size;
121 chunk_tab->bytes_per_chunk_entry = (size >= (1ULL << 32)) ? 8 : 4;
122 chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
124 chunk_tab->cur_offset = 0;
125 chunk_tab->cur_offset_p = chunk_tab->offsets;
127 if (full_write(out_fd, chunk_tab,
128 chunk_tab->table_disk_size) != chunk_tab->table_disk_size)
130 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
133 return WIMLIB_ERR_WRITE;
135 *chunk_tab_ret = chunk_tab;
140 * compress_func_t- Pointer to a function to compresses a chunk
141 * of a WIM resource. This may be either
142 * wimlib_xpress_compress() (xpress-compress.c) or
143 * wimlib_lzx_compress() (lzx-compress.c).
145 * @chunk: Uncompressed data of the chunk.
146 * @chunk_size: Size of the uncompressed chunk, in bytes.
147 * @out: Pointer to output buffer of size at least (@chunk_size - 1) bytes.
149 * Returns the size of the compressed data written to @out in bytes, or 0 if the
150 * data could not be compressed to (@chunk_size - 1) bytes or fewer.
152 * As a special requirement, the compression code is optimized for the WIM
153 * format and therefore requires (@chunk_size <= 32768).
155 * As another special requirement, the compression code will read up to 8 bytes
156 * off the end of the @chunk array for performance reasons. The values of these
157 * bytes will not affect the output of the compression, but the calling code
158 * must make sure that the buffer holding the uncompressed chunk is actually at
159 * least (@chunk_size + 8) bytes, or at least that these extra bytes are in
160 * mapped memory that will not cause a memory access violation if accessed.
162 typedef unsigned (*compress_func_t)(const void *chunk, unsigned chunk_size,
165 static compress_func_t
166 get_compress_func(int out_ctype)
168 if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX)
169 return wimlib_lzx_compress;
171 return wimlib_xpress_compress;
175 * Writes a chunk of a WIM resource to an output file.
177 * @chunk: Uncompressed data of the chunk.
178 * @chunk_size: Size of the chunk (<= WIM_CHUNK_SIZE)
179 * @out_fd: File descriptor to write the chunk to.
180 * @compress: Compression function to use (NULL if writing uncompressed
182 * @chunk_tab: Pointer to chunk table being created. It is updated with the
183 * offset of the chunk we write.
185 * Returns 0 on success; nonzero on failure.
188 write_wim_resource_chunk(const void * restrict chunk,
191 compress_func_t compress,
192 struct chunk_table * restrict chunk_tab)
194 const void *out_chunk;
195 unsigned out_chunk_size;
197 void *compressed_chunk = alloca(chunk_size);
199 out_chunk_size = (*compress)(chunk, chunk_size, compressed_chunk);
200 if (out_chunk_size) {
201 /* Write compressed */
202 out_chunk = compressed_chunk;
204 /* Write uncompressed */
206 out_chunk_size = chunk_size;
208 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
209 chunk_tab->cur_offset += out_chunk_size;
211 /* Write uncompressed */
213 out_chunk_size = chunk_size;
215 if (full_write(out_fd, out_chunk, out_chunk_size) != out_chunk_size) {
216 ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
217 return WIMLIB_ERR_WRITE;
223 * Finishes a WIM chunk table and writes it to the output file at the correct
226 * The final size of the full compressed resource is returned in the
227 * @compressed_size_p.
230 finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
231 int out_fd, u64 *compressed_size_p)
233 size_t bytes_written;
235 if (chunk_tab->bytes_per_chunk_entry == 8) {
236 array_cpu_to_le64(chunk_tab->offsets, chunk_tab->num_chunks);
238 for (u64 i = 0; i < chunk_tab->num_chunks; i++)
239 chunk_tab->u32_offsets[i] = cpu_to_le32(chunk_tab->offsets[i]);
241 bytes_written = full_pwrite(out_fd,
242 (u8*)chunk_tab->offsets + chunk_tab->bytes_per_chunk_entry,
243 chunk_tab->table_disk_size,
244 chunk_tab->file_offset);
245 if (bytes_written != chunk_tab->table_disk_size) {
246 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
248 return WIMLIB_ERR_WRITE;
250 *compressed_size_p = chunk_tab->cur_offset + chunk_tab->table_disk_size;
255 seek_and_truncate(int out_fd, off_t offset)
257 if (lseek(out_fd, offset, SEEK_SET) == -1 ||
258 ftruncate(out_fd, offset))
260 ERROR_WITH_ERRNO("Failed to truncate output WIM file");
261 return WIMLIB_ERR_WRITE;
268 finalize_and_check_sha1(SHA_CTX * restrict sha_ctx,
269 struct wim_lookup_table_entry * restrict lte)
271 u8 md[SHA1_HASH_SIZE];
272 sha1_final(md, sha_ctx);
274 copy_hash(lte->hash, md);
275 } else if (!hashes_equal(md, lte->hash)) {
276 ERROR("WIM resource has incorrect hash!");
277 if (lte_filename_valid(lte)) {
278 ERROR("We were reading it from \"%"TS"\"; maybe "
279 "it changed while we were reading it.",
282 return WIMLIB_ERR_INVALID_RESOURCE_HASH;
288 struct write_resource_ctx {
289 compress_func_t compress;
290 struct chunk_table *chunk_tab;
297 write_resource_cb(const void *restrict chunk, size_t chunk_size,
300 struct write_resource_ctx *ctx = _ctx;
303 sha1_update(&ctx->sha_ctx, chunk, chunk_size);
304 return write_wim_resource_chunk(chunk, chunk_size,
305 ctx->out_fd, ctx->compress,
310 * Write a resource to an output WIM.
312 * @lte: Lookup table entry for the resource, which could be in another WIM,
313 * in an external file, or in another location.
315 * @out_fd: File descriptor opened to the output WIM.
317 * @out_ctype: One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate
318 * which compression algorithm to use.
320 * @out_res_entry: On success, this is filled in with the offset, flags,
321 * compressed size, and uncompressed size of the resource
324 * @flags: WIMLIB_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed
325 * even if it could otherwise be copied directly from the input.
327 * Additional notes: The SHA1 message digest of the uncompressed data is
328 * calculated (except when doing a raw copy --- see below). If the @unhashed
329 * flag is set on the lookup table entry, this message digest is simply copied
330 * to it; otherwise, the message digest is compared with the existing one, and
331 * the function will fail if they do not match.
334 write_wim_resource(struct wim_lookup_table_entry *lte,
335 int out_fd, int out_ctype,
336 struct resource_entry *out_res_entry,
339 struct write_resource_ctx write_ctx;
345 flags &= ~WIMLIB_RESOURCE_FLAG_RECOMPRESS;
347 /* Get current position in output WIM */
348 offset = filedes_offset(out_fd);
350 ERROR_WITH_ERRNO("Can't get position in output WIM");
351 return WIMLIB_ERR_WRITE;
354 /* If we are not forcing the data to be recompressed, and the input
355 * resource is located in a WIM with the same compression type as that
356 * desired other than no compression, we can simply copy the compressed
357 * data without recompressing it. This also means we must skip
358 * calculating the SHA1, as we never will see the uncompressed data. */
359 if (!(flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
360 lte->resource_location == RESOURCE_IN_WIM &&
361 out_ctype != WIMLIB_COMPRESSION_TYPE_NONE &&
362 wimlib_get_compression_type(lte->wim) == out_ctype)
364 flags |= WIMLIB_RESOURCE_FLAG_RAW;
365 write_ctx.doing_sha = false;
366 read_size = lte->resource_entry.size;
368 write_ctx.doing_sha = true;
369 sha1_init(&write_ctx.sha_ctx);
370 read_size = lte->resource_entry.original_size;
373 /* Initialize the chunk table and set the compression function if
374 * compressing the resource. */
375 if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
376 (flags & WIMLIB_RESOURCE_FLAG_RAW)) {
377 write_ctx.compress = NULL;
378 write_ctx.chunk_tab = NULL;
380 write_ctx.compress = get_compress_func(out_ctype);
381 ret = begin_wim_resource_chunk_tab(lte, out_fd,
383 &write_ctx.chunk_tab);
388 /* Write the entire resource by reading the entire resource and feeding
389 * the data through the write_resource_cb function. */
390 write_ctx.out_fd = out_fd;
392 ret = read_resource_prefix(lte, read_size,
393 write_resource_cb, &write_ctx, flags);
395 goto out_free_chunk_tab;
397 /* Verify SHA1 message digest of the resource, or set the hash for the
399 if (write_ctx.doing_sha) {
400 ret = finalize_and_check_sha1(&write_ctx.sha_ctx, lte);
402 goto out_free_chunk_tab;
405 out_res_entry->flags = lte->resource_entry.flags;
406 out_res_entry->original_size = wim_resource_size(lte);
407 out_res_entry->offset = offset;
408 if (flags & WIMLIB_RESOURCE_FLAG_RAW) {
409 /* Doing a raw write: The new compressed size is the same as
410 * the compressed size in the other WIM. */
411 new_size = lte->resource_entry.size;
412 } else if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE) {
413 /* Using WIMLIB_COMPRESSION_TYPE_NONE: The new compressed size
414 * is the original size. */
415 new_size = lte->resource_entry.original_size;
416 out_res_entry->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
418 /* Using a different compression type: Call
419 * finish_wim_resource_chunk_tab() and it will provide the new
420 * compressed size. */
421 ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab, out_fd,
424 goto out_free_chunk_tab;
425 if (new_size >= wim_resource_size(lte)) {
426 /* Oops! We compressed the resource to larger than the original
427 * size. Write the resource uncompressed instead. */
428 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
429 "writing uncompressed instead",
430 wim_resource_size(lte), new_size);
431 ret = seek_and_truncate(out_fd, offset);
433 goto out_free_chunk_tab;
434 write_ctx.compress = NULL;
435 write_ctx.doing_sha = false;
436 out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
437 goto try_write_again;
439 out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
441 out_res_entry->size = new_size;
444 FREE(write_ctx.chunk_tab);
448 #ifdef ENABLE_MULTITHREADED_COMPRESSION
450 /* Blocking shared queue (solves the producer-consumer problem) */
451 struct shared_queue {
455 unsigned filled_slots;
457 pthread_mutex_t lock;
458 pthread_cond_t msg_avail_cond;
459 pthread_cond_t space_avail_cond;
463 shared_queue_init(struct shared_queue *q, unsigned size)
465 wimlib_assert(size != 0);
466 q->array = CALLOC(sizeof(q->array[0]), size);
473 if (pthread_mutex_init(&q->lock, NULL)) {
474 ERROR_WITH_ERRNO("Failed to initialize mutex");
477 if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
478 ERROR_WITH_ERRNO("Failed to initialize condition variable");
479 goto err_destroy_lock;
481 if (pthread_cond_init(&q->space_avail_cond, NULL)) {
482 ERROR_WITH_ERRNO("Failed to initialize condition variable");
483 goto err_destroy_msg_avail_cond;
486 err_destroy_msg_avail_cond:
487 pthread_cond_destroy(&q->msg_avail_cond);
489 pthread_mutex_destroy(&q->lock);
491 return WIMLIB_ERR_NOMEM;
495 shared_queue_destroy(struct shared_queue *q)
498 pthread_mutex_destroy(&q->lock);
499 pthread_cond_destroy(&q->msg_avail_cond);
500 pthread_cond_destroy(&q->space_avail_cond);
504 shared_queue_put(struct shared_queue *q, void *obj)
506 pthread_mutex_lock(&q->lock);
507 while (q->filled_slots == q->size)
508 pthread_cond_wait(&q->space_avail_cond, &q->lock);
510 q->back = (q->back + 1) % q->size;
511 q->array[q->back] = obj;
514 pthread_cond_broadcast(&q->msg_avail_cond);
515 pthread_mutex_unlock(&q->lock);
519 shared_queue_get(struct shared_queue *q)
523 pthread_mutex_lock(&q->lock);
524 while (q->filled_slots == 0)
525 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
527 obj = q->array[q->front];
528 q->array[q->front] = NULL;
529 q->front = (q->front + 1) % q->size;
532 pthread_cond_broadcast(&q->space_avail_cond);
533 pthread_mutex_unlock(&q->lock);
537 struct compressor_thread_params {
538 struct shared_queue *res_to_compress_queue;
539 struct shared_queue *compressed_res_queue;
540 compress_func_t compress;
543 #define MAX_CHUNKS_PER_MSG 2
546 struct wim_lookup_table_entry *lte;
547 u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
548 u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
549 unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
550 struct iovec out_chunks[MAX_CHUNKS_PER_MSG];
551 size_t total_out_bytes;
553 struct list_head list;
559 compress_chunks(struct message *msg, compress_func_t compress)
561 msg->total_out_bytes = 0;
562 for (unsigned i = 0; i < msg->num_chunks; i++) {
563 unsigned len = compress(msg->uncompressed_chunks[i],
564 msg->uncompressed_chunk_sizes[i],
565 msg->compressed_chunks[i]);
569 /* To be written compressed */
570 out_chunk = msg->compressed_chunks[i];
573 /* To be written uncompressed */
574 out_chunk = msg->uncompressed_chunks[i];
575 out_len = msg->uncompressed_chunk_sizes[i];
577 msg->out_chunks[i].iov_base = out_chunk;
578 msg->out_chunks[i].iov_len = out_len;
579 msg->total_out_bytes += out_len;
583 /* Compressor thread routine. This is a lot simpler than the main thread
584 * routine: just repeatedly get a group of chunks from the
585 * res_to_compress_queue, compress them, and put them in the
586 * compressed_res_queue. A NULL pointer indicates that the thread should stop.
589 compressor_thread_proc(void *arg)
591 struct compressor_thread_params *params = arg;
592 struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
593 struct shared_queue *compressed_res_queue = params->compressed_res_queue;
594 compress_func_t compress = params->compress;
597 DEBUG("Compressor thread ready");
598 while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
599 compress_chunks(msg, compress);
600 shared_queue_put(compressed_res_queue, msg);
602 DEBUG("Compressor thread terminating");
605 #endif /* ENABLE_MULTITHREADED_COMPRESSION */
608 do_write_streams_progress(union wimlib_progress_info *progress,
609 wimlib_progress_func_t progress_func,
611 bool stream_discarded)
613 if (stream_discarded) {
614 progress->write_streams.total_bytes -= size_added;
615 if (progress->write_streams._private != ~(uint64_t)0 &&
616 progress->write_streams._private > progress->write_streams.total_bytes)
618 progress->write_streams._private = progress->write_streams.total_bytes;
621 progress->write_streams.completed_bytes += size_added;
623 progress->write_streams.completed_streams++;
625 progress->write_streams.completed_bytes >= progress->write_streams._private)
627 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
629 if (progress->write_streams._private == progress->write_streams.total_bytes) {
630 progress->write_streams._private = ~(uint64_t)0;
632 progress->write_streams._private =
633 min(progress->write_streams.total_bytes,
634 progress->write_streams.completed_bytes +
635 progress->write_streams.total_bytes / 100);
640 struct serial_write_stream_ctx {
643 int write_resource_flags;
647 serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
649 struct serial_write_stream_ctx *ctx = _ctx;
650 return write_wim_resource(lte, ctx->out_fd,
651 ctx->out_ctype, <e->output_resource_entry,
652 ctx->write_resource_flags);
655 /* Write a list of streams, taking into account that some streams may be
656 * duplicates that are checksummed and discarded on the fly, and also delegating
657 * the actual writing of a stream to a function @write_stream_cb, which is
658 * passed the context @write_stream_ctx. */
660 do_write_stream_list(struct list_head *stream_list,
661 struct wim_lookup_table *lookup_table,
662 int (*write_stream_cb)(struct wim_lookup_table_entry *, void *),
663 void *write_stream_ctx,
664 wimlib_progress_func_t progress_func,
665 union wimlib_progress_info *progress)
668 struct wim_lookup_table_entry *lte;
669 bool stream_discarded;
671 /* For each stream in @stream_list ... */
672 while (!list_empty(stream_list)) {
673 stream_discarded = false;
674 lte = container_of(stream_list->next,
675 struct wim_lookup_table_entry,
677 list_del(<e->write_streams_list);
678 if (lte->unhashed && !lte->unique_size) {
679 /* Unhashed stream that shares a size with some other
680 * stream in the WIM we are writing. The stream must be
681 * checksummed to know if we need to write it or not. */
682 struct wim_lookup_table_entry *tmp;
683 u32 orig_refcnt = lte->out_refcnt;
685 ret = hash_unhashed_stream(lte, lookup_table, &tmp);
690 /* We found a duplicate stream. */
691 if (orig_refcnt != tmp->out_refcnt) {
692 /* We have already written, or are going
693 * to write, the duplicate stream. So
694 * just skip to the next stream. */
695 DEBUG("Discarding duplicate stream of length %"PRIu64,
696 wim_resource_size(lte));
697 lte->no_progress = 0;
698 stream_discarded = true;
699 goto skip_to_progress;
704 /* Here, @lte is either a hashed stream or an unhashed stream
705 * with a unique size. In either case we know that the stream
706 * has to be written. In either case the SHA1 message digest
707 * will be calculated over the stream while writing it; however,
708 * in the former case this is done merely to check the data,
709 * while in the latter case this is done because we do not have
710 * the SHA1 message digest yet. */
711 wimlib_assert(lte->out_refcnt != 0);
713 lte->no_progress = 0;
714 ret = (*write_stream_cb)(lte, write_stream_ctx);
717 /* In parallel mode, some streams are deferred for later,
718 * serialized processing; ignore them here. */
722 list_del(<e->unhashed_list);
723 lookup_table_insert(lookup_table, lte);
727 if (!lte->no_progress) {
728 do_write_streams_progress(progress,
730 wim_resource_size(lte),
738 do_write_stream_list_serial(struct list_head *stream_list,
739 struct wim_lookup_table *lookup_table,
742 int write_resource_flags,
743 wimlib_progress_func_t progress_func,
744 union wimlib_progress_info *progress)
746 struct serial_write_stream_ctx ctx = {
748 .out_ctype = out_ctype,
749 .write_resource_flags = write_resource_flags,
751 return do_write_stream_list(stream_list,
760 write_flags_to_resource_flags(int write_flags)
762 int resource_flags = 0;
764 if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
765 resource_flags |= WIMLIB_RESOURCE_FLAG_RECOMPRESS;
766 return resource_flags;
770 write_stream_list_serial(struct list_head *stream_list,
771 struct wim_lookup_table *lookup_table,
774 int write_resource_flags,
775 wimlib_progress_func_t progress_func,
776 union wimlib_progress_info *progress)
778 DEBUG("Writing stream list (serial version)");
779 progress->write_streams.num_threads = 1;
781 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
782 return do_write_stream_list_serial(stream_list,
786 write_resource_flags,
791 #ifdef ENABLE_MULTITHREADED_COMPRESSION
793 write_wim_chunks(struct message *msg, int out_fd,
794 struct chunk_table *chunk_tab)
796 for (unsigned i = 0; i < msg->num_chunks; i++) {
797 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
798 chunk_tab->cur_offset += msg->out_chunks[i].iov_len;
800 if (full_writev(out_fd, msg->out_chunks,
801 msg->num_chunks) != msg->total_out_bytes)
803 ERROR_WITH_ERRNO("Failed to write WIM chunks");
804 return WIMLIB_ERR_WRITE;
809 struct main_writer_thread_ctx {
810 struct list_head *stream_list;
811 struct wim_lookup_table *lookup_table;
814 int write_resource_flags;
815 struct shared_queue *res_to_compress_queue;
816 struct shared_queue *compressed_res_queue;
818 wimlib_progress_func_t progress_func;
819 union wimlib_progress_info *progress;
821 struct list_head available_msgs;
822 struct list_head outstanding_streams;
823 struct list_head serial_streams;
824 size_t num_outstanding_messages;
826 SHA_CTX next_sha_ctx;
829 struct wim_lookup_table_entry *next_lte;
831 struct message *msgs;
832 struct message *next_msg;
833 struct chunk_table *cur_chunk_tab;
837 init_message(struct message *msg)
839 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
840 msg->compressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
841 msg->uncompressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
842 if (msg->compressed_chunks[i] == NULL ||
843 msg->uncompressed_chunks[i] == NULL)
844 return WIMLIB_ERR_NOMEM;
850 destroy_message(struct message *msg)
852 for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
853 FREE(msg->compressed_chunks[i]);
854 FREE(msg->uncompressed_chunks[i]);
859 free_messages(struct message *msgs, size_t num_messages)
862 for (size_t i = 0; i < num_messages; i++)
863 destroy_message(&msgs[i]);
868 static struct message *
869 allocate_messages(size_t num_messages)
871 struct message *msgs;
873 msgs = CALLOC(num_messages, sizeof(struct message));
876 for (size_t i = 0; i < num_messages; i++) {
877 if (init_message(&msgs[i])) {
878 free_messages(msgs, num_messages);
886 main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
888 while (ctx->num_outstanding_messages--)
889 shared_queue_get(ctx->compressed_res_queue);
890 free_messages(ctx->msgs, ctx->num_messages);
891 FREE(ctx->cur_chunk_tab);
895 main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
897 /* Pre-allocate all the buffers that will be needed to do the chunk
899 ctx->msgs = allocate_messages(ctx->num_messages);
901 return WIMLIB_ERR_NOMEM;
903 /* Initially, all the messages are available to use. */
904 INIT_LIST_HEAD(&ctx->available_msgs);
905 for (size_t i = 0; i < ctx->num_messages; i++)
906 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
908 /* outstanding_streams is the list of streams that currently have had
909 * chunks sent off for compression.
911 * The first stream in outstanding_streams is the stream that is
912 * currently being written.
914 * The last stream in outstanding_streams is the stream that is
915 * currently being read and having chunks fed to the compressor threads.
917 INIT_LIST_HEAD(&ctx->outstanding_streams);
918 ctx->num_outstanding_messages = 0;
920 ctx->next_msg = NULL;
922 /* Resources that don't need any chunks compressed are added to this
923 * list and written directly by the main thread. */
924 INIT_LIST_HEAD(&ctx->serial_streams);
926 ctx->cur_chunk_tab = NULL;
932 receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
935 struct wim_lookup_table_entry *cur_lte;
938 wimlib_assert(!list_empty(&ctx->outstanding_streams));
939 wimlib_assert(ctx->num_outstanding_messages != 0);
941 cur_lte = container_of(ctx->outstanding_streams.next,
942 struct wim_lookup_table_entry,
943 being_compressed_list);
945 /* Get the next message from the queue and process it.
946 * The message will contain 1 or more data chunks that have been
948 msg = shared_queue_get(ctx->compressed_res_queue);
949 msg->complete = true;
950 --ctx->num_outstanding_messages;
952 /* Is this the next chunk in the current resource? If it's not
953 * (i.e., an earlier chunk in a same or different resource
954 * hasn't been compressed yet), do nothing, and keep this
955 * message around until all earlier chunks are received.
957 * Otherwise, write all the chunks we can. */
958 while (cur_lte != NULL &&
959 !list_empty(&cur_lte->msg_list)
960 && (msg = container_of(cur_lte->msg_list.next,
964 list_move(&msg->list, &ctx->available_msgs);
965 if (msg->begin_chunk == 0) {
966 /* This is the first set of chunks. Leave space
967 * for the chunk table in the output file. */
968 off_t cur_offset = filedes_offset(ctx->out_fd);
969 if (cur_offset == -1)
970 return WIMLIB_ERR_WRITE;
971 ret = begin_wim_resource_chunk_tab(cur_lte,
974 &ctx->cur_chunk_tab);
979 /* Write the compressed chunks from the message. */
980 ret = write_wim_chunks(msg, ctx->out_fd, ctx->cur_chunk_tab);
984 /* Was this the last chunk of the stream? If so, finish
986 if (list_empty(&cur_lte->msg_list) &&
987 msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
992 ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
998 list_del(&cur_lte->being_compressed_list);
1000 /* Grab the offset of this stream in the output file
1001 * from the chunk table before we free it. */
1002 offset = ctx->cur_chunk_tab->file_offset;
1004 FREE(ctx->cur_chunk_tab);
1005 ctx->cur_chunk_tab = NULL;
1007 if (res_csize >= wim_resource_size(cur_lte)) {
1008 /* Oops! We compressed the resource to
1009 * larger than the original size. Write
1010 * the resource uncompressed instead. */
1011 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
1012 "writing uncompressed instead",
1013 wim_resource_size(cur_lte), res_csize);
1014 ret = seek_and_truncate(ctx->out_fd, offset);
1017 ret = write_wim_resource(cur_lte,
1019 WIMLIB_COMPRESSION_TYPE_NONE,
1020 &cur_lte->output_resource_entry,
1021 ctx->write_resource_flags);
1025 cur_lte->output_resource_entry.size =
1028 cur_lte->output_resource_entry.original_size =
1029 cur_lte->resource_entry.original_size;
1031 cur_lte->output_resource_entry.offset =
1034 cur_lte->output_resource_entry.flags =
1035 cur_lte->resource_entry.flags |
1036 WIM_RESHDR_FLAG_COMPRESSED;
1039 do_write_streams_progress(ctx->progress,
1041 wim_resource_size(cur_lte),
1044 /* Since we just finished writing a stream, write any
1045 * streams that have been added to the serial_streams
1046 * list for direct writing by the main thread (e.g.
1047 * resources that don't need to be compressed because
1048 * the desired compression type is the same as the
1049 * previous compression type). */
1050 if (!list_empty(&ctx->serial_streams)) {
1051 ret = do_write_stream_list_serial(&ctx->serial_streams,
1055 ctx->write_resource_flags,
1062 /* Advance to the next stream to write. */
1063 if (list_empty(&ctx->outstanding_streams)) {
1066 cur_lte = container_of(ctx->outstanding_streams.next,
1067 struct wim_lookup_table_entry,
1068 being_compressed_list);
1075 /* Called when the main thread has read a new chunk of data. */
1077 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
1079 struct main_writer_thread_ctx *ctx = _ctx;
1081 struct message *next_msg;
1082 u64 next_chunk_in_msg;
1084 /* Update SHA1 message digest for the stream currently being read by the
1086 sha1_update(&ctx->next_sha_ctx, chunk, chunk_size);
1088 /* We send chunks of data to the compressor chunks in batches which we
1089 * refer to as "messages". @next_msg is the message that is currently
1090 * being prepared to send off. If it is NULL, that indicates that we
1091 * need to start a new message. */
1092 next_msg = ctx->next_msg;
1094 /* We need to start a new message. First check to see if there
1095 * is a message available in the list of available messages. If
1096 * so, we can just take one. If not, all the messages (there is
1097 * a fixed number of them, proportional to the number of
1098 * threads) have been sent off to the compressor threads, so we
1099 * receive messages from the compressor threads containing
1100 * compressed chunks of data.
1102 * We may need to receive multiple messages before one is
1103 * actually available to use because messages received that are
1104 * *not* for the very next set of chunks to compress must be
1105 * buffered until it's time to write those chunks. */
1106 while (list_empty(&ctx->available_msgs)) {
1107 ret = receive_compressed_chunks(ctx);
1112 next_msg = container_of(ctx->available_msgs.next,
1113 struct message, list);
1114 list_del(&next_msg->list);
1115 next_msg->complete = false;
1116 next_msg->begin_chunk = ctx->next_chunk;
1117 next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
1118 ctx->next_num_chunks - ctx->next_chunk);
1119 ctx->next_msg = next_msg;
1122 /* Fill in the next chunk to compress */
1123 next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;
1125 next_msg->uncompressed_chunk_sizes[next_chunk_in_msg] = chunk_size;
1126 memcpy(next_msg->uncompressed_chunks[next_chunk_in_msg],
1129 if (++next_chunk_in_msg == next_msg->num_chunks) {
1130 /* Send off an array of chunks to compress */
1131 list_add_tail(&next_msg->list, &ctx->next_lte->msg_list);
1132 shared_queue_put(ctx->res_to_compress_queue, next_msg);
1133 ++ctx->num_outstanding_messages;
1134 ctx->next_msg = NULL;
1140 main_writer_thread_finish(void *_ctx)
1142 struct main_writer_thread_ctx *ctx = _ctx;
1144 while (ctx->num_outstanding_messages != 0) {
1145 ret = receive_compressed_chunks(ctx);
1149 wimlib_assert(list_empty(&ctx->outstanding_streams));
1150 return do_write_stream_list_serial(&ctx->serial_streams,
1154 ctx->write_resource_flags,
1160 submit_stream_for_compression(struct wim_lookup_table_entry *lte,
1161 struct main_writer_thread_ctx *ctx)
1165 /* Read the entire stream @lte, feeding its data chunks to the
1166 * compressor threads. Also SHA1-sum the stream; this is required in
1167 * the case that @lte is unhashed, and a nice additional verification
1168 * when @lte is already hashed. */
1169 sha1_init(&ctx->next_sha_ctx);
1170 ctx->next_chunk = 0;
1171 ctx->next_num_chunks = wim_resource_chunks(lte);
1172 ctx->next_lte = lte;
1173 INIT_LIST_HEAD(<e->msg_list);
1174 list_add_tail(<e->being_compressed_list, &ctx->outstanding_streams);
1175 ret = read_resource_prefix(lte, wim_resource_size(lte),
1176 main_writer_thread_cb, ctx, 0);
1178 wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
1179 ret = finalize_and_check_sha1(&ctx->next_sha_ctx, lte);
1185 main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
1187 struct main_writer_thread_ctx *ctx = _ctx;
1190 if (wim_resource_size(lte) < 1000 ||
1191 ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
1192 (lte->resource_location == RESOURCE_IN_WIM &&
1193 !(ctx->write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
1194 wimlib_get_compression_type(lte->wim) == ctx->out_ctype))
1196 /* Stream is too small or isn't being compressed. Process it by
1197 * the main thread when we have a chance. We can't necessarily
1198 * process it right here, as the main thread could be in the
1199 * middle of writing a different stream. */
1200 list_add_tail(<e->write_streams_list, &ctx->serial_streams);
1204 ret = submit_stream_for_compression(lte, ctx);
1206 lte->no_progress = 1;
1211 get_default_num_threads(void)
1214 return win32_get_number_of_processors();
1216 return sysconf(_SC_NPROCESSORS_ONLN);
1220 /* Equivalent to write_stream_list_serial(), except this takes a @num_threads
1221 * parameter and will perform compression using that many threads. Falls
1222 * back to write_stream_list_serial() on certain errors, such as a failure to
1223 * create the number of threads requested.
1225 * High level description of the algorithm for writing compressed streams in
1226 * parallel: We perform compression on chunks of size WIM_CHUNK_SIZE bytes
1227 * rather than on full files. The currently executing thread becomes the main
1228 * thread and is entirely in charge of reading the data to compress (which may
1229 * be in any location understood by the resource code--- such as in an external
1230 * file being captured, or in another WIM file from which an image is being
1231 * exported) and actually writing the compressed data to the output file.
1232 * Additional threads are "compressor threads" and all execute the
1233 * compressor_thread_proc, where they repeatedly retrieve buffers of data from
1234 * the main thread, compress them, and hand them back to the main thread.
1236 * Certain streams, such as streams that do not need to be compressed (e.g.
1237 * input compression type same as output compression type) or streams of very
1238 * small size are placed in a list (main_writer_thread_ctx.serial_list) and
1239 * handled entirely by the main thread at an appropriate time.
1241 * At any given point in time, multiple streams may be having chunks compressed
1242 * concurrently. The stream that the main thread is currently *reading* may be
1243 * later in the list that the stream that the main thread is currently
1247 write_stream_list_parallel(struct list_head *stream_list,
1248 struct wim_lookup_table *lookup_table,
1251 int write_resource_flags,
1252 wimlib_progress_func_t progress_func,
1253 union wimlib_progress_info *progress,
1254 unsigned num_threads)
1257 struct shared_queue res_to_compress_queue;
1258 struct shared_queue compressed_res_queue;
1259 pthread_t *compressor_threads = NULL;
1261 if (num_threads == 0) {
1262 long nthreads = get_default_num_threads();
1263 if (nthreads < 1 || nthreads > UINT_MAX) {
1264 WARNING("Could not determine number of processors! Assuming 1");
1266 } else if (nthreads == 1) {
1267 goto out_serial_quiet;
1269 num_threads = nthreads;
1273 DEBUG("Writing stream list (parallel version, num_threads=%u)",
1276 progress->write_streams.num_threads = num_threads;
1278 static const size_t MESSAGES_PER_THREAD = 2;
1279 size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1281 DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1283 ret = shared_queue_init(&res_to_compress_queue, queue_size);
1287 ret = shared_queue_init(&compressed_res_queue, queue_size);
1289 goto out_destroy_res_to_compress_queue;
1291 struct compressor_thread_params params;
1292 params.res_to_compress_queue = &res_to_compress_queue;
1293 params.compressed_res_queue = &compressed_res_queue;
1294 params.compress = get_compress_func(out_ctype);
1296 compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1297 if (!compressor_threads) {
1298 ret = WIMLIB_ERR_NOMEM;
1299 goto out_destroy_compressed_res_queue;
1302 for (unsigned i = 0; i < num_threads; i++) {
1303 DEBUG("pthread_create thread %u of %u", i + 1, num_threads);
1304 ret = pthread_create(&compressor_threads[i], NULL,
1305 compressor_thread_proc, ¶ms);
1308 ERROR_WITH_ERRNO("Failed to create compressor "
1310 i + 1, num_threads);
1317 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
1319 struct main_writer_thread_ctx ctx;
1320 ctx.stream_list = stream_list;
1321 ctx.lookup_table = lookup_table;
1322 ctx.out_fd = out_fd;
1323 ctx.out_ctype = out_ctype;
1324 ctx.res_to_compress_queue = &res_to_compress_queue;
1325 ctx.compressed_res_queue = &compressed_res_queue;
1326 ctx.num_messages = queue_size;
1327 ctx.write_resource_flags = write_resource_flags;
1328 ctx.progress_func = progress_func;
1329 ctx.progress = progress;
1330 ret = main_writer_thread_init_ctx(&ctx);
1333 ret = do_write_stream_list(stream_list, lookup_table,
1334 main_thread_process_next_stream,
1335 &ctx, progress_func, progress);
1337 goto out_destroy_ctx;
1339 /* The main thread has finished reading all streams that are going to be
1340 * compressed in parallel, and it now needs to wait for all remaining
1341 * chunks to be compressed so that the remaining streams can actually be
1342 * written to the output file. Furthermore, any remaining streams that
1343 * had processing deferred to the main thread need to be handled. These
1344 * tasks are done by the main_writer_thread_finish() function. */
1345 ret = main_writer_thread_finish(&ctx);
1347 main_writer_thread_destroy_ctx(&ctx);
1349 for (unsigned i = 0; i < num_threads; i++)
1350 shared_queue_put(&res_to_compress_queue, NULL);
1352 for (unsigned i = 0; i < num_threads; i++) {
1353 if (pthread_join(compressor_threads[i], NULL)) {
1354 WARNING_WITH_ERRNO("Failed to join compressor "
1356 i + 1, num_threads);
1359 FREE(compressor_threads);
1360 out_destroy_compressed_res_queue:
1361 shared_queue_destroy(&compressed_res_queue);
1362 out_destroy_res_to_compress_queue:
1363 shared_queue_destroy(&res_to_compress_queue);
1364 if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
1367 WARNING("Falling back to single-threaded compression");
1369 return write_stream_list_serial(stream_list,
1373 write_resource_flags,
1381 * Write a list of streams to a WIM (@out_fd) using the compression type
1382 * @out_ctype and up to @num_threads compressor threads.
1385 write_stream_list(struct list_head *stream_list,
1386 struct wim_lookup_table *lookup_table,
1387 int out_fd, int out_ctype, int write_flags,
1388 unsigned num_threads, wimlib_progress_func_t progress_func)
1390 struct wim_lookup_table_entry *lte;
1391 size_t num_streams = 0;
1392 u64 total_bytes = 0;
1393 u64 total_compression_bytes = 0;
1394 union wimlib_progress_info progress;
1396 int write_resource_flags;
1398 if (list_empty(stream_list))
1401 write_resource_flags = write_flags_to_resource_flags(write_flags);
1403 /* Calculate the total size of the streams to be written. Note: this
1404 * will be the uncompressed size, as we may not know the compressed size
1405 * yet, and also this will assume that every unhashed stream will be
1406 * written (which will not necessarily be the case). */
1407 list_for_each_entry(lte, stream_list, write_streams_list) {
1409 total_bytes += wim_resource_size(lte);
1410 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
1411 && (wim_resource_compression_type(lte) != out_ctype ||
1412 (write_resource_flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS)))
1414 total_compression_bytes += wim_resource_size(lte);
1417 progress.write_streams.total_bytes = total_bytes;
1418 progress.write_streams.total_streams = num_streams;
1419 progress.write_streams.completed_bytes = 0;
1420 progress.write_streams.completed_streams = 0;
1421 progress.write_streams.num_threads = num_threads;
1422 progress.write_streams.compression_type = out_ctype;
1423 progress.write_streams._private = 0;
1425 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1426 if (total_compression_bytes >= 2000000 && num_threads != 1)
1427 ret = write_stream_list_parallel(stream_list,
1431 write_resource_flags,
1437 ret = write_stream_list_serial(stream_list,
1441 write_resource_flags,
1447 struct stream_size_table {
1448 struct hlist_head *array;
1454 init_stream_size_table(struct stream_size_table *tab, size_t capacity)
1456 tab->array = CALLOC(capacity, sizeof(tab->array[0]));
1458 return WIMLIB_ERR_NOMEM;
1459 tab->num_entries = 0;
1460 tab->capacity = capacity;
1465 destroy_stream_size_table(struct stream_size_table *tab)
1471 stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
1473 struct stream_size_table *tab = _tab;
1475 struct wim_lookup_table_entry *same_size_lte;
1476 struct hlist_node *tmp;
1478 pos = hash_u64(wim_resource_size(lte)) % tab->capacity;
1479 lte->unique_size = 1;
1480 hlist_for_each_entry(same_size_lte, tmp, &tab->array[pos], hash_list_2) {
1481 if (wim_resource_size(same_size_lte) == wim_resource_size(lte)) {
1482 lte->unique_size = 0;
1483 same_size_lte->unique_size = 0;
1488 hlist_add_head(<e->hash_list_2, &tab->array[pos]);
1494 struct lte_overwrite_prepare_args {
1497 struct list_head stream_list;
1498 struct stream_size_table stream_size_tab;
1501 /* First phase of preparing streams for an in-place overwrite. This is called
1502 * on all streams, both hashed and unhashed, except the metadata resources. */
1504 lte_overwrite_prepare(struct wim_lookup_table_entry *lte, void *_args)
1506 struct lte_overwrite_prepare_args *args = _args;
1508 wimlib_assert(!(lte->resource_entry.flags & WIM_RESHDR_FLAG_METADATA));
1509 if (lte->resource_location != RESOURCE_IN_WIM || lte->wim != args->wim)
1510 list_add_tail(<e->write_streams_list, &args->stream_list);
1511 lte->out_refcnt = lte->refcnt;
1512 stream_size_table_insert(lte, &args->stream_size_tab);
1516 /* Second phase of preparing streams for an in-place overwrite. This is called
1517 * on existing metadata resources and hashed streams, but not unhashed streams.
1519 * NOTE: lte->output_resource_entry is in union with lte->hash_list_2, so
1520 * lte_overwrite_prepare_2() must be called after lte_overwrite_prepare(), as
1521 * the latter uses lte->hash_list_2, while the former expects to set
1522 * lte->output_resource_entry. */
1524 lte_overwrite_prepare_2(struct wim_lookup_table_entry *lte, void *_args)
1526 struct lte_overwrite_prepare_args *args = _args;
1528 if (lte->resource_location == RESOURCE_IN_WIM && lte->wim == args->wim) {
1529 /* We can't do an in place overwrite on the WIM if there are
1530 * streams after the XML data. */
1531 if (lte->resource_entry.offset +
1532 lte->resource_entry.size > args->end_offset)
1534 #ifdef ENABLE_ERROR_MESSAGES
1535 ERROR("The following resource is after the XML data:");
1536 print_lookup_table_entry(lte, stderr);
1538 return WIMLIB_ERR_RESOURCE_ORDER;
1540 copy_resource_entry(<e->output_resource_entry,
1541 <e->resource_entry);
1546 /* Given a WIM that we are going to overwrite in place with zero or more
1547 * additional streams added, construct a list the list of new unique streams
1548 * ('struct wim_lookup_table_entry's) that must be written, plus any unhashed
1549 * streams that need to be added but may be identical to other hashed or
1550 * unhashed streams. These unhashed streams are checksummed while the streams
1551 * are being written. To aid this process, the member @unique_size is set to 1
1552 * on streams that have a unique size and therefore must be written.
1554 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1555 * indicate the number of times the stream is referenced in only the streams
1556 * that are being written; this may still be adjusted later when unhashed
1557 * streams are being resolved.
1560 prepare_streams_for_overwrite(WIMStruct *wim, off_t end_offset,
1561 struct list_head *stream_list)
1564 struct lte_overwrite_prepare_args args;
1568 args.end_offset = end_offset;
1569 ret = init_stream_size_table(&args.stream_size_tab,
1570 wim->lookup_table->capacity);
1574 INIT_LIST_HEAD(&args.stream_list);
1575 for (i = 0; i < wim->hdr.image_count; i++) {
1576 struct wim_image_metadata *imd;
1577 struct wim_lookup_table_entry *lte;
1579 imd = wim->image_metadata[i];
1580 image_for_each_unhashed_stream(lte, imd)
1581 lte_overwrite_prepare(lte, &args);
1583 for_lookup_table_entry(wim->lookup_table, lte_overwrite_prepare, &args);
1584 list_transfer(&args.stream_list, stream_list);
1586 for (i = 0; i < wim->hdr.image_count; i++) {
1587 ret = lte_overwrite_prepare_2(wim->image_metadata[i]->metadata_lte,
1590 goto out_destroy_stream_size_table;
1592 ret = for_lookup_table_entry(wim->lookup_table,
1593 lte_overwrite_prepare_2, &args);
1594 out_destroy_stream_size_table:
1595 destroy_stream_size_table(&args.stream_size_tab);
1600 struct find_streams_ctx {
1601 struct list_head stream_list;
1602 struct stream_size_table stream_size_tab;
1606 inode_find_streams_to_write(struct wim_inode *inode,
1607 struct wim_lookup_table *table,
1608 struct list_head *stream_list,
1609 struct stream_size_table *tab)
1611 struct wim_lookup_table_entry *lte;
1612 for (unsigned i = 0; i <= inode->i_num_ads; i++) {
1613 lte = inode_stream_lte(inode, i, table);
1615 if (lte->out_refcnt == 0) {
1617 stream_size_table_insert(lte, tab);
1618 list_add_tail(<e->write_streams_list, stream_list);
1620 lte->out_refcnt += inode->i_nlink;
1626 image_find_streams_to_write(WIMStruct *w)
1628 struct find_streams_ctx *ctx;
1629 struct wim_image_metadata *imd;
1630 struct wim_inode *inode;
1631 struct wim_lookup_table_entry *lte;
1634 imd = wim_get_current_image_metadata(w);
1636 image_for_each_unhashed_stream(lte, imd)
1637 lte->out_refcnt = 0;
1639 /* Go through this image's inodes to find any streams that have not been
1641 image_for_each_inode(inode, imd) {
1642 inode_find_streams_to_write(inode, w->lookup_table,
1644 &ctx->stream_size_tab);
1649 /* Given a WIM that from which one or all of the images is being written, build
1650 * the list of unique streams ('struct wim_lookup_table_entry's) that must be
1651 * written, plus any unhashed streams that need to be written but may be
1652 * identical to other hashed or unhashed streams being written. These unhashed
1653 * streams are checksummed while the streams are being written. To aid this
1654 * process, the member @unique_size is set to 1 on streams that have a unique
1655 * size and therefore must be written.
1657 * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1658 * indicate the number of times the stream is referenced in only the streams
1659 * that are being written; this may still be adjusted later when unhashed
1660 * streams are being resolved.
1663 prepare_stream_list(WIMStruct *wim, int image, struct list_head *stream_list)
1666 struct find_streams_ctx ctx;
1668 for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
1669 ret = init_stream_size_table(&ctx.stream_size_tab,
1670 wim->lookup_table->capacity);
1673 for_lookup_table_entry(wim->lookup_table, stream_size_table_insert,
1674 &ctx.stream_size_tab);
1675 INIT_LIST_HEAD(&ctx.stream_list);
1676 wim->private = &ctx;
1677 ret = for_image(wim, image, image_find_streams_to_write);
1678 destroy_stream_size_table(&ctx.stream_size_tab);
1680 list_transfer(&ctx.stream_list, stream_list);
1684 /* Writes the streams for the specified @image in @wim to @wim->out_fd.
1687 write_wim_streams(WIMStruct *wim, int image, int write_flags,
1688 unsigned num_threads,
1689 wimlib_progress_func_t progress_func)
1692 struct list_head stream_list;
1694 ret = prepare_stream_list(wim, image, &stream_list);
1697 return write_stream_list(&stream_list,
1700 wimlib_get_compression_type(wim),
1707 * Finish writing a WIM file: write the lookup table, xml data, and integrity
1708 * table (optional), then overwrite the WIM header.
1710 * write_flags is a bitwise OR of the following:
1712 * (public) WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
1713 * Include an integrity table.
1715 * (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
1716 * Don't write the lookup table.
1718 * (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
1719 * When (if) writing the integrity table, re-use entries from the
1720 * existing integrity table, if possible.
1722 * (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
1723 * After writing the XML data but before writing the integrity
1724 * table, write a temporary WIM header and flush the stream so that
1725 * the WIM is less likely to become corrupted upon abrupt program
1728 * (private) WIMLIB_WRITE_FLAG_FSYNC:
1729 * fsync() the output file before closing it.
1733 finish_write(WIMStruct *w, int image, int write_flags,
1734 wimlib_progress_func_t progress_func)
1737 struct wim_header hdr;
1739 /* @hdr will be the header for the new WIM. First copy all the data
1740 * from the header in the WIMStruct; then set all the fields that may
1741 * have changed, including the resource entries, boot index, and image
1743 memcpy(&hdr, &w->hdr, sizeof(struct wim_header));
1745 /* Set image count and boot index correctly for single image writes */
1746 if (image != WIMLIB_ALL_IMAGES) {
1747 hdr.image_count = 1;
1748 if (hdr.boot_idx == image)
1754 /* In the WIM header, there is room for the resource entry for a
1755 * metadata resource labeled as the "boot metadata". This entry should
1756 * be zeroed out if there is no bootable image (boot_idx 0). Otherwise,
1757 * it should be a copy of the resource entry for the image that is
1758 * marked as bootable. This is not well documented... */
1759 if (hdr.boot_idx == 0) {
1760 zero_resource_entry(&hdr.boot_metadata_res_entry);
1762 copy_resource_entry(&hdr.boot_metadata_res_entry,
1763 &w->image_metadata[ hdr.boot_idx- 1
1764 ]->metadata_lte->output_resource_entry);
1767 if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
1768 ret = write_lookup_table(w, image, &hdr.lookup_table_res_entry);
1773 ret = write_xml_data(w->wim_info, image, w->out_fd,
1774 (write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE) ?
1775 wim_info_get_total_bytes(w->wim_info) : 0,
1776 &hdr.xml_res_entry);
1780 if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
1781 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
1782 struct wim_header checkpoint_hdr;
1783 memcpy(&checkpoint_hdr, &hdr, sizeof(struct wim_header));
1784 zero_resource_entry(&checkpoint_hdr.integrity);
1785 ret = write_header(&checkpoint_hdr, w->out_fd);
1790 off_t old_lookup_table_end;
1791 off_t new_lookup_table_end;
1792 if (write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE) {
1793 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
1794 w->hdr.lookup_table_res_entry.size;
1796 old_lookup_table_end = 0;
1798 new_lookup_table_end = hdr.lookup_table_res_entry.offset +
1799 hdr.lookup_table_res_entry.size;
1801 ret = write_integrity_table(w->out_fd,
1803 new_lookup_table_end,
1804 old_lookup_table_end,
1809 zero_resource_entry(&hdr.integrity);
1812 ret = write_header(&hdr, w->out_fd);
1816 if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
1817 if (fsync(w->out_fd)) {
1818 ERROR_WITH_ERRNO("Error syncing data to WIM file");
1819 ret = WIMLIB_ERR_WRITE;
1823 if (close(w->out_fd)) {
1824 ERROR_WITH_ERRNO("Failed to close the output WIM file");
1826 ret = WIMLIB_ERR_WRITE;
1832 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
1834 lock_wim(WIMStruct *w, int fd)
1837 if (fd != -1 && !w->wim_locked) {
1838 ret = flock(fd, LOCK_EX | LOCK_NB);
1840 if (errno == EWOULDBLOCK) {
1841 ERROR("`%"TS"' is already being modified or has been "
1842 "mounted read-write\n"
1843 " by another process!", w->filename);
1844 ret = WIMLIB_ERR_ALREADY_LOCKED;
1846 WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
1859 open_wim_writable(WIMStruct *w, const tchar *path, int open_flags)
1861 w->out_fd = topen(path, open_flags | O_BINARY, 0644);
1862 if (w->out_fd == -1) {
1863 ERROR_WITH_ERRNO("Failed to open `%"TS"' for writing", path);
1864 return WIMLIB_ERR_OPEN;
1871 close_wim_writable(WIMStruct *w)
1873 if (w->out_fd != -1) {
1874 if (close(w->out_fd))
1875 WARNING_WITH_ERRNO("Failed to close output WIM");
1880 /* Open file stream and write dummy header for WIM. */
1882 begin_write(WIMStruct *w, const tchar *path, int write_flags)
1885 int open_flags = O_TRUNC | O_CREAT;
1886 if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
1887 open_flags |= O_RDWR;
1889 open_flags |= O_WRONLY;
1890 ret = open_wim_writable(w, path, open_flags);
1893 /* Write dummy header. It will be overwritten later. */
1894 ret = write_header(&w->hdr, w->out_fd);
1897 if (lseek(w->out_fd, WIM_HEADER_DISK_SIZE, SEEK_SET) == -1) {
1898 ERROR_WITH_ERRNO("Failed to seek to end of WIM header");
1899 return WIMLIB_ERR_WRITE;
1904 /* Writes a stand-alone WIM to a file. */
1906 wimlib_write(WIMStruct *w, const tchar *path,
1907 int image, int write_flags, unsigned num_threads,
1908 wimlib_progress_func_t progress_func)
1913 return WIMLIB_ERR_INVALID_PARAM;
1915 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1917 if (image != WIMLIB_ALL_IMAGES &&
1918 (image < 1 || image > w->hdr.image_count))
1919 return WIMLIB_ERR_INVALID_IMAGE;
1921 if (w->hdr.total_parts != 1) {
1922 ERROR("Cannot call wimlib_write() on part of a split WIM");
1923 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1926 ret = begin_write(w, path, write_flags);
1930 ret = write_wim_streams(w, image, write_flags, num_threads,
1936 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
1938 ret = for_image(w, image, write_metadata_resource);
1943 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
1945 ret = finish_write(w, image, write_flags, progress_func);
1946 /* finish_write() closed the WIM for us */
1949 close_wim_writable(w);
1951 DEBUG("wimlib_write(path=%"TS") = %d", path, ret);
1956 any_images_modified(WIMStruct *w)
1958 for (int i = 0; i < w->hdr.image_count; i++)
1959 if (w->image_metadata[i]->modified)
1965 * Overwrite a WIM, possibly appending streams to it.
1967 * A WIM looks like (or is supposed to look like) the following:
1969 * Header (212 bytes)
1970 * Streams and metadata resources (variable size)
1971 * Lookup table (variable size)
1972 * XML data (variable size)
1973 * Integrity table (optional) (variable size)
1975 * If we are not adding any streams or metadata resources, the lookup table is
1976 * unchanged--- so we only need to overwrite the XML data, integrity table, and
1977 * header. This operation is potentially unsafe if the program is abruptly
1978 * terminated while the XML data or integrity table are being overwritten, but
1979 * before the new header has been written. To partially alleviate this problem,
1980 * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
1981 * finish_write() to cause a temporary WIM header to be written after the XML
1982 * data has been written. This may prevent the WIM from becoming corrupted if
1983 * the program is terminated while the integrity table is being calculated (but
1984 * no guarantees, due to write re-ordering...).
1986 * If we are adding new streams or images (metadata resources), the lookup table
1987 * needs to be changed, and those streams need to be written. In this case, we
1988 * try to perform a safe update of the WIM file by writing the streams *after*
1989 * the end of the previous WIM, then writing the new lookup table, XML data, and
1990 * (optionally) integrity table following the new streams. This will produce a
1991 * layout like the following:
1993 * Header (212 bytes)
1994 * (OLD) Streams and metadata resources (variable size)
1995 * (OLD) Lookup table (variable size)
1996 * (OLD) XML data (variable size)
1997 * (OLD) Integrity table (optional) (variable size)
1998 * (NEW) Streams and metadata resources (variable size)
1999 * (NEW) Lookup table (variable size)
2000 * (NEW) XML data (variable size)
2001 * (NEW) Integrity table (optional) (variable size)
2003 * At all points, the WIM is valid as nothing points to the new data yet. Then,
2004 * the header is overwritten to point to the new lookup table, XML data, and
2005 * integrity table, to produce the following layout:
2007 * Header (212 bytes)
2008 * Streams and metadata resources (variable size)
2009 * Nothing (variable size)
2010 * More Streams and metadata resources (variable size)
2011 * Lookup table (variable size)
2012 * XML data (variable size)
2013 * Integrity table (optional) (variable size)
2015 * This method allows an image to be appended to a large WIM very quickly, and
2016 * is is crash-safe except in the case of write re-ordering, but the
2017 * disadvantage is that a small hole is left in the WIM where the old lookup
2018 * table, xml data, and integrity table were. (These usually only take up a
2019 * small amount of space compared to the streams, however.)
2022 overwrite_wim_inplace(WIMStruct *w, int write_flags,
2023 unsigned num_threads,
2024 wimlib_progress_func_t progress_func)
2027 struct list_head stream_list;
2029 u64 old_lookup_table_end, old_xml_begin, old_xml_end;
2032 DEBUG("Overwriting `%"TS"' in-place", w->filename);
2034 /* Make sure that the integrity table (if present) is after the XML
2035 * data, and that there are no stream resources, metadata resources, or
2036 * lookup tables after the XML data. Otherwise, these data would be
2038 old_xml_begin = w->hdr.xml_res_entry.offset;
2039 old_xml_end = old_xml_begin + w->hdr.xml_res_entry.size;
2040 old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
2041 w->hdr.lookup_table_res_entry.size;
2042 if (w->hdr.integrity.offset != 0 && w->hdr.integrity.offset < old_xml_end) {
2043 ERROR("Didn't expect the integrity table to be before the XML data");
2044 return WIMLIB_ERR_RESOURCE_ORDER;
2047 if (old_lookup_table_end > old_xml_begin) {
2048 ERROR("Didn't expect the lookup table to be after the XML data");
2049 return WIMLIB_ERR_RESOURCE_ORDER;
2052 /* Set @old_wim_end, which indicates the point beyond which we don't
2053 * allow any file and metadata resources to appear without returning
2054 * WIMLIB_ERR_RESOURCE_ORDER (due to the fact that we would otherwise
2055 * overwrite these resources). */
2056 if (!w->deletion_occurred && !any_images_modified(w)) {
2057 /* If no images have been modified and no images have been
2058 * deleted, a new lookup table does not need to be written. We
2059 * shall write the new XML data and optional integrity table
2060 * immediately after the lookup table. Note that this may
2061 * overwrite an existing integrity table. */
2062 DEBUG("Skipping writing lookup table "
2063 "(no images modified or deleted)");
2064 old_wim_end = old_lookup_table_end;
2065 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
2066 WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
2067 } else if (w->hdr.integrity.offset) {
2068 /* Old WIM has an integrity table; begin writing new streams
2070 old_wim_end = w->hdr.integrity.offset + w->hdr.integrity.size;
2072 /* No existing integrity table; begin writing new streams after
2073 * the old XML data. */
2074 old_wim_end = old_xml_end;
2077 ret = prepare_streams_for_overwrite(w, old_wim_end, &stream_list);
2082 if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
2083 open_flags |= O_RDWR;
2085 open_flags |= O_WRONLY;
2086 ret = open_wim_writable(w, w->filename, open_flags);
2090 ret = lock_wim(w, w->out_fd);
2092 close_wim_writable(w);
2096 if (lseek(w->out_fd, old_wim_end, SEEK_SET) == -1) {
2097 ERROR_WITH_ERRNO("Can't seek to end of WIM");
2098 close_wim_writable(w);
2100 return WIMLIB_ERR_WRITE;
2103 DEBUG("Writing newly added streams (offset = %"PRIu64")",
2105 ret = write_stream_list(&stream_list,
2108 wimlib_get_compression_type(w),
2115 for (int i = 0; i < w->hdr.image_count; i++) {
2116 if (w->image_metadata[i]->modified) {
2117 select_wim_image(w, i + 1);
2118 ret = write_metadata_resource(w);
2123 write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
2124 ret = finish_write(w, WIMLIB_ALL_IMAGES, write_flags,
2127 close_wim_writable(w);
2128 if (ret != 0 && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
2129 WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
2130 w->filename, old_wim_end);
2131 /* Return value of truncate() is ignored because this is already
2133 (void)ttruncate(w->filename, old_wim_end);
2140 overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
2141 unsigned num_threads,
2142 wimlib_progress_func_t progress_func)
2144 size_t wim_name_len;
2147 DEBUG("Overwriting `%"TS"' via a temporary file", w->filename);
2149 /* Write the WIM to a temporary file in the same directory as the
2151 wim_name_len = tstrlen(w->filename);
2152 tchar tmpfile[wim_name_len + 10];
2153 tmemcpy(tmpfile, w->filename, wim_name_len);
2154 randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
2155 tmpfile[wim_name_len + 9] = T('\0');
2157 ret = wimlib_write(w, tmpfile, WIMLIB_ALL_IMAGES,
2158 write_flags | WIMLIB_WRITE_FLAG_FSYNC,
2159 num_threads, progress_func);
2161 ERROR("Failed to write the WIM file `%"TS"'", tmpfile);
2167 DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, w->filename);
2168 /* Rename the new file to the old file .*/
2169 if (trename(tmpfile, w->filename) != 0) {
2170 ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
2171 tmpfile, w->filename);
2172 ret = WIMLIB_ERR_RENAME;
2176 if (progress_func) {
2177 union wimlib_progress_info progress;
2178 progress.rename.from = tmpfile;
2179 progress.rename.to = w->filename;
2180 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
2184 /* Remove temporary file. */
2185 if (tunlink(tmpfile) != 0)
2186 WARNING_WITH_ERRNO("Failed to remove `%"TS"'", tmpfile);
2192 * Writes a WIM file to the original file that it was read from, overwriting it.
2195 wimlib_overwrite(WIMStruct *w, int write_flags,
2196 unsigned num_threads,
2197 wimlib_progress_func_t progress_func)
2199 write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2202 return WIMLIB_ERR_NO_FILENAME;
2204 if (w->hdr.total_parts != 1) {
2205 ERROR("Cannot modify a split WIM");
2206 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
2209 if ((!w->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
2210 && !(write_flags & WIMLIB_WRITE_FLAG_REBUILD))
2213 ret = overwrite_wim_inplace(w, write_flags, num_threads,
2215 if (ret == WIMLIB_ERR_RESOURCE_ORDER)
2216 WARNING("Falling back to re-building entire WIM");
2220 return overwrite_wim_via_tmpfile(w, write_flags, num_threads,