write.c, lookup table.c: cleanup
[wimlib] / src / write.c
1 /*
2  * write.c
3  *
4  * Support for writing WIM files; write a WIM file, overwrite a WIM file, write
5  * compressed file resources, etc.
6  */
7
8 /*
9  * Copyright (C) 2012, 2013 Eric Biggers
10  *
11  * This file is part of wimlib, a library for working with WIM files.
12  *
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)
16  * any later version.
17  *
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
21  * details.
22  *
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/.
25  */
26
27 #include "config.h"
28
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>
33 #endif
34
35 #ifdef __WIN32__
36 #  include "win32.h"
37 #endif
38
39 #include "list.h"
40 #include "wimlib_internal.h"
41 #include "buffer_io.h"
42 #include "dentry.h"
43 #include "lookup_table.h"
44 #include "xml.h"
45
46 #ifdef ENABLE_MULTITHREADED_COMPRESSION
47 #  include <pthread.h>
48 #endif
49
50 #include <unistd.h>
51 #include <errno.h>
52
53 #ifdef WITH_NTFS_3G
54 #  include <time.h>
55 #  include <ntfs-3g/attrib.h>
56 #  include <ntfs-3g/inode.h>
57 #  include <ntfs-3g/dir.h>
58 #endif
59
60 #ifdef HAVE_ALLOCA_H
61 #  include <alloca.h>
62 #else
63 #  include <stdlib.h>
64 #endif
65
66 #include <limits.h>
67
68 #if defined(__WIN32__) && !defined(INVALID_HANDLE_VALUE)
69 #  define INVALID_HANDLE_VALUE ((HANDLE)(-1))
70 #endif
71
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.) */
75 struct chunk_table {
76         off_t file_offset;
77         u64 num_chunks;
78         u64 original_resource_size;
79         u64 bytes_per_chunk_entry;
80         u64 table_disk_size;
81         u64 cur_offset;
82         u64 *cur_offset_p;
83         u64 offsets[0];
84 };
85
86 /*
87  * Allocates and initializes a chunk table, and reserves space for it in the
88  * output file.
89  */
90 static int
91 begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
92                              FILE *out_fp,
93                              off_t file_offset,
94                              struct chunk_table **chunk_tab_ret)
95 {
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);
100         int ret;
101
102         if (!chunk_tab) {
103                 ERROR("Failed to allocate chunk table for %"PRIu64" byte "
104                       "resource", size);
105                 ret = WIMLIB_ERR_NOMEM;
106                 goto out;
107         }
108         chunk_tab->file_offset = file_offset;
109         chunk_tab->num_chunks = num_chunks;
110         chunk_tab->original_resource_size = size;
111         chunk_tab->bytes_per_chunk_entry = (size >= (1ULL << 32)) ? 8 : 4;
112         chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
113                                      (num_chunks - 1);
114         chunk_tab->cur_offset = 0;
115         chunk_tab->cur_offset_p = chunk_tab->offsets;
116
117         if (fwrite(chunk_tab, 1, chunk_tab->table_disk_size, out_fp) !=
118                    chunk_tab->table_disk_size) {
119                 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
120                                  "file resource");
121                 FREE(chunk_tab);
122                 ret = WIMLIB_ERR_WRITE;
123                 goto out;
124         }
125
126         ret = 0;
127         *chunk_tab_ret = chunk_tab;
128 out:
129         return ret;
130 }
131
132 /*
133  * compress_func_t- Pointer to a function to compresses a chunk
134  *                  of a WIM resource.  This may be either
135  *                  wimlib_xpress_compress() (xpress-compress.c) or
136  *                  wimlib_lzx_compress() (lzx-compress.c).
137  *
138  * @chunk:        Uncompressed data of the chunk.
139  * @chunk_size:   Size of the uncompressed chunk, in bytes.
140  * @out:          Pointer to output buffer of size at least (@chunk_size - 1) bytes.
141  *
142  * Returns the size of the compressed data written to @out in bytes, or 0 if the
143  * data could not be compressed to (@chunk_size - 1) bytes or fewer.
144  *
145  * As a special requirement, the compression code is optimized for the WIM
146  * format and therefore requires (@chunk_size <= 32768).
147  *
148  * As another special requirement, the compression code will read up to 8 bytes
149  * off the end of the @chunk array for performance reasons.  The values of these
150  * bytes will not affect the output of the compression, but the calling code
151  * must make sure that the buffer holding the uncompressed chunk is actually at
152  * least (@chunk_size + 8) bytes, or at least that these extra bytes are in
153  * mapped memory that will not cause a memory access violation if accessed.
154  */
155 typedef unsigned (*compress_func_t)(const void *chunk, unsigned chunk_size,
156                                     void *out);
157
158 compress_func_t
159 get_compress_func(int out_ctype)
160 {
161         if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX)
162                 return wimlib_lzx_compress;
163         else
164                 return wimlib_xpress_compress;
165 }
166
167 /*
168  * Writes a chunk of a WIM resource to an output file.
169  *
170  * @chunk:        Uncompressed data of the chunk.
171  * @chunk_size:   Size of the chunk (<= WIM_CHUNK_SIZE)
172  * @out_fp:       FILE * to write the chunk to.
173  * @compress:     Compression function to use (NULL if writing uncompressed
174  *                      data).
175  * @chunk_tab:    Pointer to chunk table being created.  It is updated with the
176  *                      offset of the chunk we write.
177  *
178  * Returns 0 on success; nonzero on failure.
179  */
180 static int
181 write_wim_resource_chunk(const void *chunk, unsigned chunk_size,
182                          FILE *out_fp, compress_func_t compress,
183                          struct chunk_table *chunk_tab)
184 {
185         const u8 *out_chunk;
186         unsigned out_chunk_size;
187         if (compress) {
188                 u8 *compressed_chunk = alloca(chunk_size);
189
190                 out_chunk_size = compress(chunk, chunk_size, compressed_chunk);
191                 if (out_chunk_size) {
192                         /* Write compressed */
193                         out_chunk = compressed_chunk;
194                 } else {
195                         /* Write uncompressed */
196                         out_chunk = chunk;
197                         out_chunk_size = chunk_size;
198                 }
199                 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
200                 chunk_tab->cur_offset += out_chunk_size;
201         } else {
202                 /* Write uncompressed */
203                 out_chunk = chunk;
204                 out_chunk_size = chunk_size;
205         }
206         if (fwrite(out_chunk, 1, out_chunk_size, out_fp) != out_chunk_size) {
207                 ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
208                 return WIMLIB_ERR_WRITE;
209         }
210         return 0;
211 }
212
213 /*
214  * Finishes a WIM chunk table and writes it to the output file at the correct
215  * offset.
216  *
217  * The final size of the full compressed resource is returned in the
218  * @compressed_size_p.
219  */
220 static int
221 finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
222                               FILE *out_fp, u64 *compressed_size_p)
223 {
224         size_t bytes_written;
225         if (fseeko(out_fp, chunk_tab->file_offset, SEEK_SET) != 0) {
226                 ERROR_WITH_ERRNO("Failed to seek to byte %"PRIu64" of output "
227                                  "WIM file", chunk_tab->file_offset);
228                 return WIMLIB_ERR_WRITE;
229         }
230
231         if (chunk_tab->bytes_per_chunk_entry == 8) {
232                 array_cpu_to_le64(chunk_tab->offsets, chunk_tab->num_chunks);
233         } else {
234                 for (u64 i = 0; i < chunk_tab->num_chunks; i++)
235                         ((u32*)chunk_tab->offsets)[i] =
236                                 cpu_to_le32(chunk_tab->offsets[i]);
237         }
238         bytes_written = fwrite((u8*)chunk_tab->offsets +
239                                         chunk_tab->bytes_per_chunk_entry,
240                                1, chunk_tab->table_disk_size, out_fp);
241         if (bytes_written != chunk_tab->table_disk_size) {
242                 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
243                                  "file resource");
244                 return WIMLIB_ERR_WRITE;
245         }
246         if (fseeko(out_fp, 0, SEEK_END) != 0) {
247                 ERROR_WITH_ERRNO("Failed to seek to end of output WIM file");
248                 return WIMLIB_ERR_WRITE;
249         }
250         *compressed_size_p = chunk_tab->cur_offset + chunk_tab->table_disk_size;
251         return 0;
252 }
253
254 static int
255 finalize_and_check_sha1(SHA_CTX *sha_ctx, struct wim_lookup_table_entry *lte)
256 {
257         u8 md[SHA1_HASH_SIZE];
258         sha1_final(md, sha_ctx);
259         if (lte->unhashed) {
260                 copy_hash(lte->hash, md);
261         } else if (!hashes_equal(md, lte->hash)) {
262                 ERROR("WIM resource has incorrect hash!");
263                 if (lte_filename_valid(lte)) {
264                         ERROR("We were reading it from \"%"TS"\"; maybe "
265                               "it changed while we were reading it.",
266                               lte->file_on_disk);
267                 }
268                 return WIMLIB_ERR_INVALID_RESOURCE_HASH;
269         }
270         return 0;
271 }
272
273
274 struct write_resource_ctx {
275         compress_func_t compress;
276         struct chunk_table *chunk_tab;
277         FILE *out_fp;
278         SHA_CTX sha_ctx;
279         bool doing_sha;
280 };
281
282 static int
283 write_resource_cb(const void *chunk, size_t chunk_size, void *_ctx)
284 {
285         struct write_resource_ctx *ctx = _ctx;
286
287         if (ctx->doing_sha)
288                 sha1_update(&ctx->sha_ctx, chunk, chunk_size);
289
290         if (ctx->compress) {
291                 return write_wim_resource_chunk(chunk, chunk_size,
292                                                 ctx->out_fp, ctx->compress,
293                                                 ctx->chunk_tab);
294         } else {
295                 if (fwrite(chunk, 1, chunk_size, ctx->out_fp) != chunk_size) {
296                         ERROR_WITH_ERRNO("Error writing to output WIM");
297                         return WIMLIB_ERR_WRITE;
298                 } else {
299                         return 0;
300                 }
301         }
302 }
303
304 /*
305  * Write a resource to an output WIM.
306  *
307  * @lte:  Lookup table entry for the resource, which could be in another WIM,
308  *        in an external file, or in another location.
309  *
310  * @out_fp:  FILE * opened to the output WIM.
311  *
312  * @out_ctype:  One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate
313  *              which compression algorithm to use.
314  *
315  * @out_res_entry:  On success, this is filled in with the offset, flags,
316  *                  compressed size, and uncompressed size of the resource
317  *                  in the output WIM.
318  *
319  * @flags:  WIMLIB_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed
320  *          even if it could otherwise be copied directly from the input.
321  *
322  * Additional notes:  The SHA1 message digest of the uncompressed data is
323  * calculated (except when doing a raw copy --- see below).  If the @unhashed
324  * flag is set on the lookup table entry, this message digest is simply copied
325  * to it; otherwise, the message digest is compared with the existing one, and
326  * the function will fail if they do not match.
327  */
328 int
329 write_wim_resource(struct wim_lookup_table_entry *lte,
330                    FILE *out_fp, int out_ctype,
331                    struct resource_entry *out_res_entry,
332                    int flags)
333 {
334         struct write_resource_ctx write_ctx;
335         u64 read_size;
336         u64 new_size;
337         off_t offset;
338         int ret;
339
340         flags &= ~WIMLIB_RESOURCE_FLAG_RECOMPRESS;
341
342         if (wim_resource_size(lte) == 0) {
343                 /* Empty resource; nothing needs to be done, so just return
344                  * success. */
345                 return 0;
346         }
347
348         /* Get current position in output WIM */
349         offset = ftello(out_fp);
350         if (offset == -1) {
351                 ERROR_WITH_ERRNO("Can't get position in output WIM");
352                 return WIMLIB_ERR_WRITE;
353         }
354
355         /* If we are not forcing the data to be recompressed, and the input
356          * resource is located in a WIM with the same compression type as that
357          * desired other than no compression, we can simply copy the compressed
358          * data without recompressing it.  This also means we must skip
359          * calculating the SHA1, as we never will see the uncompressed data. */
360         if (!(flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS) &&
361             lte->resource_location == RESOURCE_IN_WIM &&
362             out_ctype != WIMLIB_COMPRESSION_TYPE_NONE &&
363             wimlib_get_compression_type(lte->wim) == out_ctype)
364         {
365                 flags |= WIMLIB_RESOURCE_FLAG_RAW;
366                 write_ctx.doing_sha = false;
367                 read_size = lte->resource_entry.size;
368         } else {
369                 write_ctx.doing_sha = true;
370                 sha1_init(&write_ctx.sha_ctx);
371                 read_size = lte->resource_entry.original_size;
372         }
373
374         /* Initialize the chunk table and set the compression function if
375          * compressing the resource. */
376         if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
377             (flags & WIMLIB_RESOURCE_FLAG_RAW)) {
378                 write_ctx.compress = NULL;
379                 write_ctx.chunk_tab = NULL;
380         } else {
381                 write_ctx.compress = get_compress_func(out_ctype);
382                 ret = begin_wim_resource_chunk_tab(lte, out_fp,
383                                                    offset,
384                                                    &write_ctx.chunk_tab);
385                 if (ret)
386                         return ret;
387         }
388
389         /* Write the entire resource by reading the entire resource and feeding
390          * the data through the write_resource_cb function. */
391         write_ctx.out_fp = out_fp;
392 try_write_again:
393         ret = read_resource_prefix(lte, read_size,
394                                    write_resource_cb, &write_ctx, flags);
395         if (ret)
396                 goto out_free_chunk_tab;
397
398         /* Verify SHA1 message digest of the resource, or set the hash for the
399          * first time. */
400         if (write_ctx.doing_sha) {
401                 ret = finalize_and_check_sha1(&write_ctx.sha_ctx, lte);
402                 if (ret)
403                         goto out_free_chunk_tab;
404         }
405
406         out_res_entry->flags = lte->resource_entry.flags;
407         out_res_entry->original_size = wim_resource_size(lte);
408         out_res_entry->offset = offset;
409         if (flags & WIMLIB_RESOURCE_FLAG_RAW) {
410                 /* Doing a raw write:  The new compressed size is the same as
411                  * the compressed size in the other WIM. */
412                 new_size = lte->resource_entry.size;
413         } else if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE) {
414                 /* Using WIMLIB_COMPRESSION_TYPE_NONE:  The new compressed size
415                  * is the original size. */
416                 new_size = lte->resource_entry.original_size;
417                 out_res_entry->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
418         } else {
419                 /* Using a different compression type:  Call
420                  * finish_wim_resource_chunk_tab() and it will provide the new
421                  * compressed size. */
422                 ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab, out_fp,
423                                                     &new_size);
424                 if (ret)
425                         goto out_free_chunk_tab;
426                 if (new_size >= wim_resource_size(lte)) {
427                         /* Oops!  We compressed the resource to larger than the original
428                          * size.  Write the resource uncompressed instead. */
429                         if (fseeko(out_fp, offset, SEEK_SET) ||
430                             fflush(out_fp) ||
431                             ftruncate(fileno(out_fp),
432                                       offset + wim_resource_size(lte)))
433                         {
434                                 ERROR_WITH_ERRNO("Failed to flush and/or truncate "
435                                                  "output WIM file");
436                                 ret = WIMLIB_ERR_WRITE;
437                                 goto out_free_chunk_tab;
438                         }
439                         DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
440                               "writing uncompressed instead",
441                               wim_resource_size(lte), new_size);
442                         write_ctx.compress = NULL;
443                         write_ctx.doing_sha = false;
444                         out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
445                         goto try_write_again;
446                 }
447                 out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
448         }
449         out_res_entry->size = new_size;
450         ret = 0;
451 out_free_chunk_tab:
452         FREE(write_ctx.chunk_tab);
453         return ret;
454 }
455
456 #ifdef ENABLE_MULTITHREADED_COMPRESSION
457
458 /* Blocking shared queue (solves the producer-consumer problem) */
459 struct shared_queue {
460         unsigned size;
461         unsigned front;
462         unsigned back;
463         unsigned filled_slots;
464         void **array;
465         pthread_mutex_t lock;
466         pthread_cond_t msg_avail_cond;
467         pthread_cond_t space_avail_cond;
468 };
469
470 static int
471 shared_queue_init(struct shared_queue *q, unsigned size)
472 {
473         wimlib_assert(size != 0);
474         q->array = CALLOC(sizeof(q->array[0]), size);
475         if (!q->array)
476                 return WIMLIB_ERR_NOMEM;
477         q->filled_slots = 0;
478         q->front = 0;
479         q->back = size - 1;
480         q->size = size;
481         pthread_mutex_init(&q->lock, NULL);
482         pthread_cond_init(&q->msg_avail_cond, NULL);
483         pthread_cond_init(&q->space_avail_cond, NULL);
484         return 0;
485 }
486
487 static void
488 shared_queue_destroy(struct shared_queue *q)
489 {
490         FREE(q->array);
491         pthread_mutex_destroy(&q->lock);
492         pthread_cond_destroy(&q->msg_avail_cond);
493         pthread_cond_destroy(&q->space_avail_cond);
494 }
495
496 static void
497 shared_queue_put(struct shared_queue *q, void *obj)
498 {
499         pthread_mutex_lock(&q->lock);
500         while (q->filled_slots == q->size)
501                 pthread_cond_wait(&q->space_avail_cond, &q->lock);
502
503         q->back = (q->back + 1) % q->size;
504         q->array[q->back] = obj;
505         q->filled_slots++;
506
507         pthread_cond_broadcast(&q->msg_avail_cond);
508         pthread_mutex_unlock(&q->lock);
509 }
510
511 static void *
512 shared_queue_get(struct shared_queue *q)
513 {
514         void *obj;
515
516         pthread_mutex_lock(&q->lock);
517         while (q->filled_slots == 0)
518                 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
519
520         obj = q->array[q->front];
521         q->array[q->front] = NULL;
522         q->front = (q->front + 1) % q->size;
523         q->filled_slots--;
524
525         pthread_cond_broadcast(&q->space_avail_cond);
526         pthread_mutex_unlock(&q->lock);
527         return obj;
528 }
529
530 struct compressor_thread_params {
531         struct shared_queue *res_to_compress_queue;
532         struct shared_queue *compressed_res_queue;
533         compress_func_t compress;
534 };
535
536 #define MAX_CHUNKS_PER_MSG 2
537
538 struct message {
539         struct wim_lookup_table_entry *lte;
540         u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
541         u8 *out_compressed_chunks[MAX_CHUNKS_PER_MSG];
542         u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
543         unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
544         unsigned compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
545         unsigned num_chunks;
546         struct list_head list;
547         bool complete;
548         u64 begin_chunk;
549 };
550
551 static void
552 compress_chunks(struct message *msg, compress_func_t compress)
553 {
554         for (unsigned i = 0; i < msg->num_chunks; i++) {
555                 DEBUG2("compress chunk %u of %u", i, msg->num_chunks);
556                 unsigned len = compress(msg->uncompressed_chunks[i],
557                                         msg->uncompressed_chunk_sizes[i],
558                                         msg->compressed_chunks[i]);
559                 if (len) {
560                         /* To be written compressed */
561                         msg->out_compressed_chunks[i] = msg->compressed_chunks[i];
562                         msg->compressed_chunk_sizes[i] = len;
563                 } else {
564                         /* To be written uncompressed */
565                         msg->out_compressed_chunks[i] = msg->uncompressed_chunks[i];
566                         msg->compressed_chunk_sizes[i] = msg->uncompressed_chunk_sizes[i];
567
568                 }
569         }
570 }
571
572 /* Compressor thread routine.  This is a lot simpler than the main thread
573  * routine: just repeatedly get a group of chunks from the
574  * res_to_compress_queue, compress them, and put them in the
575  * compressed_res_queue.  A NULL pointer indicates that the thread should stop.
576  * */
577 static void *
578 compressor_thread_proc(void *arg)
579 {
580         struct compressor_thread_params *params = arg;
581         struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
582         struct shared_queue *compressed_res_queue = params->compressed_res_queue;
583         compress_func_t compress = params->compress;
584         struct message *msg;
585
586         DEBUG("Compressor thread ready");
587         while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
588                 compress_chunks(msg, compress);
589                 shared_queue_put(compressed_res_queue, msg);
590         }
591         DEBUG("Compressor thread terminating");
592         return NULL;
593 }
594 #endif /* ENABLE_MULTITHREADED_COMPRESSION */
595
596 static void
597 do_write_streams_progress(union wimlib_progress_info *progress,
598                           wimlib_progress_func_t progress_func,
599                           uint64_t size_added)
600 {
601         progress->write_streams.completed_bytes += size_added;
602         progress->write_streams.completed_streams++;
603         if (progress_func &&
604             progress->write_streams.completed_bytes >= progress->write_streams._private)
605         {
606                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
607                               progress);
608                 if (progress->write_streams._private == progress->write_streams.total_bytes) {
609                         progress->write_streams._private = ~0;
610                 } else {
611                         progress->write_streams._private =
612                                 min(progress->write_streams.total_bytes,
613                                     progress->write_streams.completed_bytes +
614                                         progress->write_streams.total_bytes / 100);
615                 }
616         }
617 }
618
619 static int
620 do_write_stream_list(struct list_head *stream_list,
621                      struct wim_lookup_table *lookup_table,
622                      FILE *out_fp,
623                      int out_ctype,
624                      int write_resource_flags,
625                      wimlib_progress_func_t progress_func,
626                      union wimlib_progress_info *progress)
627 {
628         int ret = 0;
629         struct wim_lookup_table_entry *lte;
630
631         /* For each stream in @stream_list ... */
632         while (!list_empty(stream_list)) {
633                 lte = container_of(stream_list->next,
634                                    struct wim_lookup_table_entry,
635                                    write_streams_list);
636                 list_del(&lte->write_streams_list);
637                 if (lte->unhashed && !lte->unique_size) {
638                         /* Unhashed stream that shares a size with some other
639                          * stream in the WIM we are writing.  The stream must be
640                          * checksummed to know if we need to write it or not. */
641                         struct wim_lookup_table_entry *tmp;
642                         u32 orig_refcnt = lte->out_refcnt;
643
644                         ret = hash_unhashed_stream(lte,
645                                                    lookup_table,
646                                                    &tmp);
647                         if (ret)
648                                 break;
649                         if (tmp != lte) {
650                                 lte = tmp;
651                                 /* We found a duplicate stream. */
652                                 if (orig_refcnt != tmp->out_refcnt) {
653                                         /* We have already written, or are going
654                                          * to write, the duplicate stream.  So
655                                          * just skip to the next stream. */
656                                         DEBUG("Discarding duplicate stream of length %"PRIu64,
657                                               wim_resource_size(lte));
658                                         goto skip_to_progress;
659                                 }
660                         }
661                 }
662
663                 /* Here, @lte is either a hashed stream or an unhashed stream
664                  * with a unique size.  In either case we know that the stream
665                  * has to be written.  In either case the SHA1 message digest
666                  * will be calculated over the stream while writing it; however,
667                  * in the former case this is done merely to check the data,
668                  * while in the latter case this is done because we do not have
669                  * the SHA1 message digest yet.  */
670                 wimlib_assert(lte->out_refcnt != 0);
671                 ret = write_wim_resource(lte,
672                                          out_fp,
673                                          out_ctype,
674                                          &lte->output_resource_entry,
675                                          write_resource_flags);
676                 if (ret)
677                         break;
678                 if (lte->unhashed) {
679                         list_del(&lte->unhashed_list);
680                         lookup_table_insert(lookup_table, lte);
681                         lte->unhashed = 0;
682                 }
683         skip_to_progress:
684                 do_write_streams_progress(progress,
685                                           progress_func,
686                                           wim_resource_size(lte));
687         }
688         return ret;
689 }
690
691 static int
692 write_stream_list_serial(struct list_head *stream_list,
693                          struct wim_lookup_table *lookup_table,
694                          FILE *out_fp,
695                          int out_ctype,
696                          int write_flags,
697                          wimlib_progress_func_t progress_func,
698                          union wimlib_progress_info *progress)
699 {
700         int write_resource_flags = 0;
701         if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
702                 write_resource_flags |= WIMLIB_RESOURCE_FLAG_RECOMPRESS;
703
704         progress->write_streams.num_threads = 1;
705         if (progress_func)
706                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
707         return do_write_stream_list(stream_list,
708                                     lookup_table,
709                                     out_fp,
710                                     out_ctype,
711                                     write_resource_flags,
712                                     progress_func,
713                                     progress);
714 }
715
716 #ifdef ENABLE_MULTITHREADED_COMPRESSION
717 static int
718 write_wim_chunks(struct message *msg, FILE *out_fp,
719                  struct chunk_table *chunk_tab)
720 {
721         for (unsigned i = 0; i < msg->num_chunks; i++) {
722                 unsigned chunk_csize = msg->compressed_chunk_sizes[i];
723
724                 DEBUG2("Write wim chunk %u of %u (csize = %u)",
725                       i, msg->num_chunks, chunk_csize);
726
727                 if (fwrite(msg->out_compressed_chunks[i], 1, chunk_csize, out_fp)
728                     != chunk_csize)
729                 {
730                         ERROR_WITH_ERRNO("Failed to write WIM chunk");
731                         return WIMLIB_ERR_WRITE;
732                 }
733
734                 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
735                 chunk_tab->cur_offset += chunk_csize;
736         }
737         return 0;
738 }
739
740 struct main_writer_thread_ctx {
741         struct list_head *stream_list;
742         struct wim_lookup_table *lookup_table;
743         FILE *out_fp;
744         int out_ctype;
745         struct shared_queue *res_to_compress_queue;
746         struct shared_queue *compressed_res_queue;
747         size_t num_messages;
748         int write_flags;
749         wimlib_progress_func_t progress_func;
750         union wimlib_progress_info *progress;
751
752         struct list_head available_msgs;
753         struct list_head outstanding_streams;
754         struct list_head serial_streams;
755         u64 next_chunk;
756         u64 next_num_chunks;
757         struct message *msgs;
758         struct message *next_msg;
759         size_t next_chunk_in_msg;
760         struct wim_lookup_table_entry *cur_lte;
761         struct chunk_table *cur_chunk_tab;
762         struct wim_lookup_table_entry *next_lte;
763         SHA_CTX sha_ctx;
764         u8 next_hash[20];
765 };
766
767 static int
768 init_message(struct message *msg)
769 {
770         for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
771                 msg->compressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
772                 msg->uncompressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
773                 if (msg->compressed_chunks[i] == NULL ||
774                     msg->uncompressed_chunks[i] == NULL)
775                         return WIMLIB_ERR_NOMEM;
776         }
777         return 0;
778 }
779
780 static void
781 destroy_message(struct message *msg)
782 {
783         for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
784                 FREE(msg->compressed_chunks[i]);
785                 FREE(msg->uncompressed_chunks[i]);
786         }
787 }
788
789 static void
790 free_messages(struct message *msgs, size_t num_messages)
791 {
792         if (msgs) {
793                 for (size_t i = 0; i < num_messages; i++)
794                         destroy_message(&msgs[i]);
795                 FREE(msgs);
796         }
797 }
798
799 static struct message *
800 allocate_messages(size_t num_messages)
801 {
802         struct message *msgs;
803
804         msgs = CALLOC(num_messages, sizeof(struct message));
805         if (!msgs)
806                 return NULL;
807         for (size_t i = 0; i < num_messages; i++) {
808                 if (init_message(&msgs[i])) {
809                         free_messages(msgs, num_messages);
810                         return NULL;
811                 }
812         }
813         return msgs;
814 }
815
816 static void
817 main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
818 {
819         free_messages(ctx->msgs, ctx->num_messages);
820         FREE(ctx->cur_chunk_tab);
821 }
822
823
824 static int
825 main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
826 {
827         /* Pre-allocate all the buffers that will be needed to do the chunk
828          * compression. */
829         ctx->msgs = allocate_messages(ctx->num_messages);
830         if (!ctx->msgs)
831                 return WIMLIB_ERR_NOMEM;
832
833         /* Initially, all the messages are available to use. */
834         INIT_LIST_HEAD(&ctx->available_msgs);
835         for (size_t i = 0; i < ctx->num_messages; i++)
836                 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
837
838         /* outstanding_streams is the list of streams that currently have had
839          * chunks sent off for compression.
840          *
841          * The first stream in outstanding_streams is the stream that is
842          * currently being written (cur_lte).
843          *
844          * The last stream in outstanding_streams is the stream that is
845          * currently being read and chunks fed to the compressor threads.  */
846         INIT_LIST_HEAD(&ctx->outstanding_streams);
847
848         /* Resources that don't need any chunks compressed are added to this
849          * list and written directly by the main thread. */
850         INIT_LIST_HEAD(&ctx->serial_streams);
851
852         ctx->cur_lte = NULL;
853         return 0;
854 }
855
856 static int
857 receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
858 {
859         struct message *msg;
860         struct wim_lookup_table_entry *cur_lte;
861         int ret;
862
863         wimlib_assert(!list_empty(&ctx->outstanding_streams));
864
865         /* Get the next message from the queue and process it.
866          * The message will contain 1 or more data chunks that have been
867          * compressed. */
868         msg = shared_queue_get(ctx->compressed_res_queue);
869         msg->complete = true;
870         cur_lte = ctx->cur_lte;
871
872         /* Is this the next chunk in the current resource?  If it's not
873          * (i.e., an earlier chunk in a same or different resource
874          * hasn't been compressed yet), do nothing, and keep this
875          * message around until all earlier chunks are received.
876          *
877          * Otherwise, write all the chunks we can. */
878         while (cur_lte != NULL &&
879                !list_empty(&cur_lte->msg_list) &&
880                (msg = container_of(cur_lte->msg_list.next,
881                                    struct message,
882                                    list))->complete)
883         {
884                 if (msg->begin_chunk == 0) {
885
886                         /* This is the first set of chunks.  Leave space
887                          * for the chunk table in the output file. */
888                         off_t cur_offset = ftello(ctx->out_fp);
889                         if (cur_offset == -1) {
890                                 ret = WIMLIB_ERR_WRITE;
891                                 goto out;
892                         }
893                         ret = begin_wim_resource_chunk_tab(cur_lte,
894                                                            ctx->out_fp,
895                                                            cur_offset,
896                                                            &ctx->cur_chunk_tab);
897                         if (ret)
898                                 goto out;
899                 }
900
901                 /* Write the compressed chunks from the message. */
902                 ret = write_wim_chunks(msg, ctx->out_fp, ctx->cur_chunk_tab);
903                 if (ret)
904                         goto out;
905
906                 list_del(&msg->list);
907
908                 /* This message is available to use for different chunks
909                  * now. */
910                 list_add(&msg->list, &ctx->available_msgs);
911
912                 /* Was this the last chunk of the stream?  If so, finish
913                  * it. */
914                 if (list_empty(&cur_lte->msg_list) &&
915                     msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
916                 {
917                         DEBUG2("Finish wim chunk tab");
918                         u64 res_csize;
919                         ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
920                                                             ctx->out_fp,
921                                                             &res_csize);
922                         if (ret)
923                                 goto out;
924
925 #if 0
926                         if (res_csize >= wim_resource_size(cur_lte)) {
927                                 /* Oops!  We compressed the resource to
928                                  * larger than the original size.  Write
929                                  * the resource uncompressed instead. */
930                                 ret = write_uncompressed_resource_and_truncate(
931                                                  cur_lte,
932                                                  ctx->out_fp,
933                                                  ctx->cur_chunk_tab->file_offset,
934                                                  &cur_lte->output_resource_entry);
935                                 if (ret)
936                                         goto out;
937                         } else
938 #endif
939                         {
940                                 cur_lte->output_resource_entry.size =
941                                         res_csize;
942
943                                 cur_lte->output_resource_entry.original_size =
944                                         cur_lte->resource_entry.original_size;
945
946                                 cur_lte->output_resource_entry.offset =
947                                         ctx->cur_chunk_tab->file_offset;
948
949                                 cur_lte->output_resource_entry.flags =
950                                         cur_lte->resource_entry.flags |
951                                                 WIM_RESHDR_FLAG_COMPRESSED;
952                         }
953                         do_write_streams_progress(ctx->progress, ctx->progress_func,
954                                                   wim_resource_size(cur_lte));
955                         FREE(ctx->cur_chunk_tab);
956                         ctx->cur_chunk_tab = NULL;
957
958                         struct list_head *next = cur_lte->write_streams_list.next;
959                         list_del(&cur_lte->write_streams_list);
960
961                         if (next == &ctx->outstanding_streams)
962                                 cur_lte = NULL;
963                         else
964                                 cur_lte = container_of(cur_lte->write_streams_list.next,
965                                                        struct wim_lookup_table_entry,
966                                                        write_streams_list);
967
968                         /* Since we just finished writing a stream, write any
969                          * streams that have been added to the serial_streams
970                          * list for direct writing by the main thread (e.g.
971                          * resources that don't need to be compressed because
972                          * the desired compression type is the same as the
973                          * previous compression type). */
974                         ret = do_write_stream_list(&ctx->serial_streams,
975                                                    ctx->lookup_table,
976                                                    ctx->out_fp,
977                                                    ctx->out_ctype,
978                                                    ctx->progress_func,
979                                                    ctx->progress,
980                                                    0);
981                         if (ret)
982                                 goto out;
983                 }
984         }
985 out:
986         ctx->cur_lte = cur_lte;
987         return ret;
988 }
989
990 static int
991 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
992 {
993         struct main_writer_thread_ctx *ctx = _ctx;
994         int ret;
995         struct message *next_msg;
996
997         next_msg = ctx->next_msg;
998
999         sha1_update(&ctx->sha_ctx, chunk, chunk_size);
1000
1001         if (!next_msg) {
1002                 if (list_empty(&ctx->available_msgs)) {
1003                         ret = receive_compressed_chunks(ctx);
1004                         if (ret)
1005                                 return ret;
1006                 }
1007
1008                 wimlib_assert(!list_empty(&ctx->available_msgs));
1009
1010                 next_msg = container_of(ctx->available_msgs.next,
1011                                        struct message,
1012                                        list);
1013                 list_del(&next_msg->list);
1014                 next_msg->complete = false;
1015                 next_msg->begin_chunk = ctx->next_chunk;
1016                 next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
1017                                            ctx->next_num_chunks - ctx->next_chunk);
1018                 ctx->next_chunk_in_msg = 0;
1019         }
1020
1021         wimlib_assert(next_msg != NULL);
1022         wimlib_assert(ctx->next_chunk_in_msg < next_msg->num_chunks);
1023
1024         next_msg->uncompressed_chunk_sizes[ctx->next_chunk_in_msg] = chunk_size;
1025         memcpy(next_msg->uncompressed_chunks[ctx->next_chunk_in_msg],
1026                chunk, chunk_size);
1027
1028         if (++ctx->next_chunk_in_msg == next_msg->num_chunks) {
1029                 shared_queue_put(ctx->res_to_compress_queue,
1030                                  next_msg);
1031                 ctx->next_msg = NULL;
1032         }
1033         return 0;
1034 }
1035
1036 static int
1037 submit_stream_for_compression(struct wim_lookup_table_entry *lte,
1038                               struct main_writer_thread_ctx *ctx)
1039 {
1040         int ret;
1041
1042         sha1_init(&ctx->sha_ctx);
1043         ctx->next_num_chunks = wim_resource_chunks(lte);
1044         ret = read_resource_prefix(lte, wim_resource_size(lte),
1045                                    main_writer_thread_cb, ctx, 0);
1046         if (ret)
1047                 return ret;
1048         ret = finalize_and_check_sha1(&ctx->sha_ctx, lte);
1049         if (ret)
1050                 return ret;
1051 }
1052
1053 /*
1054  * This function is executed by the main thread when the resources are being
1055  * compressed in parallel.  The main thread is in change of all reading of the
1056  * uncompressed data and writing of the compressed data.  The compressor threads
1057  * *only* do compression from/to in-memory buffers.
1058  *
1059  * Each unit of work given to a compressor thread is up to MAX_CHUNKS_PER_MSG
1060  * chunks of compressed data to compress, represented in a `struct message'.
1061  * Each message is passed from the main thread to a worker thread through the
1062  * res_to_compress_queue, and it is passed back through the
1063  * compressed_res_queue.
1064  */
1065 static int
1066 main_writer_thread_proc(struct main_writer_thread_ctx *ctx)
1067 {
1068         int ret;
1069         struct list_head *stream_list;
1070         struct wim_lookup_table_entry *lte;
1071
1072         ret = main_writer_thread_init_ctx(ctx);
1073         if (ret)
1074                 goto out_destroy_ctx;
1075
1076         stream_list = ctx->stream_list;
1077         while (!list_empty(stream_list)) {
1078                 lte = container_of(stream_list->next,
1079                                    struct wim_lookup_table_entry,
1080                                    write_streams_list);
1081                 list_del(&lte->write_streams_list);
1082                 if (lte->unhashed && !lte->unique_size) {
1083                         struct wim_lookup_table_entry *tmp;
1084                         u32 orig_refcnt = lte->out_refcnt;
1085
1086                         ret = hash_unhashed_stream(lte, ctx->lookup_table, &tmp);
1087                         if (ret)
1088                                 goto out_destroy_ctx;
1089                         if (tmp != lte) {
1090                                 lte = tmp;
1091                                 if (orig_refcnt != tmp->out_refcnt) {
1092                                         DEBUG("Discarding duplicate stream of length %"PRIu64,
1093                                               wim_resource_size(lte));
1094                                         goto skip_to_progress;
1095                                 }
1096                         }
1097                 }
1098
1099                 if (wim_resource_size(lte) < 1000 ||
1100                     ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
1101                     (lte->resource_location == RESOURCE_IN_WIM &&
1102                      wimlib_get_compression_type(lte->wim) == ctx->out_ctype))
1103                 {
1104                         list_add(&lte->write_streams_list,
1105                                  &ctx->serial_streams);
1106                 } else {
1107                         ret = submit_stream_for_compression(lte, ctx);
1108                         if (ret)
1109                                 goto out_destroy_ctx;
1110                         if (lte->unhashed) {
1111                                 list_del(&lte->unhashed_list);
1112                                 lookup_table_insert(ctx->lookup_table, lte);
1113                                 lte->unhashed = 0;
1114                         }
1115                 }
1116         skip_to_progress:
1117                 do_write_streams_progress(ctx->progress,
1118                                           ctx->progress_func,
1119                                           wim_resource_size(lte));
1120         }
1121
1122         while (!list_empty(&ctx->outstanding_streams)) {
1123                 ret = receive_compressed_chunks(ctx);
1124                 if (ret)
1125                         goto out_destroy_ctx;
1126         }
1127         ret = 0;
1128 out_destroy_ctx:
1129         main_writer_thread_destroy_ctx(ctx);
1130         return ret;
1131 }
1132
1133 static long
1134 get_default_num_threads()
1135 {
1136 #ifdef __WIN32__
1137         return win32_get_number_of_processors();
1138 #else
1139         return sysconf(_SC_NPROCESSORS_ONLN);
1140 #endif
1141 }
1142
1143 static int
1144 write_stream_list_parallel(struct list_head *stream_list,
1145                            struct wim_lookup_table *lookup_table,
1146                            FILE *out_fp,
1147                            int out_ctype,
1148                            int write_flags,
1149                            unsigned num_threads,
1150                            wimlib_progress_func_t progress_func,
1151                            union wimlib_progress_info *progress)
1152 {
1153         int ret;
1154         struct shared_queue res_to_compress_queue;
1155         struct shared_queue compressed_res_queue;
1156         pthread_t *compressor_threads = NULL;
1157
1158         if (num_threads == 0) {
1159                 long nthreads = get_default_num_threads();
1160                 if (nthreads < 1 || nthreads > UINT_MAX) {
1161                         WARNING("Could not determine number of processors! Assuming 1");
1162                         goto out_serial;
1163                 } else {
1164                         num_threads = nthreads;
1165                 }
1166         }
1167
1168         progress->write_streams.num_threads = num_threads;
1169
1170         static const double MESSAGES_PER_THREAD = 2.0;
1171         size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1172
1173         DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1174
1175         ret = shared_queue_init(&res_to_compress_queue, queue_size);
1176         if (ret != 0)
1177                 goto out_serial;
1178
1179         ret = shared_queue_init(&compressed_res_queue, queue_size);
1180         if (ret != 0)
1181                 goto out_destroy_res_to_compress_queue;
1182
1183         struct compressor_thread_params params;
1184         params.res_to_compress_queue = &res_to_compress_queue;
1185         params.compressed_res_queue = &compressed_res_queue;
1186         params.compress = get_compress_func(out_ctype);
1187
1188         compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1189         if (!compressor_threads) {
1190                 ret = WIMLIB_ERR_NOMEM;
1191                 goto out_destroy_compressed_res_queue;
1192         }
1193
1194         for (unsigned i = 0; i < num_threads; i++) {
1195                 DEBUG("pthread_create thread %u", i);
1196                 ret = pthread_create(&compressor_threads[i], NULL,
1197                                      compressor_thread_proc, &params);
1198                 if (ret != 0) {
1199                         ret = -1;
1200                         ERROR_WITH_ERRNO("Failed to create compressor "
1201                                          "thread %u", i);
1202                         num_threads = i;
1203                         goto out_join;
1204                 }
1205         }
1206
1207         if (progress_func)
1208                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
1209
1210         struct main_writer_thread_ctx ctx;
1211         memset(&ctx, 0, sizeof(ctx));
1212         ctx.stream_list           = stream_list;
1213         ctx.lookup_table          = lookup_table;
1214         ctx.out_fp                = out_fp;
1215         ctx.out_ctype             = out_ctype;
1216         ctx.res_to_compress_queue = &res_to_compress_queue;
1217         ctx.compressed_res_queue  = &compressed_res_queue;
1218         ctx.num_messages          = queue_size;
1219         ctx.write_flags           = write_flags;
1220         ctx.progress_func         = progress_func;
1221         ctx.progress              = progress;
1222         ret = main_writer_thread_proc(&ctx);
1223 out_join:
1224         for (unsigned i = 0; i < num_threads; i++)
1225                 shared_queue_put(&res_to_compress_queue, NULL);
1226
1227         for (unsigned i = 0; i < num_threads; i++) {
1228                 if (pthread_join(compressor_threads[i], NULL)) {
1229                         WARNING_WITH_ERRNO("Failed to join compressor "
1230                                            "thread %u", i);
1231                 }
1232         }
1233         FREE(compressor_threads);
1234 out_destroy_compressed_res_queue:
1235         shared_queue_destroy(&compressed_res_queue);
1236 out_destroy_res_to_compress_queue:
1237         shared_queue_destroy(&res_to_compress_queue);
1238         if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
1239                 return ret;
1240 out_serial:
1241         WARNING("Falling back to single-threaded compression");
1242         return write_stream_list_serial(stream_list,
1243                                         lookup_table,
1244                                         out_fp,
1245                                         out_ctype,
1246                                         write_flags,
1247                                         progress_func,
1248                                         progress);
1249
1250 }
1251 #endif
1252
1253 /*
1254  * Write a list of streams to a WIM (@out_fp) using the compression type
1255  * @out_ctype and up to @num_threads compressor threads.
1256  */
1257 static int
1258 write_stream_list(struct list_head *stream_list,
1259                   struct wim_lookup_table *lookup_table,
1260                   FILE *out_fp, int out_ctype, int write_flags,
1261                   unsigned num_threads, wimlib_progress_func_t progress_func)
1262 {
1263         struct wim_lookup_table_entry *lte;
1264         size_t num_streams = 0;
1265         u64 total_bytes = 0;
1266         u64 total_compression_bytes = 0;
1267         union wimlib_progress_info progress;
1268         int ret;
1269
1270         if (list_empty(stream_list))
1271                 return 0;
1272
1273         /* Calculate the total size of the streams to be written.  Note: this
1274          * will be the uncompressed size, as we may not know the compressed size
1275          * yet, and also this will assume that every unhashed stream will be
1276          * written (which will not necessarily be the case). */
1277         list_for_each_entry(lte, stream_list, write_streams_list) {
1278                 num_streams++;
1279                 total_bytes += wim_resource_size(lte);
1280                 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
1281                        && (wim_resource_compression_type(lte) != out_ctype ||
1282                            (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)))
1283                 {
1284                         total_compression_bytes += wim_resource_size(lte);
1285                 }
1286         }
1287         progress.write_streams.total_bytes       = total_bytes;
1288         progress.write_streams.total_streams     = num_streams;
1289         progress.write_streams.completed_bytes   = 0;
1290         progress.write_streams.completed_streams = 0;
1291         progress.write_streams.num_threads       = num_threads;
1292         progress.write_streams.compression_type  = out_ctype;
1293         progress.write_streams._private          = 0;
1294
1295 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1296         if (total_compression_bytes >= 1000000 && num_threads != 1)
1297                 ret = write_stream_list_parallel(stream_list,
1298                                                  lookup_table,
1299                                                  out_fp,
1300                                                  out_ctype,
1301                                                  write_flags,
1302                                                  num_threads,
1303                                                  progress_func,
1304                                                  &progress);
1305         else
1306 #endif
1307                 ret = write_stream_list_serial(stream_list,
1308                                                lookup_table,
1309                                                out_fp,
1310                                                out_ctype,
1311                                                write_flags,
1312                                                progress_func,
1313                                                &progress);
1314         return ret;
1315 }
1316
1317 struct stream_size_table {
1318         struct hlist_head *array;
1319         size_t num_entries;
1320         size_t capacity;
1321 };
1322
1323 static int
1324 init_stream_size_table(struct stream_size_table *tab, size_t capacity)
1325 {
1326         tab->array = CALLOC(capacity, sizeof(tab->array[0]));
1327         if (!tab->array)
1328                 return WIMLIB_ERR_NOMEM;
1329         tab->num_entries = 0;
1330         tab->capacity = capacity;
1331         return 0;
1332 }
1333
1334 static void
1335 destroy_stream_size_table(struct stream_size_table *tab)
1336 {
1337         FREE(tab->array);
1338 }
1339
1340 static int
1341 stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
1342 {
1343         struct stream_size_table *tab = _tab;
1344         size_t pos;
1345         struct wim_lookup_table_entry *hashed_lte;
1346         struct hlist_node *tmp;
1347
1348         pos = hash_u64(wim_resource_size(lte)) % tab->capacity;
1349         lte->unique_size = 1;
1350         hlist_for_each_entry(hashed_lte, tmp, &tab->array[pos], hash_list_2) {
1351                 if (wim_resource_size(hashed_lte) == wim_resource_size(lte)) {
1352                         lte->unique_size = 0;
1353                         hashed_lte->unique_size = 0;
1354                         break;
1355                 }
1356         }
1357
1358         hlist_add_head(&lte->hash_list_2, &tab->array[pos]);
1359         tab->num_entries++;
1360         return 0;
1361 }
1362
1363
1364 struct lte_overwrite_prepare_args {
1365         WIMStruct *wim;
1366         off_t end_offset;
1367         struct list_head stream_list;
1368         struct stream_size_table stream_size_tab;
1369 };
1370
1371 static int
1372 lte_overwrite_prepare(struct wim_lookup_table_entry *lte, void *arg)
1373 {
1374         struct lte_overwrite_prepare_args *args = arg;
1375
1376         if (lte->resource_location == RESOURCE_IN_WIM &&
1377             lte->wim == args->wim)
1378         {
1379                 /* We can't do an in place overwrite on the WIM if there are
1380                  * streams after the XML data. */
1381                 if (lte->resource_entry.offset +
1382                     lte->resource_entry.size > args->end_offset)
1383                 {
1384                 #ifdef ENABLE_ERROR_MESSAGES
1385                         ERROR("The following resource is after the XML data:");
1386                         print_lookup_table_entry(lte, stderr);
1387                 #endif
1388                         return WIMLIB_ERR_RESOURCE_ORDER;
1389                 }
1390         } else {
1391                 wimlib_assert(!(lte->resource_entry.flags & WIM_RESHDR_FLAG_METADATA));
1392                 list_add_tail(&lte->write_streams_list, &args->stream_list);
1393         }
1394         lte->out_refcnt = lte->refcnt;
1395         stream_size_table_insert(lte, &args->stream_size_tab);
1396         return 0;
1397 }
1398
1399 static int
1400 lte_set_output_res_entry(struct wim_lookup_table_entry *lte, void *_wim)
1401 {
1402         if (lte->resource_location == RESOURCE_IN_WIM && lte->wim == _wim) {
1403                 copy_resource_entry(&lte->output_resource_entry,
1404                                     &lte->resource_entry);
1405         }
1406         return 0;
1407 }
1408
1409 /* Given a WIM that we are going to overwrite in place with zero or more
1410  * additional streams added, construct a list the list of new unique streams
1411  * ('struct wim_lookup_table_entry's) that must be written, plus any unhashed
1412  * streams that need to be added but may be identical to other hashed or
1413  * unhashed streams.  These unhashed streams are checksummed while the streams
1414  * are being written.  To aid this process, the member @unique_size is set to 1
1415  * on streams that have a unique size and therefore must be written.
1416  *
1417  * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1418  * indicate the number of times the stream is referenced in only the streams
1419  * that are being written; this may still be adjusted later when unhashed
1420  * streams are being resolved.
1421  */
1422 static int
1423 prepare_streams_for_overwrite(WIMStruct *wim, off_t end_offset,
1424                               struct list_head *stream_list)
1425 {
1426         int ret;
1427         struct lte_overwrite_prepare_args args;
1428
1429         args.wim = wim;
1430         args.end_offset = end_offset;
1431         ret = init_stream_size_table(&args.stream_size_tab,
1432                                      wim->lookup_table->capacity);
1433         if (ret)
1434                 return ret;
1435
1436         INIT_LIST_HEAD(&args.stream_list);
1437         for (int i = 0; i < wim->hdr.image_count; i++) {
1438                 struct wim_image_metadata *imd;
1439                 struct wim_lookup_table_entry *lte;
1440
1441                 imd = wim->image_metadata[i];
1442                 image_for_each_unhashed_stream(lte, imd) {
1443                         ret = lte_overwrite_prepare(lte, &args);
1444                         if (ret)
1445                                 goto out_destroy_stream_size_table;
1446                 }
1447         }
1448         ret = for_lookup_table_entry(wim->lookup_table,
1449                                      lte_overwrite_prepare, &args);
1450         if (ret)
1451                 goto out_destroy_stream_size_table;
1452
1453         for (int i = 0; i < wim->hdr.image_count; i++)
1454                 lte_set_output_res_entry(wim->image_metadata[i]->metadata_lte,
1455                                          wim);
1456         for_lookup_table_entry(wim->lookup_table, lte_set_output_res_entry, wim);
1457         list_transfer(&args.stream_list, stream_list);
1458 out_destroy_stream_size_table:
1459         destroy_stream_size_table(&args.stream_size_tab);
1460         return ret;
1461 }
1462
1463
1464 struct find_streams_ctx {
1465         struct list_head stream_list;
1466         struct stream_size_table stream_size_tab;
1467 };
1468
1469 static void
1470 inode_find_streams_to_write(struct wim_inode *inode,
1471                             struct wim_lookup_table *table,
1472                             struct list_head *stream_list,
1473                             struct stream_size_table *tab)
1474 {
1475         struct wim_lookup_table_entry *lte;
1476         for (unsigned i = 0; i <= inode->i_num_ads; i++) {
1477                 lte = inode_stream_lte(inode, i, table);
1478                 if (lte) {
1479                         if (lte->out_refcnt == 0) {
1480                                 if (lte->unhashed)
1481                                         stream_size_table_insert(lte, tab);
1482                                 list_add_tail(&lte->write_streams_list, stream_list);
1483                         }
1484                         lte->out_refcnt += inode->i_nlink;
1485                 }
1486         }
1487 }
1488
1489 static int
1490 image_find_streams_to_write(WIMStruct *w)
1491 {
1492         struct find_streams_ctx *ctx;
1493         struct wim_image_metadata *imd;
1494         struct wim_inode *inode;
1495         struct wim_lookup_table_entry *lte;
1496
1497         ctx = w->private;
1498         imd = wim_get_current_image_metadata(w);
1499
1500         image_for_each_unhashed_stream(lte, imd)
1501                 lte->out_refcnt = 0;
1502
1503         /* Go through this image's inodes to find any streams that have not been
1504          * found yet. */
1505         image_for_each_inode(inode, imd) {
1506                 inode_find_streams_to_write(inode, w->lookup_table,
1507                                             &ctx->stream_list,
1508                                             &ctx->stream_size_tab);
1509         }
1510         return 0;
1511 }
1512
1513 /* Given a WIM that from which one or all of the images is being written, build
1514  * the list of unique streams ('struct wim_lookup_table_entry's) that must be
1515  * written, plus any unhashed streams that need to be written but may be
1516  * identical to other hashed or unhashed streams being written.  These unhashed
1517  * streams are checksummed while the streams are being written.  To aid this
1518  * process, the member @unique_size is set to 1 on streams that have a unique
1519  * size and therefore must be written.
1520  *
1521  * The out_refcnt member of each 'struct wim_lookup_table_entry' is set to
1522  * indicate the number of times the stream is referenced in only the streams
1523  * that are being written; this may still be adjusted later when unhashed
1524  * streams are being resolved.
1525  */
1526 static int
1527 prepare_stream_list(WIMStruct *wim, int image, struct list_head *stream_list)
1528 {
1529         int ret;
1530         struct find_streams_ctx ctx;
1531
1532         for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
1533         ret = init_stream_size_table(&ctx.stream_size_tab,
1534                                      wim->lookup_table->capacity);
1535         if (ret)
1536                 return ret;
1537         for_lookup_table_entry(wim->lookup_table, stream_size_table_insert,
1538                                &ctx.stream_size_tab);
1539         INIT_LIST_HEAD(&ctx.stream_list);
1540         wim->private = &ctx;
1541         ret = for_image(wim, image, image_find_streams_to_write);
1542         destroy_stream_size_table(&ctx.stream_size_tab);
1543         if (ret == 0)
1544                 list_transfer(&ctx.stream_list, stream_list);
1545         return ret;
1546 }
1547
1548 /* Writes the streams for the specified @image in @wim to @wim->out_fp.
1549  */
1550 static int
1551 write_wim_streams(WIMStruct *wim, int image, int write_flags,
1552                   unsigned num_threads,
1553                   wimlib_progress_func_t progress_func)
1554 {
1555         int ret;
1556         struct list_head stream_list;
1557
1558         ret = prepare_stream_list(wim, image, &stream_list);
1559         if (ret)
1560                 return ret;
1561         return write_stream_list(&stream_list,
1562                                  wim->lookup_table,
1563                                  wim->out_fp,
1564                                  wimlib_get_compression_type(wim),
1565                                  write_flags,
1566                                  num_threads,
1567                                  progress_func);
1568 }
1569
1570 /*
1571  * Finish writing a WIM file: write the lookup table, xml data, and integrity
1572  * table (optional), then overwrite the WIM header.
1573  *
1574  * write_flags is a bitwise OR of the following:
1575  *
1576  *      (public)  WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
1577  *              Include an integrity table.
1578  *
1579  *      (public)  WIMLIB_WRITE_FLAG_SHOW_PROGRESS:
1580  *              Show progress information when (if) writing the integrity table.
1581  *
1582  *      (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
1583  *              Don't write the lookup table.
1584  *
1585  *      (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
1586  *              When (if) writing the integrity table, re-use entries from the
1587  *              existing integrity table, if possible.
1588  *
1589  *      (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
1590  *              After writing the XML data but before writing the integrity
1591  *              table, write a temporary WIM header and flush the stream so that
1592  *              the WIM is less likely to become corrupted upon abrupt program
1593  *              termination.
1594  *
1595  *      (private) WIMLIB_WRITE_FLAG_FSYNC:
1596  *              fsync() the output file before closing it.
1597  *
1598  */
1599 int
1600 finish_write(WIMStruct *w, int image, int write_flags,
1601              wimlib_progress_func_t progress_func)
1602 {
1603         int ret;
1604         struct wim_header hdr;
1605         FILE *out = w->out_fp;
1606
1607         /* @hdr will be the header for the new WIM.  First copy all the data
1608          * from the header in the WIMStruct; then set all the fields that may
1609          * have changed, including the resource entries, boot index, and image
1610          * count.  */
1611         memcpy(&hdr, &w->hdr, sizeof(struct wim_header));
1612
1613         /* Set image count and boot index correctly for single image writes */
1614         if (image != WIMLIB_ALL_IMAGES) {
1615                 hdr.image_count = 1;
1616                 if (hdr.boot_idx == image)
1617                         hdr.boot_idx = 1;
1618                 else
1619                         hdr.boot_idx = 0;
1620         }
1621
1622         /* In the WIM header, there is room for the resource entry for a
1623          * metadata resource labeled as the "boot metadata".  This entry should
1624          * be zeroed out if there is no bootable image (boot_idx 0).  Otherwise,
1625          * it should be a copy of the resource entry for the image that is
1626          * marked as bootable.  This is not well documented...  */
1627         if (hdr.boot_idx == 0) {
1628                 zero_resource_entry(&hdr.boot_metadata_res_entry);
1629         } else {
1630                 copy_resource_entry(&hdr.boot_metadata_res_entry,
1631                             &w->image_metadata[ hdr.boot_idx- 1
1632                                         ]->metadata_lte->output_resource_entry);
1633         }
1634
1635         if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
1636                 ret = write_lookup_table(w, image, &hdr.lookup_table_res_entry);
1637                 if (ret)
1638                         goto out_close_wim;
1639         }
1640
1641         ret = write_xml_data(w->wim_info, image, out,
1642                              (write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE) ?
1643                               wim_info_get_total_bytes(w->wim_info) : 0,
1644                              &hdr.xml_res_entry);
1645         if (ret)
1646                 goto out_close_wim;
1647
1648         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
1649                 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
1650                         struct wim_header checkpoint_hdr;
1651                         memcpy(&checkpoint_hdr, &hdr, sizeof(struct wim_header));
1652                         zero_resource_entry(&checkpoint_hdr.integrity);
1653                         if (fseeko(out, 0, SEEK_SET)) {
1654                                 ERROR_WITH_ERRNO("Failed to seek to beginning "
1655                                                  "of WIM being written");
1656                                 ret = WIMLIB_ERR_WRITE;
1657                                 goto out_close_wim;
1658                         }
1659                         ret = write_header(&checkpoint_hdr, out);
1660                         if (ret)
1661                                 goto out_close_wim;
1662
1663                         if (fflush(out) != 0) {
1664                                 ERROR_WITH_ERRNO("Can't write data to WIM");
1665                                 ret = WIMLIB_ERR_WRITE;
1666                                 goto out_close_wim;
1667                         }
1668
1669                         if (fseeko(out, 0, SEEK_END) != 0) {
1670                                 ERROR_WITH_ERRNO("Failed to seek to end "
1671                                                  "of WIM being written");
1672                                 ret = WIMLIB_ERR_WRITE;
1673                                 goto out_close_wim;
1674                         }
1675                 }
1676
1677                 off_t old_lookup_table_end;
1678                 off_t new_lookup_table_end;
1679                 if (write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE) {
1680                         old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
1681                                                w->hdr.lookup_table_res_entry.size;
1682                 } else {
1683                         old_lookup_table_end = 0;
1684                 }
1685                 new_lookup_table_end = hdr.lookup_table_res_entry.offset +
1686                                        hdr.lookup_table_res_entry.size;
1687
1688                 ret = write_integrity_table(out,
1689                                             &hdr.integrity,
1690                                             new_lookup_table_end,
1691                                             old_lookup_table_end,
1692                                             progress_func);
1693                 if (ret)
1694                         goto out_close_wim;
1695         } else {
1696                 zero_resource_entry(&hdr.integrity);
1697         }
1698
1699         if (fseeko(out, 0, SEEK_SET) != 0) {
1700                 ERROR_WITH_ERRNO("Failed to seek to beginning of WIM "
1701                                  "being written");
1702                 ret = WIMLIB_ERR_WRITE;
1703                 goto out_close_wim;
1704         }
1705
1706         ret = write_header(&hdr, out);
1707         if (ret)
1708                 goto out_close_wim;
1709
1710         if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
1711                 if (fflush(out) != 0
1712                     || fsync(fileno(out)) != 0)
1713                 {
1714                         ERROR_WITH_ERRNO("Error flushing data to WIM file");
1715                         ret = WIMLIB_ERR_WRITE;
1716                 }
1717         }
1718 out_close_wim:
1719         if (fclose(out) != 0) {
1720                 ERROR_WITH_ERRNO("Failed to close the output WIM file");
1721                 if (ret == 0)
1722                         ret = WIMLIB_ERR_WRITE;
1723         }
1724         w->out_fp = NULL;
1725         return ret;
1726 }
1727
1728 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
1729 int
1730 lock_wim(WIMStruct *w, FILE *fp)
1731 {
1732         int ret = 0;
1733         if (fp && !w->wim_locked) {
1734                 ret = flock(fileno(fp), LOCK_EX | LOCK_NB);
1735                 if (ret != 0) {
1736                         if (errno == EWOULDBLOCK) {
1737                                 ERROR("`%"TS"' is already being modified or has been "
1738                                       "mounted read-write\n"
1739                                       "        by another process!", w->filename);
1740                                 ret = WIMLIB_ERR_ALREADY_LOCKED;
1741                         } else {
1742                                 WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
1743                                                    w->filename);
1744                                 ret = 0;
1745                         }
1746                 } else {
1747                         w->wim_locked = 1;
1748                 }
1749         }
1750         return ret;
1751 }
1752 #endif
1753
1754 static int
1755 open_wim_writable(WIMStruct *w, const tchar *path,
1756                   bool trunc, bool also_readable)
1757 {
1758         const tchar *mode;
1759         if (trunc)
1760                 if (also_readable)
1761                         mode = T("w+b");
1762                 else
1763                         mode = T("wb");
1764         else
1765                 mode = T("r+b");
1766
1767         wimlib_assert(w->out_fp == NULL);
1768         w->out_fp = tfopen(path, mode);
1769         if (w->out_fp) {
1770                 return 0;
1771         } else {
1772                 ERROR_WITH_ERRNO("Failed to open `%"TS"' for writing", path);
1773                 return WIMLIB_ERR_OPEN;
1774         }
1775 }
1776
1777
1778 void
1779 close_wim_writable(WIMStruct *w)
1780 {
1781         if (w->out_fp) {
1782                 if (fclose(w->out_fp) != 0) {
1783                         WARNING_WITH_ERRNO("Failed to close output WIM");
1784                 }
1785                 w->out_fp = NULL;
1786         }
1787 }
1788
1789 /* Open file stream and write dummy header for WIM. */
1790 int
1791 begin_write(WIMStruct *w, const tchar *path, int write_flags)
1792 {
1793         int ret;
1794         ret = open_wim_writable(w, path, true,
1795                                 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
1796         if (ret)
1797                 return ret;
1798         /* Write dummy header. It will be overwritten later. */
1799         return write_header(&w->hdr, w->out_fp);
1800 }
1801
1802 /* Writes a stand-alone WIM to a file.  */
1803 WIMLIBAPI int
1804 wimlib_write(WIMStruct *w, const tchar *path,
1805              int image, int write_flags, unsigned num_threads,
1806              wimlib_progress_func_t progress_func)
1807 {
1808         int ret;
1809
1810         if (!path)
1811                 return WIMLIB_ERR_INVALID_PARAM;
1812
1813         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1814
1815         if (image != WIMLIB_ALL_IMAGES &&
1816              (image < 1 || image > w->hdr.image_count))
1817                 return WIMLIB_ERR_INVALID_IMAGE;
1818
1819         if (w->hdr.total_parts != 1) {
1820                 ERROR("Cannot call wimlib_write() on part of a split WIM");
1821                 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1822         }
1823
1824         ret = begin_write(w, path, write_flags);
1825         if (ret)
1826                 goto out_close_wim;
1827
1828         ret = write_wim_streams(w, image, write_flags, num_threads,
1829                                 progress_func);
1830         if (ret)
1831                 goto out_close_wim;
1832
1833         if (progress_func)
1834                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
1835
1836         ret = for_image(w, image, write_metadata_resource);
1837         if (ret)
1838                 goto out_close_wim;
1839
1840         if (progress_func)
1841                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
1842
1843         ret = finish_write(w, image, write_flags, progress_func);
1844         /* finish_write() closed the WIM for us */
1845         goto out;
1846 out_close_wim:
1847         close_wim_writable(w);
1848 out:
1849         DEBUG("wimlib_write(path=%"TS") = %d", path, ret);
1850         return ret;
1851 }
1852
1853 static bool
1854 any_images_modified(WIMStruct *w)
1855 {
1856         for (int i = 0; i < w->hdr.image_count; i++)
1857                 if (w->image_metadata[i]->modified)
1858                         return true;
1859         return false;
1860 }
1861
1862 /*
1863  * Overwrite a WIM, possibly appending streams to it.
1864  *
1865  * A WIM looks like (or is supposed to look like) the following:
1866  *
1867  *                   Header (212 bytes)
1868  *                   Streams and metadata resources (variable size)
1869  *                   Lookup table (variable size)
1870  *                   XML data (variable size)
1871  *                   Integrity table (optional) (variable size)
1872  *
1873  * If we are not adding any streams or metadata resources, the lookup table is
1874  * unchanged--- so we only need to overwrite the XML data, integrity table, and
1875  * header.  This operation is potentially unsafe if the program is abruptly
1876  * terminated while the XML data or integrity table are being overwritten, but
1877  * before the new header has been written.  To partially alleviate this problem,
1878  * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
1879  * finish_write() to cause a temporary WIM header to be written after the XML
1880  * data has been written.  This may prevent the WIM from becoming corrupted if
1881  * the program is terminated while the integrity table is being calculated (but
1882  * no guarantees, due to write re-ordering...).
1883  *
1884  * If we are adding new streams or images (metadata resources), the lookup table
1885  * needs to be changed, and those streams need to be written.  In this case, we
1886  * try to perform a safe update of the WIM file by writing the streams *after*
1887  * the end of the previous WIM, then writing the new lookup table, XML data, and
1888  * (optionally) integrity table following the new streams.  This will produce a
1889  * layout like the following:
1890  *
1891  *                   Header (212 bytes)
1892  *                   (OLD) Streams and metadata resources (variable size)
1893  *                   (OLD) Lookup table (variable size)
1894  *                   (OLD) XML data (variable size)
1895  *                   (OLD) Integrity table (optional) (variable size)
1896  *                   (NEW) Streams and metadata resources (variable size)
1897  *                   (NEW) Lookup table (variable size)
1898  *                   (NEW) XML data (variable size)
1899  *                   (NEW) Integrity table (optional) (variable size)
1900  *
1901  * At all points, the WIM is valid as nothing points to the new data yet.  Then,
1902  * the header is overwritten to point to the new lookup table, XML data, and
1903  * integrity table, to produce the following layout:
1904  *
1905  *                   Header (212 bytes)
1906  *                   Streams and metadata resources (variable size)
1907  *                   Nothing (variable size)
1908  *                   More Streams and metadata resources (variable size)
1909  *                   Lookup table (variable size)
1910  *                   XML data (variable size)
1911  *                   Integrity table (optional) (variable size)
1912  *
1913  * This method allows an image to be appended to a large WIM very quickly, and
1914  * is is crash-safe except in the case of write re-ordering, but the
1915  * disadvantage is that a small hole is left in the WIM where the old lookup
1916  * table, xml data, and integrity table were.  (These usually only take up a
1917  * small amount of space compared to the streams, however.)
1918  */
1919 static int
1920 overwrite_wim_inplace(WIMStruct *w, int write_flags,
1921                       unsigned num_threads,
1922                       wimlib_progress_func_t progress_func)
1923 {
1924         int ret;
1925         struct list_head stream_list;
1926         off_t old_wim_end;
1927
1928         DEBUG("Overwriting `%"TS"' in-place", w->filename);
1929
1930         /* Make sure that the integrity table (if present) is after the XML
1931          * data, and that there are no stream resources, metadata resources, or
1932          * lookup tables after the XML data.  Otherwise, these data would be
1933          * overwritten. */
1934         if (w->hdr.integrity.offset != 0 &&
1935             w->hdr.integrity.offset < w->hdr.xml_res_entry.offset) {
1936                 ERROR("Didn't expect the integrity table to be before the XML data");
1937                 return WIMLIB_ERR_RESOURCE_ORDER;
1938         }
1939
1940         if (w->hdr.lookup_table_res_entry.offset > w->hdr.xml_res_entry.offset) {
1941                 ERROR("Didn't expect the lookup table to be after the XML data");
1942                 return WIMLIB_ERR_RESOURCE_ORDER;
1943         }
1944
1945
1946         if (w->hdr.integrity.offset)
1947                 old_wim_end = w->hdr.integrity.offset + w->hdr.integrity.size;
1948         else
1949                 old_wim_end = w->hdr.xml_res_entry.offset + w->hdr.xml_res_entry.size;
1950
1951         if (!w->deletion_occurred && !any_images_modified(w)) {
1952                 /* If no images have been modified and no images have been
1953                  * deleted, a new lookup table does not need to be written. */
1954                 DEBUG("Skipping writing lookup table "
1955                       "(no images modified or deleted)");
1956                 old_wim_end = w->hdr.lookup_table_res_entry.offset +
1957                               w->hdr.lookup_table_res_entry.size;
1958                 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
1959                                WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
1960         }
1961         ret = prepare_streams_for_overwrite(w, old_wim_end, &stream_list);
1962         if (ret)
1963                 return ret;
1964
1965         ret = open_wim_writable(w, w->filename, false,
1966                                 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
1967         if (ret)
1968                 return ret;
1969
1970         ret = lock_wim(w, w->out_fp);
1971         if (ret) {
1972                 fclose(w->out_fp);
1973                 w->out_fp = NULL;
1974                 return ret;
1975         }
1976
1977         if (fseeko(w->out_fp, old_wim_end, SEEK_SET) != 0) {
1978                 ERROR_WITH_ERRNO("Can't seek to end of WIM");
1979                 fclose(w->out_fp);
1980                 w->out_fp = NULL;
1981                 w->wim_locked = 0;
1982                 return WIMLIB_ERR_WRITE;
1983         }
1984
1985         DEBUG("Writing newly added streams (offset = %"PRIu64")",
1986               old_wim_end);
1987         ret = write_stream_list(&stream_list,
1988                                 w->lookup_table,
1989                                 w->out_fp,
1990                                 wimlib_get_compression_type(w),
1991                                 write_flags,
1992                                 num_threads,
1993                                 progress_func);
1994         if (ret)
1995                 goto out_ftruncate;
1996
1997         for (int i = 0; i < w->hdr.image_count; i++) {
1998                 if (w->image_metadata[i]->modified) {
1999                         select_wim_image(w, i + 1);
2000                         ret = write_metadata_resource(w);
2001                         if (ret)
2002                                 goto out_ftruncate;
2003                 }
2004         }
2005         write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
2006         ret = finish_write(w, WIMLIB_ALL_IMAGES, write_flags,
2007                            progress_func);
2008 out_ftruncate:
2009         close_wim_writable(w);
2010         if (ret != 0 && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
2011                 WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
2012                         w->filename, old_wim_end);
2013                 /* Return value of truncate() is ignored because this is already
2014                  * an error path. */
2015                 (void)ttruncate(w->filename, old_wim_end);
2016         }
2017         w->wim_locked = 0;
2018         return ret;
2019 }
2020
2021 static int
2022 overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
2023                           unsigned num_threads,
2024                           wimlib_progress_func_t progress_func)
2025 {
2026         size_t wim_name_len;
2027         int ret;
2028
2029         DEBUG("Overwriting `%"TS"' via a temporary file", w->filename);
2030
2031         /* Write the WIM to a temporary file in the same directory as the
2032          * original WIM. */
2033         wim_name_len = tstrlen(w->filename);
2034         tchar tmpfile[wim_name_len + 10];
2035         tmemcpy(tmpfile, w->filename, wim_name_len);
2036         randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
2037         tmpfile[wim_name_len + 9] = T('\0');
2038
2039         ret = wimlib_write(w, tmpfile, WIMLIB_ALL_IMAGES,
2040                            write_flags | WIMLIB_WRITE_FLAG_FSYNC,
2041                            num_threads, progress_func);
2042         if (ret) {
2043                 ERROR("Failed to write the WIM file `%"TS"'", tmpfile);
2044                 goto out_unlink;
2045         }
2046
2047         DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, w->filename);
2048
2049 #ifdef __WIN32__
2050         /* Windows won't let you delete open files unless FILE_SHARE_DELETE was
2051          * specified to CreateFile().  The WIM was opened with fopen(), which
2052          * didn't provided this flag to CreateFile, so the handle must be closed
2053          * before executing the rename(). */
2054         if (w->fp != NULL) {
2055                 fclose(w->fp);
2056                 w->fp = NULL;
2057         }
2058 #endif
2059
2060         /* Rename the new file to the old file .*/
2061         if (trename(tmpfile, w->filename) != 0) {
2062                 ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
2063                                  tmpfile, w->filename);
2064                 ret = WIMLIB_ERR_RENAME;
2065                 goto out_unlink;
2066         }
2067
2068         if (progress_func) {
2069                 union wimlib_progress_info progress;
2070                 progress.rename.from = tmpfile;
2071                 progress.rename.to = w->filename;
2072                 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
2073         }
2074
2075         /* Close the original WIM file that was opened for reading. */
2076         if (w->fp != NULL) {
2077                 fclose(w->fp);
2078                 w->fp = NULL;
2079         }
2080
2081         /* Re-open the WIM read-only. */
2082         w->fp = tfopen(w->filename, T("rb"));
2083         if (w->fp == NULL) {
2084                 ret = WIMLIB_ERR_REOPEN;
2085                 WARNING_WITH_ERRNO("Failed to re-open `%"TS"' read-only",
2086                                    w->filename);
2087                 FREE(w->filename);
2088                 w->filename = NULL;
2089         }
2090         goto out;
2091 out_unlink:
2092         /* Remove temporary file. */
2093         if (tunlink(tmpfile) != 0)
2094                 WARNING_WITH_ERRNO("Failed to remove `%"TS"'", tmpfile);
2095 out:
2096         return ret;
2097 }
2098
2099 /*
2100  * Writes a WIM file to the original file that it was read from, overwriting it.
2101  */
2102 WIMLIBAPI int
2103 wimlib_overwrite(WIMStruct *w, int write_flags,
2104                  unsigned num_threads,
2105                  wimlib_progress_func_t progress_func)
2106 {
2107         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2108
2109         if (!w->filename)
2110                 return WIMLIB_ERR_NO_FILENAME;
2111
2112         if (w->hdr.total_parts != 1) {
2113                 ERROR("Cannot modify a split WIM");
2114                 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
2115         }
2116
2117         if ((!w->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
2118             && !(write_flags & WIMLIB_WRITE_FLAG_REBUILD))
2119         {
2120                 int ret;
2121                 ret = overwrite_wim_inplace(w, write_flags, num_threads,
2122                                             progress_func);
2123                 if (ret == WIMLIB_ERR_RESOURCE_ORDER)
2124                         WARNING("Falling back to re-building entire WIM");
2125                 else
2126                         return ret;
2127         }
2128         return overwrite_wim_via_tmpfile(w, write_flags, num_threads,
2129                                          progress_func);
2130 }