]> wimlib.net Git - wimlib/blob - src/write.c
a08bc8503a6c5a57f908ab7f535787205fe3fded
[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) 2010 Carl Thijssen
10  * Copyright (C) 2012 Eric Biggers
11  *
12  * This file is part of wimlib, a library for working with WIM files.
13  *
14  * wimlib is free software; you can redistribute it and/or modify it under the
15  * terms of the GNU General Public License as published by the Free
16  * Software Foundation; either version 3 of the License, or (at your option)
17  * any later version.
18  *
19  * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
20  * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
21  * A PARTICULAR PURPOSE. See the GNU General Public License for more
22  * details.
23  *
24  * You should have received a copy of the GNU General Public License
25  * along with wimlib; if not, see http://www.gnu.org/licenses/.
26  */
27
28 #include "wimlib_internal.h"
29 #include "io.h"
30 #include "dentry.h"
31 #include "lookup_table.h"
32 #include "xml.h"
33 #include "lzx.h"
34 #include "xpress.h"
35 #include <unistd.h>
36
37 #ifdef ENABLE_MULTITHREADED_COMPRESSION
38 #include <semaphore.h>
39 #include <pthread.h>
40 #include <errno.h>
41 #endif
42
43 #ifdef WITH_NTFS_3G
44 #include <time.h>
45 #include <ntfs-3g/attrib.h>
46 #include <ntfs-3g/inode.h>
47 #include <ntfs-3g/dir.h>
48 #endif
49
50
51 #ifdef HAVE_ALLOCA_H
52 #include <alloca.h>
53 #else
54 #include <stdlib.h>
55 #endif
56
57 static int do_fflush(FILE *fp)
58 {
59         int ret = fflush(fp);
60         if (ret != 0) {
61                 ERROR_WITH_ERRNO("Failed to flush data to output WIM file");
62                 return WIMLIB_ERR_WRITE;
63         }
64         return 0;
65 }
66
67 static int fflush_and_ftruncate(FILE *fp, off_t size)
68 {
69         int ret;
70
71         ret = do_fflush(fp);
72         if (ret != 0)
73                 return ret;
74         ret = ftruncate(fileno(fp), size);
75         if (ret != 0) {
76                 ERROR_WITH_ERRNO("Failed to truncate output WIM file to "
77                                  "%"PRIu64" bytes", size);
78                 return WIMLIB_ERR_WRITE;
79         }
80         return 0;
81 }
82
83 /* Chunk table that's located at the beginning of each compressed resource in
84  * the WIM.  (This is not the on-disk format; the on-disk format just has an
85  * array of offsets.) */
86 struct chunk_table {
87         off_t file_offset;
88         u64 num_chunks;
89         u64 original_resource_size;
90         u64 bytes_per_chunk_entry;
91         u64 table_disk_size;
92         u64 cur_offset;
93         u64 *cur_offset_p;
94         u64 offsets[0];
95 };
96
97 /*
98  * Allocates and initializes a chunk table, and reserves space for it in the
99  * output file.
100  */
101 static int
102 begin_wim_resource_chunk_tab(const struct lookup_table_entry *lte,
103                              FILE *out_fp,
104                              off_t file_offset,
105                              struct chunk_table **chunk_tab_ret)
106 {
107         u64 size = wim_resource_size(lte);
108         u64 num_chunks = (size + WIM_CHUNK_SIZE - 1) / WIM_CHUNK_SIZE;
109         size_t alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
110         struct chunk_table *chunk_tab = CALLOC(1, alloc_size);
111         int ret;
112
113         if (!chunk_tab) {
114                 ERROR("Failed to allocate chunk table for %"PRIu64" byte "
115                       "resource", size);
116                 ret = WIMLIB_ERR_NOMEM;
117                 goto out;
118         }
119         chunk_tab->file_offset = file_offset;
120         chunk_tab->num_chunks = num_chunks;
121         chunk_tab->original_resource_size = size;
122         chunk_tab->bytes_per_chunk_entry = (size >= (1ULL << 32)) ? 8 : 4;
123         chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
124                                      (num_chunks - 1);
125         chunk_tab->cur_offset = 0;
126         chunk_tab->cur_offset_p = chunk_tab->offsets;
127
128         if (fwrite(chunk_tab, 1, chunk_tab->table_disk_size, out_fp) !=
129                    chunk_tab->table_disk_size) {
130                 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
131                                  "file resource");
132                 ret = WIMLIB_ERR_WRITE;
133                 goto out;
134         }
135
136         ret = 0;
137 out:
138         *chunk_tab_ret = chunk_tab;
139         return ret;
140 }
141
142 /*
143  * Pointer to function to compresses a chunk of a WIM resource.
144  *
145  * @chunk:              Uncompressed data of the chunk.
146  * @chunk_size:         Size of the uncompressed chunk in bytes.
147  * @compressed_chunk:   Pointer to output buffer of size at least
148  *                              (@chunk_size - 1) bytes.
149  * @compressed_chunk_len_ret:   Pointer to an unsigned int into which the size
150  *                                      of the compressed chunk will be
151  *                                      returned.
152  *
153  * Returns zero if compressed succeeded, and nonzero if the chunk could not be
154  * compressed to any smaller than @chunk_size.  This function cannot fail for
155  * any other reasons.
156  */
157 typedef int (*compress_func_t)(const void *, unsigned, void *, unsigned *);
158
159 compress_func_t get_compress_func(int out_ctype)
160 {
161         if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX)
162                 return lzx_compress;
163         else
164                 return 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 tho chunk to.
173  * @out_ctype:    Compression type to use when writing the chunk (ignored if no
174  *                      chunk table provided)
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 write_wim_resource_chunk(const u8 chunk[], unsigned chunk_size,
181                                     FILE *out_fp, compress_func_t compress,
182                                     struct chunk_table *chunk_tab)
183 {
184         const u8 *out_chunk;
185         unsigned out_chunk_size;
186         if (chunk_tab) {
187                 u8 *compressed_chunk = alloca(chunk_size);
188                 int ret;
189
190                 ret = compress(chunk, chunk_size, compressed_chunk,
191                                &out_chunk_size);
192                 if (ret == 0) {
193                         out_chunk = compressed_chunk;
194                 } else {
195                         out_chunk = chunk;
196                         out_chunk_size = chunk_size;
197                 }
198                 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
199                 chunk_tab->cur_offset += out_chunk_size;
200         } else {
201                 out_chunk = chunk;
202                 out_chunk_size = chunk_size;
203         }
204         if (fwrite(out_chunk, 1, out_chunk_size, out_fp) != out_chunk_size) {
205                 ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
206                 return WIMLIB_ERR_WRITE;
207         }
208         return 0;
209 }
210
211 /*
212  * Finishes a WIM chunk tale and writes it to the output file at the correct
213  * offset.
214  *
215  * The final size of the full compressed resource is returned in the
216  * @compressed_size_p.
217  */
218 static int
219 finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
220                               FILE *out_fp, u64 *compressed_size_p)
221 {
222         size_t bytes_written;
223         if (fseeko(out_fp, chunk_tab->file_offset, SEEK_SET) != 0) {
224                 ERROR_WITH_ERRNO("Failed to seek to byte %"PRIu64" of output "
225                                  "WIM file", chunk_tab->file_offset);
226                 return WIMLIB_ERR_WRITE;
227         }
228
229         if (chunk_tab->bytes_per_chunk_entry == 8) {
230                 array_cpu_to_le64(chunk_tab->offsets, chunk_tab->num_chunks);
231         } else {
232                 for (u64 i = 0; i < chunk_tab->num_chunks; i++)
233                         ((u32*)chunk_tab->offsets)[i] =
234                                 cpu_to_le32(chunk_tab->offsets[i]);
235         }
236         bytes_written = fwrite((u8*)chunk_tab->offsets +
237                                         chunk_tab->bytes_per_chunk_entry,
238                                1, chunk_tab->table_disk_size, out_fp);
239         if (bytes_written != chunk_tab->table_disk_size) {
240                 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
241                                  "file resource");
242                 return WIMLIB_ERR_WRITE;
243         }
244         if (fseeko(out_fp, 0, SEEK_END) != 0) {
245                 ERROR_WITH_ERRNO("Failed to seek to end of output WIM file");
246                 return WIMLIB_ERR_WRITE;
247         }
248         *compressed_size_p = chunk_tab->cur_offset + chunk_tab->table_disk_size;
249         return 0;
250 }
251
252 /* Prepare for multiple reads to a resource by caching a FILE * or NTFS
253  * attribute pointer in the lookup table entry. */
254 static int prepare_resource_for_read(struct lookup_table_entry *lte
255
256                                         #ifdef WITH_NTFS_3G
257                                         , ntfs_inode **ni_ret
258                                         #endif
259                 )
260 {
261         if (lte->resource_location == RESOURCE_IN_FILE_ON_DISK
262              && !lte->file_on_disk_fp)
263         {
264                 wimlib_assert(lte->file_on_disk);
265                 lte->file_on_disk_fp = fopen(lte->file_on_disk, "rb");
266                 if (!lte->file_on_disk_fp) {
267                         ERROR_WITH_ERRNO("Failed to open the file `%s' for "
268                                          "reading", lte->file_on_disk);
269                         return WIMLIB_ERR_OPEN;
270                 }
271         }
272 #ifdef WITH_NTFS_3G
273         else if (lte->resource_location == RESOURCE_IN_NTFS_VOLUME
274                   && !lte->attr)
275         {
276                 struct ntfs_location *loc = lte->ntfs_loc;
277                 ntfs_inode *ni;
278                 wimlib_assert(loc);
279                 ni = ntfs_pathname_to_inode(*loc->ntfs_vol_p, NULL, loc->path_utf8);
280                 if (!ni) {
281                         ERROR_WITH_ERRNO("Failed to open inode `%s' in NTFS "
282                                          "volume", loc->path_utf8);
283                         return WIMLIB_ERR_NTFS_3G;
284                 }
285                 lte->attr = ntfs_attr_open(ni,
286                                            loc->is_reparse_point ? AT_REPARSE_POINT : AT_DATA,
287                                            (ntfschar*)loc->stream_name_utf16,
288                                            loc->stream_name_utf16_num_chars);
289                 if (!lte->attr) {
290                         ERROR_WITH_ERRNO("Failed to open attribute of `%s' in "
291                                          "NTFS volume", loc->path_utf8);
292                         ntfs_inode_close(ni);
293                         return WIMLIB_ERR_NTFS_3G;
294                 }
295                 *ni_ret = ni;
296         }
297 #endif
298         return 0;
299 }
300
301 /* Undo prepare_resource_for_read() by closing the cached FILE * or NTFS
302  * attribute. */
303 static void end_wim_resource_read(struct lookup_table_entry *lte
304                                 #ifdef WITH_NTFS_3G
305                                         , ntfs_inode *ni
306                                 #endif
307                                         )
308 {
309         if (lte->resource_location == RESOURCE_IN_FILE_ON_DISK
310             && lte->file_on_disk_fp) {
311                 fclose(lte->file_on_disk_fp);
312                 lte->file_on_disk_fp = NULL;
313         }
314 #ifdef WITH_NTFS_3G
315         else if (lte->resource_location == RESOURCE_IN_NTFS_VOLUME) {
316                 if (lte->attr) {
317                         ntfs_attr_close(lte->attr);
318                         lte->attr = NULL;
319                 }
320                 if (ni)
321                         ntfs_inode_close(ni);
322         }
323 #endif
324 }
325
326 /*
327  * Writes a WIM resource to a FILE * opened for writing.  The resource may be
328  * written uncompressed or compressed depending on the @out_ctype parameter.
329  *
330  * If by chance the resource compresses to more than the original size (this may
331  * happen with random data or files than are pre-compressed), the resource is
332  * instead written uncompressed (and this is reflected in the @out_res_entry by
333  * removing the WIM_RESHDR_FLAG_COMPRESSED flag).
334  *
335  * @lte:        The lookup table entry for the WIM resource.
336  * @out_fp:     The FILE * to write the resource to.
337  * @out_ctype:  The compression type of the resource to write.  Note: if this is
338  *                      the same as the compression type of the WIM resource we
339  *                      need to read, we simply copy the data (i.e. we do not
340  *                      uncompress it, then compress it again).
341  * @out_res_entry:  If non-NULL, a resource entry that is filled in with the
342  *                  offset, original size, compressed size, and compression flag
343  *                  of the output resource.
344  *
345  * Returns 0 on success; nonzero on failure.
346  */
347 int write_wim_resource(struct lookup_table_entry *lte,
348                        FILE *out_fp, int out_ctype,
349                        struct resource_entry *out_res_entry,
350                        int flags)
351 {
352         u64 bytes_remaining;
353         u64 original_size;
354         u64 old_compressed_size;
355         u64 new_compressed_size;
356         u64 offset;
357         int ret;
358         struct chunk_table *chunk_tab = NULL;
359         bool raw;
360         off_t file_offset;
361         compress_func_t compress = NULL;
362 #ifdef WITH_NTFS_3G
363         ntfs_inode *ni = NULL;
364 #endif
365
366         wimlib_assert(lte);
367
368         /* Original size of the resource */
369         original_size = wim_resource_size(lte);
370
371         /* Compressed size of the resource (as it exists now) */
372         old_compressed_size = wim_resource_compressed_size(lte);
373
374         /* Current offset in output file */
375         file_offset = ftello(out_fp);
376         if (file_offset == -1) {
377                 ERROR_WITH_ERRNO("Failed to get offset in output "
378                                  "stream");
379                 return WIMLIB_ERR_WRITE;
380         }
381
382         /* Are the compression types the same?  If so, do a raw copy (copy
383          * without decompressing and recompressing the data). */
384         raw = (wim_resource_compression_type(lte) == out_ctype
385                && out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
386                && !(flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS));
387
388         if (raw) {
389                 flags |= WIMLIB_RESOURCE_FLAG_RAW;
390                 bytes_remaining = old_compressed_size;
391         } else {
392                 flags &= ~WIMLIB_RESOURCE_FLAG_RAW;
393                 bytes_remaining = original_size;
394         }
395
396         /* Empty resource; nothing needs to be done, so just return success. */
397         if (bytes_remaining == 0)
398                 return 0;
399
400         /* Buffer for reading chunks for the resource */
401         u8 buf[min(WIM_CHUNK_SIZE, bytes_remaining)];
402
403         /* If we are writing a compressed resource and not doing a raw copy, we
404          * need to initialize the chunk table */
405         if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE && !raw) {
406                 ret = begin_wim_resource_chunk_tab(lte, out_fp, file_offset,
407                                                    &chunk_tab);
408                 if (ret != 0)
409                         goto out;
410         }
411
412         /* If the WIM resource is in an external file, open a FILE * to it so we
413          * don't have to open a temporary one in read_wim_resource() for each
414          * chunk. */
415 #ifdef WITH_NTFS_3G
416         ret = prepare_resource_for_read(lte, &ni);
417 #else
418         ret = prepare_resource_for_read(lte);
419 #endif
420         if (ret != 0)
421                 goto out;
422
423         /* If we aren't doing a raw copy, we will compute the SHA1 message
424          * digest of the resource as we read it, and verify it's the same as the
425          * hash given in the lookup table entry once we've finished reading the
426          * resource. */
427         SHA_CTX ctx;
428         if (!raw) {
429                 sha1_init(&ctx);
430                 compress = get_compress_func(out_ctype);
431         }
432         offset = 0;
433
434         /* While there are still bytes remaining in the WIM resource, read a
435          * chunk of the resource, update SHA1, then write that chunk using the
436          * desired compression type. */
437         do {
438                 u64 to_read = min(bytes_remaining, WIM_CHUNK_SIZE);
439                 ret = read_wim_resource(lte, buf, to_read, offset, flags);
440                 if (ret != 0)
441                         goto out_fclose;
442                 if (!raw)
443                         sha1_update(&ctx, buf, to_read);
444                 ret = write_wim_resource_chunk(buf, to_read, out_fp,
445                                                compress, chunk_tab);
446                 if (ret != 0)
447                         goto out_fclose;
448                 bytes_remaining -= to_read;
449                 offset += to_read;
450         } while (bytes_remaining);
451
452         /* Raw copy:  The new compressed size is the same as the old compressed
453          * size
454          *
455          * Using WIMLIB_COMPRESSION_TYPE_NONE:  The new compressed size is the
456          * original size
457          *
458          * Using a different compression type:  Call
459          * finish_wim_resource_chunk_tab() and it will provide the new
460          * compressed size.
461          */
462         if (raw) {
463                 new_compressed_size = old_compressed_size;
464         } else {
465                 if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE)
466                         new_compressed_size = original_size;
467                 else {
468                         ret = finish_wim_resource_chunk_tab(chunk_tab, out_fp,
469                                                             &new_compressed_size);
470                         if (ret != 0)
471                                 goto out_fclose;
472                 }
473         }
474
475         /* Verify SHA1 message digest of the resource, unless we are doing a raw
476          * write (in which case we never even saw the uncompressed data).  Or,
477          * if the hash we had before is all 0's, just re-set it to be the new
478          * hash. */
479         if (!raw) {
480                 u8 md[SHA1_HASH_SIZE];
481                 sha1_final(md, &ctx);
482                 if (is_zero_hash(lte->hash)) {
483                         copy_hash(lte->hash, md);
484                 } else if (!hashes_equal(md, lte->hash)) {
485                         ERROR("WIM resource has incorrect hash!");
486                         if (lte->resource_location == RESOURCE_IN_FILE_ON_DISK) {
487                                 ERROR("We were reading it from `%s'; maybe it changed "
488                                       "while we were reading it.",
489                                       lte->file_on_disk);
490                         }
491                         ret = WIMLIB_ERR_INVALID_RESOURCE_HASH;
492                         goto out_fclose;
493                 }
494         }
495
496         if (!raw && new_compressed_size >= original_size &&
497             out_ctype != WIMLIB_COMPRESSION_TYPE_NONE)
498         {
499                 /* Oops!  We compressed the resource to larger than the original
500                  * size.  Write the resource uncompressed instead. */
501                 if (fseeko(out_fp, file_offset, SEEK_SET) != 0) {
502                         ERROR_WITH_ERRNO("Failed to seek to byte %"PRIu64" "
503                                          "of output WIM file", file_offset);
504                         ret = WIMLIB_ERR_WRITE;
505                         goto out_fclose;
506                 }
507                 ret = write_wim_resource(lte, out_fp, WIMLIB_COMPRESSION_TYPE_NONE,
508                                          out_res_entry, flags);
509                 if (ret != 0)
510                         goto out_fclose;
511
512                 ret = fflush_and_ftruncate(out_fp, file_offset + out_res_entry->size);
513                 if (ret != 0)
514                         goto out_fclose;
515         } else {
516                 if (out_res_entry) {
517                         out_res_entry->size          = new_compressed_size;
518                         out_res_entry->original_size = original_size;
519                         out_res_entry->offset        = file_offset;
520                         out_res_entry->flags         = lte->resource_entry.flags
521                                                         & ~WIM_RESHDR_FLAG_COMPRESSED;
522                         if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE)
523                                 out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
524                 }
525         }
526         ret = 0;
527 out_fclose:
528 #ifdef WITH_NTFS_3G
529         end_wim_resource_read(lte, ni);
530 #else
531         end_wim_resource_read(lte);
532 #endif
533 out:
534         FREE(chunk_tab);
535         return ret;
536 }
537
538 #ifdef ENABLE_MULTITHREADED_COMPRESSION
539 struct shared_queue {
540         sem_t filled_slots;
541         sem_t empty_slots;
542         pthread_mutex_t lock;
543         unsigned front;
544         unsigned back;
545         void **array;
546         unsigned size;
547 };
548
549 static int shared_queue_init(struct shared_queue *q, unsigned size)
550 {
551         q->array = CALLOC(sizeof(q->array[0]), size);
552         if (!q->array)
553                 return WIMLIB_ERR_NOMEM;
554
555         sem_init(&q->filled_slots, 0, 0);
556         sem_init(&q->empty_slots, 0, size);
557         pthread_mutex_init(&q->lock, NULL);
558         q->front = 0;
559         q->back = size - 1;
560         q->size = size;
561         return 0;
562 }
563
564 static void shared_queue_destroy(struct shared_queue *q)
565 {
566         sem_destroy(&q->filled_slots);
567         sem_destroy(&q->empty_slots);
568         pthread_mutex_destroy(&q->lock);
569         FREE(q->array);
570 }
571
572 static void shared_queue_put(struct shared_queue *q, void *obj)
573 {
574         sem_wait(&q->empty_slots);
575         pthread_mutex_lock(&q->lock);
576
577         q->back = (q->back + 1) % q->size;
578         q->array[q->back] = obj;
579
580         sem_post(&q->filled_slots);
581         pthread_mutex_unlock(&q->lock);
582 }
583
584 static void *shared_queue_get(struct shared_queue *q)
585 {
586         sem_wait(&q->filled_slots);
587         pthread_mutex_lock(&q->lock);
588
589         void *obj = q->array[q->front];
590         q->array[q->front] = NULL;
591         q->front = (q->front + 1) % q->size;
592
593         sem_post(&q->empty_slots);
594         pthread_mutex_unlock(&q->lock);
595         return obj;
596 }
597
598 struct compressor_thread_params {
599         struct shared_queue *res_to_compress_queue;
600         struct shared_queue *compressed_res_queue;
601         compress_func_t compress;
602 };
603
604 #define MAX_CHUNKS_PER_MSG 2
605
606 struct message {
607         struct lookup_table_entry *lte;
608         u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
609         u8 *out_compressed_chunks[MAX_CHUNKS_PER_MSG];
610         u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
611         unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
612         unsigned compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
613         unsigned num_chunks;
614         struct list_head list;
615         bool complete;
616         u64 begin_chunk;
617 };
618
619 static void compress_chunks(struct message *msg, compress_func_t compress)
620 {
621         for (unsigned i = 0; i < msg->num_chunks; i++) {
622                 DEBUG2("compress chunk %u of %u", i, msg->num_chunks);
623                 int ret = compress(msg->uncompressed_chunks[i],
624                                    msg->uncompressed_chunk_sizes[i],
625                                    msg->compressed_chunks[i],
626                                    &msg->compressed_chunk_sizes[i]);
627                 if (ret == 0) {
628                         msg->out_compressed_chunks[i] = msg->compressed_chunks[i];
629                 } else {
630                         msg->out_compressed_chunks[i] = msg->uncompressed_chunks[i];
631                         msg->compressed_chunk_sizes[i] = msg->uncompressed_chunk_sizes[i];
632                 }
633         }
634 }
635
636 static void *compressor_thread_proc(void *arg)
637 {
638         struct compressor_thread_params *params = arg;
639         struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
640         struct shared_queue *compressed_res_queue = params->compressed_res_queue;
641         compress_func_t compress = params->compress;
642         struct message *msg;
643
644         DEBUG("Compressor thread ready");
645         while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
646                 compress_chunks(msg, compress);
647                 shared_queue_put(compressed_res_queue, msg);
648         }
649         DEBUG("Compressor thread terminating");
650         return NULL;
651 }
652 #endif
653
654 static int do_write_stream_list(struct list_head *my_resources,
655                                 FILE *out_fp,
656                                 int out_ctype,
657                                 wimlib_progress_func_t progress_func,
658                                 union wimlib_progress_info *progress,
659                                 int write_resource_flags)
660 {
661         int ret;
662         struct lookup_table_entry *lte, *tmp;
663
664         list_for_each_entry_safe(lte, tmp, my_resources, staging_list) {
665                 ret = write_wim_resource(lte,
666                                          out_fp,
667                                          out_ctype,
668                                          &lte->output_resource_entry,
669                                          write_resource_flags);
670                 if (ret != 0)
671                         return ret;
672                 list_del(&lte->staging_list);
673                 progress->write_streams.completed_bytes +=
674                         wim_resource_size(lte);
675                 progress->write_streams.completed_streams++;
676                 if (progress_func) {
677                         progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
678                                       progress);
679                 }
680         }
681         return 0;
682 }
683
684 static int write_stream_list_serial(struct list_head *stream_list,
685                                     FILE *out_fp,
686                                     int out_ctype,
687                                     int write_flags,
688                                     wimlib_progress_func_t progress_func,
689                                     union wimlib_progress_info *progress)
690 {
691         int write_resource_flags;
692
693         if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
694                 write_resource_flags = WIMLIB_RESOURCE_FLAG_RECOMPRESS;
695         else
696                 write_resource_flags = 0;
697         progress->write_streams.num_threads = 1;
698         if (progress_func)
699                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
700         return do_write_stream_list(stream_list, out_fp,
701                                     out_ctype, progress_func,
702                                     progress, write_resource_flags);
703 }
704
705 #ifdef ENABLE_MULTITHREADED_COMPRESSION
706 static int write_wim_chunks(struct message *msg, FILE *out_fp,
707                             struct chunk_table *chunk_tab)
708 {
709         for (unsigned i = 0; i < msg->num_chunks; i++) {
710                 unsigned chunk_csize = msg->compressed_chunk_sizes[i];
711
712                 DEBUG2("Write wim chunk %u of %u (csize = %u)",
713                       i, msg->num_chunks, chunk_csize);
714
715                 if (fwrite(msg->out_compressed_chunks[i], 1, chunk_csize, out_fp)
716                     != chunk_csize)
717                 {
718                         ERROR_WITH_ERRNO("Failed to write WIM chunk");
719                         return WIMLIB_ERR_WRITE;
720                 }
721
722                 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
723                 chunk_tab->cur_offset += chunk_csize;
724         }
725         return 0;
726 }
727
728 /*
729  * This function is executed by the main thread when the resources are being
730  * compressed in parallel.  The main thread is in change of all reading of the
731  * uncompressed data and writing of the compressed data.  The compressor threads
732  * *only* do compression from/to in-memory buffers.
733  *
734  * Each unit of work given to a compressor thread is up to MAX_CHUNKS_PER_MSG
735  * chunks of compressed data to compress, represented in a `struct message'.
736  * Each message is passed from the main thread to a worker thread through the
737  * res_to_compress_queue, and it is passed back through the
738  * compressed_res_queue.
739  */
740 static int main_writer_thread_proc(struct list_head *stream_list,
741                                    FILE *out_fp,
742                                    int out_ctype,
743                                    struct shared_queue *res_to_compress_queue,
744                                    struct shared_queue *compressed_res_queue,
745                                    size_t queue_size,
746                                    int write_flags,
747                                    wimlib_progress_func_t progress_func,
748                                    union wimlib_progress_info *progress)
749 {
750         int ret;
751
752         struct message msgs[queue_size];
753         ZERO_ARRAY(msgs);
754
755         // Initially, all the messages are available to use.
756         LIST_HEAD(available_msgs);
757         for (size_t i = 0; i < ARRAY_LEN(msgs); i++)
758                 list_add(&msgs[i].list, &available_msgs);
759
760         // outstanding_resources is the list of resources that currently have
761         // had chunks sent off for compression.
762         //
763         // The first stream in outstanding_resources is the stream that is
764         // currently being written (cur_lte).
765         //
766         // The last stream in outstanding_resources is the stream that is
767         // currently being read and chunks fed to the compressor threads
768         // (next_lte).
769         //
770         // Depending on the number of threads and the sizes of the resource,
771         // the outstanding streams list may contain streams between cur_lte and
772         // next_lte that have all their chunks compressed or being compressed,
773         // but haven't been written yet.
774         //
775         LIST_HEAD(outstanding_resources);
776         struct list_head *next_resource = stream_list->next;
777         struct lookup_table_entry *next_lte = container_of(next_resource,
778                                                            struct lookup_table_entry,
779                                                            staging_list);
780         next_resource = next_resource->next;
781         u64 next_chunk = 0;
782         u64 next_num_chunks = wim_resource_chunks(next_lte);
783         INIT_LIST_HEAD(&next_lte->msg_list);
784         list_add_tail(&next_lte->staging_list, &outstanding_resources);
785
786         // As in write_wim_resource(), each resource we read is checksummed.
787         SHA_CTX next_sha_ctx;
788         sha1_init(&next_sha_ctx);
789         u8 next_hash[SHA1_HASH_SIZE];
790
791         // Resources that don't need any chunks compressed are added to this
792         // list and written directly by the main thread.
793         LIST_HEAD(my_resources);
794
795         struct lookup_table_entry *cur_lte = next_lte;
796         struct chunk_table *cur_chunk_tab = NULL;
797         struct message *msg;
798
799 #ifdef WITH_NTFS_3G
800         ntfs_inode *ni = NULL;
801 #endif
802
803 #ifdef WITH_NTFS_3G
804         ret = prepare_resource_for_read(next_lte, &ni);
805 #else
806         ret = prepare_resource_for_read(next_lte);
807 #endif
808         if (ret != 0)
809                 goto out;
810
811         DEBUG("Initializing buffers for uncompressed "
812               "and compressed data (%zu bytes needed)",
813               queue_size * MAX_CHUNKS_PER_MSG * WIM_CHUNK_SIZE * 2);
814
815         // Pre-allocate all the buffers that will be needed to do the chunk
816         // compression.
817         for (size_t i = 0; i < ARRAY_LEN(msgs); i++) {
818                 for (size_t j = 0; j < MAX_CHUNKS_PER_MSG; j++) {
819                         msgs[i].compressed_chunks[j] = MALLOC(WIM_CHUNK_SIZE);
820                         msgs[i].uncompressed_chunks[j] = MALLOC(WIM_CHUNK_SIZE);
821                         if (msgs[i].compressed_chunks[j] == NULL ||
822                             msgs[i].uncompressed_chunks[j] == NULL)
823                         {
824                                 ERROR("Could not allocate enough memory for "
825                                       "multi-threaded compression");
826                                 ret = WIMLIB_ERR_NOMEM;
827                                 goto out;
828                         }
829                 }
830         }
831
832         // This loop is executed until all resources have been written, except
833         // possibly a few that have been added to the @my_resources list for
834         // writing later.
835         while (1) {
836                 // Send chunks to the compressor threads until either (a) there
837                 // are no more messages available since they were all sent off,
838                 // or (b) there are no more resources that need to be
839                 // compressed.
840                 while (!list_empty(&available_msgs) && next_lte != NULL) {
841
842                         // Get a message from the available messages
843                         // list
844                         msg = container_of(available_msgs.next,
845                                            struct message,
846                                            list);
847
848                         // ... and delete it from the available messages
849                         // list
850                         list_del(&msg->list);
851
852                         // Initialize the message with the chunks to
853                         // compress.
854                         msg->num_chunks = min(next_num_chunks - next_chunk,
855                                               MAX_CHUNKS_PER_MSG);
856                         msg->lte = next_lte;
857                         msg->complete = false;
858                         msg->begin_chunk = next_chunk;
859
860                         unsigned size = WIM_CHUNK_SIZE;
861                         for (unsigned i = 0; i < msg->num_chunks; i++) {
862
863                                 // Read chunk @next_chunk of the stream into the
864                                 // message so that a compressor thread can
865                                 // compress it.
866
867                                 if (next_chunk == next_num_chunks - 1 &&
868                                      wim_resource_size(next_lte) % WIM_CHUNK_SIZE != 0)
869                                 {
870                                         size = wim_resource_size(next_lte) % WIM_CHUNK_SIZE;
871                                 }
872
873
874                                 DEBUG2("Read resource (size=%u, offset=%zu)",
875                                       size, next_chunk * WIM_CHUNK_SIZE);
876
877                                 msg->uncompressed_chunk_sizes[i] = size;
878
879                                 ret = read_wim_resource(next_lte,
880                                                         msg->uncompressed_chunks[i],
881                                                         size,
882                                                         next_chunk * WIM_CHUNK_SIZE,
883                                                         0);
884                                 if (ret != 0)
885                                         goto out;
886                                 sha1_update(&next_sha_ctx,
887                                             msg->uncompressed_chunks[i], size);
888                                 next_chunk++;
889                         }
890
891                         // Send the compression request
892                         list_add_tail(&msg->list, &next_lte->msg_list);
893                         shared_queue_put(res_to_compress_queue, msg);
894                         DEBUG2("Compression request sent");
895
896                         if (next_chunk != next_num_chunks)
897                                 // More chunks to send for this resource
898                                 continue;
899
900                         // Done sending compression requests for a resource!
901                         // Check the SHA1 message digest.
902                         DEBUG2("Finalize SHA1 md (next_num_chunks=%zu)", next_num_chunks);
903                         sha1_final(next_hash, &next_sha_ctx);
904                         if (!hashes_equal(next_lte->hash, next_hash)) {
905                                 ERROR("WIM resource has incorrect hash!");
906                                 if (next_lte->resource_location == RESOURCE_IN_FILE_ON_DISK) {
907                                         ERROR("We were reading it from `%s'; maybe it changed "
908                                               "while we were reading it.",
909                                               next_lte->file_on_disk);
910                                 }
911                                 ret = WIMLIB_ERR_INVALID_RESOURCE_HASH;
912                                 goto out;
913                         }
914
915                         // Advance to the next resource.
916                         //
917                         // If the next resource needs no compression, just write
918                         // it with this thread (not now though--- we could be in
919                         // the middle of writing another resource.)  Keep doing
920                         // this until we either get to the end of the resources
921                         // list, or we get to a resource that needs compression.
922
923                         while (1) {
924                                 if (next_resource == stream_list) {
925                                         next_lte = NULL;
926                                         break;
927                                 }
928                         #ifdef WITH_NTFS_3G
929                                 end_wim_resource_read(next_lte, ni);
930                                 ni = NULL;
931                         #else
932                                 end_wim_resource_read(next_lte);
933                         #endif
934
935                                 next_lte = container_of(next_resource,
936                                                         struct lookup_table_entry,
937                                                         staging_list);
938                                 next_resource = next_resource->next;
939                                 if ((!(write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
940                                       && next_lte->resource_location == RESOURCE_IN_WIM
941                                       && wimlib_get_compression_type(next_lte->wim) == out_ctype)
942                                     || wim_resource_size(next_lte) == 0)
943                                 {
944                                         list_add_tail(&next_lte->staging_list,
945                                                       &my_resources);
946                                 } else {
947                                         list_add_tail(&next_lte->staging_list,
948                                                       &outstanding_resources);
949                                         next_chunk = 0;
950                                         next_num_chunks = wim_resource_chunks(next_lte);
951                                         sha1_init(&next_sha_ctx);
952                                         INIT_LIST_HEAD(&next_lte->msg_list);
953                                 #ifdef WITH_NTFS_3G
954                                         ret = prepare_resource_for_read(next_lte, &ni);
955                                 #else
956                                         ret = prepare_resource_for_read(next_lte);
957                                 #endif
958                                         if (ret != 0)
959                                                 goto out;
960                                         DEBUG2("Updated next_lte");
961                                         break;
962                                 }
963                         }
964                 }
965
966                 // If there are no outstanding resources, there are no more
967                 // resources that need to be written.
968                 if (list_empty(&outstanding_resources)) {
969                         DEBUG("No outstanding resources! Done");
970                         ret = 0;
971                         goto out;
972                 }
973
974                 // Get the next message from the queue and process it.
975                 // The message will contain 1 or more data chunks that have been
976                 // compressed.
977                 DEBUG2("Waiting for message");
978                 msg = shared_queue_get(compressed_res_queue);
979                 msg->complete = true;
980
981                 DEBUG2("Received msg (begin_chunk=%"PRIu64")", msg->begin_chunk);
982
983                 list_for_each_entry(msg, &cur_lte->msg_list, list) {
984                         DEBUG2("complete=%d", msg->complete);
985                 }
986
987                 // Is this the next chunk in the current resource?  If it's not
988                 // (i.e., an earlier chunk in a same or different resource
989                 // hasn't been compressed yet), do nothing, and keep this
990                 // message around until all earlier chunks are received.
991                 //
992                 // Otherwise, write all the chunks we can.
993                 while (!list_empty(&cur_lte->msg_list)
994                         && (msg = container_of(cur_lte->msg_list.next,
995                                                struct message,
996                                                list))->complete)
997                 {
998                         DEBUG2("Complete msg (begin_chunk=%"PRIu64")", msg->begin_chunk);
999                         if (msg->begin_chunk == 0) {
1000                                 DEBUG2("Begin chunk tab");
1001
1002                                 // This is the first set of chunks.  Leave space
1003                                 // for the chunk table in the output file.
1004                                 off_t cur_offset = ftello(out_fp);
1005                                 if (cur_offset == -1) {
1006                                         ret = WIMLIB_ERR_WRITE;
1007                                         goto out;
1008                                 }
1009                                 ret = begin_wim_resource_chunk_tab(cur_lte,
1010                                                                    out_fp,
1011                                                                    cur_offset,
1012                                                                    &cur_chunk_tab);
1013                                 if (ret != 0)
1014                                         goto out;
1015                         }
1016
1017                         // Write the compressed chunks from the message.
1018                         ret = write_wim_chunks(msg, out_fp, cur_chunk_tab);
1019                         if (ret != 0)
1020                                 goto out;
1021
1022                         list_del(&msg->list);
1023
1024                         // This message is available to use for different chunks
1025                         // now.
1026                         list_add(&msg->list, &available_msgs);
1027
1028                         // Was this the last chunk of the stream?  If so,
1029                         // finish it.
1030                         if (list_empty(&cur_lte->msg_list) &&
1031                             msg->begin_chunk + msg->num_chunks == cur_chunk_tab->num_chunks)
1032                         {
1033                                 DEBUG2("Finish wim chunk tab");
1034                                 u64 res_csize;
1035                                 ret = finish_wim_resource_chunk_tab(cur_chunk_tab,
1036                                                                     out_fp,
1037                                                                     &res_csize);
1038                                 if (ret != 0)
1039                                         goto out;
1040
1041                                 progress->write_streams.completed_bytes +=
1042                                                 wim_resource_size(cur_lte);
1043                                 progress->write_streams.completed_streams++;
1044
1045                                 if (progress_func) {
1046                                         progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
1047                                                       progress);
1048                                 }
1049
1050                                 cur_lte->output_resource_entry.size =
1051                                         res_csize;
1052
1053                                 cur_lte->output_resource_entry.original_size =
1054                                         cur_lte->resource_entry.original_size;
1055
1056                                 cur_lte->output_resource_entry.offset =
1057                                         cur_chunk_tab->file_offset;
1058
1059                                 cur_lte->output_resource_entry.flags =
1060                                         cur_lte->resource_entry.flags |
1061                                                 WIM_RESHDR_FLAG_COMPRESSED;
1062
1063                                 FREE(cur_chunk_tab);
1064                                 cur_chunk_tab = NULL;
1065
1066                                 struct list_head *next = cur_lte->staging_list.next;
1067                                 list_del(&cur_lte->staging_list);
1068
1069                                 if (next == &outstanding_resources) {
1070                                         DEBUG("No more outstanding resources");
1071                                         ret = 0;
1072                                         goto out;
1073                                 } else {
1074                                         cur_lte = container_of(cur_lte->staging_list.next,
1075                                                                struct lookup_table_entry,
1076                                                                staging_list);
1077                                 }
1078
1079                                 // Since we just finished writing a stream,
1080                                 // write any streams that have been added to the
1081                                 // my_resources list for direct writing by the
1082                                 // main thread (e.g. resources that don't need
1083                                 // to be compressed because the desired
1084                                 // compression type is the same as the previous
1085                                 // compression type).
1086                                 ret = do_write_stream_list(&my_resources,
1087                                                            out_fp,
1088                                                            out_ctype,
1089                                                            progress_func,
1090                                                            progress,
1091                                                            0);
1092                                 if (ret != 0)
1093                                         goto out;
1094                         }
1095                 }
1096         }
1097
1098 out:
1099 #ifdef WITH_NTFS_3G
1100         end_wim_resource_read(cur_lte, ni);
1101 #else
1102         end_wim_resource_read(cur_lte);
1103 #endif
1104         if (ret == 0) {
1105                 ret = do_write_stream_list(&my_resources, out_fp,
1106                                            out_ctype, progress_func,
1107                                            progress, 0);
1108         } else {
1109                 size_t num_available_msgs = 0;
1110                 struct list_head *cur;
1111
1112                 list_for_each(cur, &available_msgs) {
1113                         num_available_msgs++;
1114                 }
1115
1116                 while (num_available_msgs < ARRAY_LEN(msgs)) {
1117                         shared_queue_get(compressed_res_queue);
1118                         num_available_msgs++;
1119                 }
1120         }
1121
1122         for (size_t i = 0; i < ARRAY_LEN(msgs); i++) {
1123                 for (size_t j = 0; j < MAX_CHUNKS_PER_MSG; j++) {
1124                         FREE(msgs[i].compressed_chunks[j]);
1125                         FREE(msgs[i].uncompressed_chunks[j]);
1126                 }
1127         }
1128
1129         if (cur_chunk_tab != NULL)
1130                 FREE(cur_chunk_tab);
1131         return ret;
1132 }
1133
1134
1135 static int write_stream_list_parallel(struct list_head *stream_list,
1136                                       FILE *out_fp,
1137                                       int out_ctype,
1138                                       int write_flags,
1139                                       unsigned num_threads,
1140                                       wimlib_progress_func_t progress_func,
1141                                       union wimlib_progress_info *progress)
1142 {
1143         int ret;
1144         struct shared_queue res_to_compress_queue;
1145         struct shared_queue compressed_res_queue;
1146         pthread_t *compressor_threads = NULL;
1147
1148         if (num_threads == 0) {
1149                 long nthreads = sysconf(_SC_NPROCESSORS_ONLN);
1150                 if (nthreads < 1) {
1151                         WARNING("Could not determine number of processors! Assuming 1");
1152                         goto out_serial;
1153                 } else {
1154                         num_threads = nthreads;
1155                 }
1156         }
1157
1158         progress->write_streams.num_threads = num_threads;
1159         wimlib_assert(stream_list->next != stream_list);
1160
1161         static const double MESSAGES_PER_THREAD = 2.0;
1162         size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1163
1164         DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1165
1166         ret = shared_queue_init(&res_to_compress_queue, queue_size);
1167         if (ret != 0)
1168                 goto out_serial;
1169
1170         ret = shared_queue_init(&compressed_res_queue, queue_size);
1171         if (ret != 0)
1172                 goto out_destroy_res_to_compress_queue;
1173
1174         struct compressor_thread_params params;
1175         params.res_to_compress_queue = &res_to_compress_queue;
1176         params.compressed_res_queue = &compressed_res_queue;
1177         params.compress = get_compress_func(out_ctype);
1178
1179         compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1180
1181         for (unsigned i = 0; i < num_threads; i++) {
1182                 DEBUG("pthread_create thread %u", i);
1183                 ret = pthread_create(&compressor_threads[i], NULL,
1184                                      compressor_thread_proc, &params);
1185                 if (ret != 0) {
1186                         ret = -1;
1187                         ERROR_WITH_ERRNO("Failed to create compressor "
1188                                          "thread %u", i);
1189                         num_threads = i;
1190                         goto out_join;
1191                 }
1192         }
1193
1194         if (progress_func)
1195                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
1196
1197         ret = main_writer_thread_proc(stream_list,
1198                                       out_fp,
1199                                       out_ctype,
1200                                       &res_to_compress_queue,
1201                                       &compressed_res_queue,
1202                                       queue_size,
1203                                       write_flags,
1204                                       progress_func,
1205                                       progress);
1206 out_join:
1207         for (unsigned i = 0; i < num_threads; i++)
1208                 shared_queue_put(&res_to_compress_queue, NULL);
1209
1210         for (unsigned i = 0; i < num_threads; i++) {
1211                 if (pthread_join(compressor_threads[i], NULL)) {
1212                         WARNING("Failed to join compressor thread %u: %s",
1213                                 i, strerror(errno));
1214                 }
1215         }
1216         FREE(compressor_threads);
1217         shared_queue_destroy(&compressed_res_queue);
1218 out_destroy_res_to_compress_queue:
1219         shared_queue_destroy(&res_to_compress_queue);
1220         if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
1221                 return ret;
1222 out_serial:
1223         WARNING("Falling back to single-threaded compression");
1224         return write_stream_list_serial(stream_list,
1225                                         out_fp,
1226                                         out_ctype,
1227                                         write_flags,
1228                                         progress_func,
1229                                         progress);
1230
1231 }
1232 #endif
1233
1234 /*
1235  * Write a list of streams to a WIM (@out_fp) using the compression type
1236  * @out_ctype and up to @num_threads compressor threads.
1237  */
1238 static int write_stream_list(struct list_head *stream_list, FILE *out_fp,
1239                              int out_ctype, int write_flags,
1240                              unsigned num_threads,
1241                              wimlib_progress_func_t progress_func)
1242 {
1243         struct lookup_table_entry *lte;
1244         size_t num_streams = 0;
1245         u64 total_bytes = 0;
1246         bool compression_needed = false;
1247         union wimlib_progress_info progress;
1248         int ret;
1249
1250         list_for_each_entry(lte, stream_list, staging_list) {
1251                 num_streams++;
1252                 total_bytes += wim_resource_size(lte);
1253                 if (!compression_needed
1254                     &&
1255                     (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
1256                        && (lte->resource_location != RESOURCE_IN_WIM
1257                            || wimlib_get_compression_type(lte->wim) != out_ctype
1258                            || (write_flags & WIMLIB_WRITE_FLAG_REBUILD)))
1259                     && wim_resource_size(lte) != 0)
1260                         compression_needed = true;
1261         }
1262         progress.write_streams.total_bytes       = total_bytes;
1263         progress.write_streams.total_streams     = num_streams;
1264         progress.write_streams.completed_bytes   = 0;
1265         progress.write_streams.completed_streams = 0;
1266         progress.write_streams.num_threads       = num_threads;
1267         progress.write_streams.compression_type  = out_ctype;
1268
1269         if (num_streams == 0) {
1270                 ret = 0;
1271                 goto out;
1272         }
1273
1274 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1275         if (compression_needed && total_bytes >= 1000000 && num_threads != 1) {
1276                 ret = write_stream_list_parallel(stream_list,
1277                                                  out_fp,
1278                                                  out_ctype,
1279                                                  write_flags,
1280                                                  num_threads,
1281                                                  progress_func,
1282                                                  &progress);
1283         }
1284         else
1285 #endif
1286         {
1287                 ret = write_stream_list_serial(stream_list,
1288                                                out_fp,
1289                                                out_ctype,
1290                                                write_flags,
1291                                                progress_func,
1292                                                &progress);
1293         }
1294 out:
1295         return ret;
1296 }
1297
1298
1299 static int dentry_find_streams_to_write(struct dentry *dentry,
1300                                         void *wim)
1301 {
1302         WIMStruct *w = wim;
1303         struct list_head *stream_list = w->private;
1304         struct lookup_table_entry *lte;
1305         for (unsigned i = 0; i <= dentry->d_inode->num_ads; i++) {
1306                 lte = inode_stream_lte(dentry->d_inode, i, w->lookup_table);
1307                 if (lte && ++lte->out_refcnt == 1)
1308                         list_add_tail(&lte->staging_list, stream_list);
1309         }
1310         return 0;
1311 }
1312
1313 static int find_streams_to_write(WIMStruct *w)
1314 {
1315         return for_dentry_in_tree(wim_root_dentry(w),
1316                                   dentry_find_streams_to_write, w);
1317 }
1318
1319 static int write_wim_streams(WIMStruct *w, int image, int write_flags,
1320                              unsigned num_threads,
1321                              wimlib_progress_func_t progress_func)
1322 {
1323
1324         for_lookup_table_entry(w->lookup_table, lte_zero_out_refcnt, NULL);
1325         LIST_HEAD(stream_list);
1326         w->private = &stream_list;
1327         for_image(w, image, find_streams_to_write);
1328         return write_stream_list(&stream_list, w->out_fp,
1329                                  wimlib_get_compression_type(w), write_flags,
1330                                  num_threads, progress_func);
1331 }
1332
1333 /*
1334  * Finish writing a WIM file: write the lookup table, xml data, and integrity
1335  * table (optional), then overwrite the WIM header.
1336  *
1337  * write_flags is a bitwise OR of the following:
1338  *
1339  *      (public)  WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
1340  *              Include an integrity table.
1341  *
1342  *      (public)  WIMLIB_WRITE_FLAG_SHOW_PROGRESS:
1343  *              Show progress information when (if) writing the integrity table.
1344  *
1345  *      (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
1346  *              Don't write the lookup table.
1347  *
1348  *      (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
1349  *              When (if) writing the integrity table, re-use entries from the
1350  *              existing integrity table, if possible.
1351  *
1352  *      (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
1353  *              After writing the XML data but before writing the integrity
1354  *              table, write a temporary WIM header and flush the stream so that
1355  *              the WIM is less likely to become corrupted upon abrupt program
1356  *              termination.
1357  *
1358  *      (private) WIMLIB_WRITE_FLAG_FSYNC:
1359  *              fsync() the output file before closing it.
1360  *
1361  */
1362 int finish_write(WIMStruct *w, int image, int write_flags,
1363                  wimlib_progress_func_t progress_func)
1364 {
1365         int ret;
1366         struct wim_header hdr;
1367         FILE *out = w->out_fp;
1368
1369         /* @hdr will be the header for the new WIM.  First copy all the data
1370          * from the header in the WIMStruct; then set all the fields that may
1371          * have changed, including the resource entries, boot index, and image
1372          * count.  */
1373         memcpy(&hdr, &w->hdr, sizeof(struct wim_header));
1374
1375         if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
1376                 ret = write_lookup_table(w->lookup_table, out, &hdr.lookup_table_res_entry);
1377                 if (ret != 0)
1378                         goto out;
1379         }
1380
1381         ret = write_xml_data(w->wim_info, image, out,
1382                              (write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE) ?
1383                               wim_info_get_total_bytes(w->wim_info) : 0,
1384                              &hdr.xml_res_entry);
1385         if (ret != 0)
1386                 goto out;
1387
1388         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
1389                 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
1390                         struct wim_header checkpoint_hdr;
1391                         memcpy(&checkpoint_hdr, &hdr, sizeof(struct wim_header));
1392                         memset(&checkpoint_hdr.integrity, 0, sizeof(struct resource_entry));
1393                         if (fseeko(out, 0, SEEK_SET) != 0) {
1394                                 ret = WIMLIB_ERR_WRITE;
1395                                 goto out;
1396                         }
1397                         ret = write_header(&checkpoint_hdr, out);
1398                         if (ret != 0)
1399                                 goto out;
1400
1401                         if (fflush(out) != 0) {
1402                                 ERROR_WITH_ERRNO("Can't write data to WIM");
1403                                 ret = WIMLIB_ERR_WRITE;
1404                                 goto out;
1405                         }
1406
1407                         if (fseeko(out, 0, SEEK_END) != 0) {
1408                                 ret = WIMLIB_ERR_WRITE;
1409                                 goto out;
1410                         }
1411                 }
1412
1413                 off_t old_lookup_table_end;
1414                 off_t new_lookup_table_end;
1415                 if (write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE) {
1416                         old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
1417                                                w->hdr.lookup_table_res_entry.size;
1418                 } else {
1419                         old_lookup_table_end = 0;
1420                 }
1421                 new_lookup_table_end = hdr.lookup_table_res_entry.offset +
1422                                        hdr.lookup_table_res_entry.size;
1423
1424                 ret = write_integrity_table(out,
1425                                             &hdr.integrity,
1426                                             new_lookup_table_end,
1427                                             old_lookup_table_end,
1428                                             progress_func);
1429                 if (ret != 0)
1430                         goto out;
1431         } else {
1432                 memset(&hdr.integrity, 0, sizeof(struct resource_entry));
1433         }
1434
1435         /*
1436          * In the WIM header, there is room for the resource entry for a
1437          * metadata resource labeled as the "boot metadata".  This entry should
1438          * be zeroed out if there is no bootable image (boot_idx 0).  Otherwise,
1439          * it should be a copy of the resource entry for the image that is
1440          * marked as bootable.  This is not well documented...
1441          */
1442         if (hdr.boot_idx == 0 || !w->image_metadata
1443                         || (image != WIMLIB_ALL_IMAGES && image != hdr.boot_idx)) {
1444                 memset(&hdr.boot_metadata_res_entry, 0,
1445                        sizeof(struct resource_entry));
1446         } else {
1447                 memcpy(&hdr.boot_metadata_res_entry,
1448                        &w->image_metadata[
1449                           hdr.boot_idx - 1].metadata_lte->output_resource_entry,
1450                        sizeof(struct resource_entry));
1451         }
1452
1453         /* Set image count and boot index correctly for single image writes */
1454         if (image != WIMLIB_ALL_IMAGES) {
1455                 hdr.image_count = 1;
1456                 if (hdr.boot_idx == image)
1457                         hdr.boot_idx = 1;
1458                 else
1459                         hdr.boot_idx = 0;
1460         }
1461
1462         if (fseeko(out, 0, SEEK_SET) != 0) {
1463                 ret = WIMLIB_ERR_WRITE;
1464                 goto out;
1465         }
1466
1467         ret = write_header(&hdr, out);
1468         if (ret != 0)
1469                 goto out;
1470
1471         if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
1472                 if (fflush(out) != 0
1473                     || fsync(fileno(out)) != 0)
1474                 {
1475                         ERROR_WITH_ERRNO("Error flushing data to WIM file");
1476                         ret = WIMLIB_ERR_WRITE;
1477                 }
1478         }
1479 out:
1480         if (fclose(out) != 0) {
1481                 ERROR_WITH_ERRNO("Failed to close the WIM file");
1482                 if (ret == 0)
1483                         ret = WIMLIB_ERR_WRITE;
1484         }
1485         w->out_fp = NULL;
1486         return ret;
1487 }
1488
1489 static void close_wim_writable(WIMStruct *w)
1490 {
1491         if (w->out_fp) {
1492                 if (fclose(w->out_fp) != 0) {
1493                         WARNING("Failed to close output WIM: %s",
1494                                 strerror(errno));
1495                 }
1496                 w->out_fp = NULL;
1497         }
1498 }
1499
1500 /* Open file stream and write dummy header for WIM. */
1501 int begin_write(WIMStruct *w, const char *path, int write_flags)
1502 {
1503         int ret;
1504         bool need_readable = false;
1505         bool trunc = true;
1506         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
1507                 need_readable = true;
1508
1509         ret = open_wim_writable(w, path, trunc, need_readable);
1510         if (ret != 0)
1511                 return ret;
1512         /* Write dummy header. It will be overwritten later. */
1513         return write_header(&w->hdr, w->out_fp);
1514 }
1515
1516 /* Writes a stand-alone WIM to a file.  */
1517 WIMLIBAPI int wimlib_write(WIMStruct *w, const char *path,
1518                            int image, int write_flags, unsigned num_threads,
1519                            wimlib_progress_func_t progress_func)
1520 {
1521         int ret;
1522
1523         if (!w || !path)
1524                 return WIMLIB_ERR_INVALID_PARAM;
1525
1526         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1527
1528         if (image != WIMLIB_ALL_IMAGES &&
1529              (image < 1 || image > w->hdr.image_count))
1530                 return WIMLIB_ERR_INVALID_IMAGE;
1531
1532         if (w->hdr.total_parts != 1) {
1533                 ERROR("Cannot call wimlib_write() on part of a split WIM");
1534                 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1535         }
1536
1537         ret = begin_write(w, path, write_flags);
1538         if (ret != 0)
1539                 goto out;
1540
1541         ret = write_wim_streams(w, image, write_flags, num_threads,
1542                                 progress_func);
1543         if (ret != 0)
1544                 goto out;
1545
1546         if (progress_func)
1547                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
1548
1549         ret = for_image(w, image, write_metadata_resource);
1550         if (ret != 0)
1551                 goto out;
1552
1553         if (progress_func)
1554                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
1555
1556         ret = finish_write(w, image, write_flags, progress_func);
1557 out:
1558         close_wim_writable(w);
1559         return ret;
1560 }
1561
1562 static int lte_overwrite_prepare(struct lookup_table_entry *lte,
1563                                  void *ignore)
1564 {
1565         memcpy(&lte->output_resource_entry, &lte->resource_entry,
1566                sizeof(struct resource_entry));
1567         lte->out_refcnt = 0;
1568         return 0;
1569 }
1570
1571 static int check_resource_offset(struct lookup_table_entry *lte, void *arg)
1572 {
1573         off_t end_offset = *(u64*)arg;
1574
1575         wimlib_assert(lte->out_refcnt <= lte->refcnt);
1576         if (lte->out_refcnt < lte->refcnt) {
1577                 if (lte->resource_entry.offset + lte->resource_entry.size > end_offset) {
1578                         ERROR("The following resource is after the XML data:");
1579                         print_lookup_table_entry(lte);
1580                         return WIMLIB_ERR_RESOURCE_ORDER;
1581                 }
1582         }
1583         return 0;
1584 }
1585
1586 static int find_new_streams(struct lookup_table_entry *lte, void *arg)
1587 {
1588         if (lte->out_refcnt == lte->refcnt)
1589                 list_add(&lte->staging_list, (struct list_head*)arg);
1590         else
1591                 lte->out_refcnt = lte->refcnt;
1592         return 0;
1593 }
1594
1595 /*
1596  * Overwrite a WIM, possibly appending streams to it.
1597  *
1598  * A WIM looks like (or is supposed to look like) the following:
1599  *
1600  *                   Header (212 bytes)
1601  *                   Streams and metadata resources (variable size)
1602  *                   Lookup table (variable size)
1603  *                   XML data (variable size)
1604  *                   Integrity table (optional) (variable size)
1605  *
1606  * If we are not adding any streams or metadata resources, the lookup table is
1607  * unchanged--- so we only need to overwrite the XML data, integrity table, and
1608  * header.  This operation is potentially unsafe if the program is abruptly
1609  * terminated while the XML data or integrity table are being overwritten, but
1610  * before the new header has been written.  To partially alleviate this problem,
1611  * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
1612  * finish_write() to cause a temporary WIM header to be written after the XML
1613  * data has been written.  This may prevent the WIM from becoming corrupted if
1614  * the program is terminated while the integrity table is being calculated (but
1615  * no guarantees, due to write re-ordering...).
1616  *
1617  * If we are adding new streams or images (metadata resources), the lookup table
1618  * needs to be changed, and those streams need to be written.  In this case, we
1619  * try to perform a safe update of the WIM file by writing the streams *after*
1620  * the end of the previous WIM, then writing the new lookup table, XML data, and
1621  * (optionally) integrity table following the new streams.  This will produce a
1622  * layout like the following:
1623  *
1624  *                   Header (212 bytes)
1625  *                   (OLD) Streams and metadata resources (variable size)
1626  *                   (OLD) Lookup table (variable size)
1627  *                   (OLD) XML data (variable size)
1628  *                   (OLD) Integrity table (optional) (variable size)
1629  *                   (NEW) Streams and metadata resources (variable size)
1630  *                   (NEW) Lookup table (variable size)
1631  *                   (NEW) XML data (variable size)
1632  *                   (NEW) Integrity table (optional) (variable size)
1633  *
1634  * At all points, the WIM is valid as nothing points to the new data yet.  Then,
1635  * the header is overwritten to point to the new lookup table, XML data, and
1636  * integrity table, to produce the following layout:
1637  *
1638  *                   Header (212 bytes)
1639  *                   Streams and metadata resources (variable size)
1640  *                   Nothing (variable size)
1641  *                   More Streams and metadata resources (variable size)
1642  *                   Lookup table (variable size)
1643  *                   XML data (variable size)
1644  *                   Integrity table (optional) (variable size)
1645  *
1646  * This method allows an image to be appended to a large WIM very quickly, and
1647  * is is crash-safe except in the case of write re-ordering, but the
1648  * disadvantage is that a small hole is left in the WIM where the old lookup
1649  * table, xml data, and integrity table were.  (These usually only take up a
1650  * small amount of space compared to the streams, however.
1651  */
1652 static int overwrite_wim_inplace(WIMStruct *w, int write_flags,
1653                                  unsigned num_threads,
1654                                  wimlib_progress_func_t progress_func,
1655                                  int modified_image_idx)
1656 {
1657         int ret;
1658         struct list_head stream_list;
1659         off_t old_wim_end;
1660
1661         DEBUG("Overwriting `%s' in-place", w->filename);
1662
1663         /* Make sure that the integrity table (if present) is after the XML
1664          * data, and that there are no stream resources, metadata resources, or
1665          * lookup tables after the XML data.  Otherwise, these data would be
1666          * overwritten. */
1667         if (w->hdr.integrity.offset != 0 &&
1668             w->hdr.integrity.offset < w->hdr.xml_res_entry.offset) {
1669                 ERROR("Didn't expect the integrity table to be before the XML data");
1670                 return WIMLIB_ERR_RESOURCE_ORDER;
1671         }
1672
1673         if (w->hdr.lookup_table_res_entry.offset > w->hdr.xml_res_entry.offset) {
1674                 ERROR("Didn't expect the lookup table to be after the XML data");
1675                 return WIMLIB_ERR_RESOURCE_ORDER;
1676         }
1677
1678         DEBUG("Identifying newly added streams");
1679         for_lookup_table_entry(w->lookup_table, lte_overwrite_prepare, NULL);
1680         INIT_LIST_HEAD(&stream_list);
1681         for (int i = modified_image_idx; i < w->hdr.image_count; i++) {
1682                 DEBUG("Identifiying streams in image %d", i + 1);
1683                 wimlib_assert(w->image_metadata[i].modified);
1684                 wimlib_assert(!w->image_metadata[i].has_been_mounted_rw);
1685                 wimlib_assert(w->image_metadata[i].root_dentry != NULL);
1686                 wimlib_assert(w->image_metadata[i].metadata_lte != NULL);
1687                 w->private = &stream_list;
1688                 for_dentry_in_tree(w->image_metadata[i].root_dentry,
1689                                    dentry_find_streams_to_write, w);
1690         }
1691
1692         if (w->hdr.integrity.offset)
1693                 old_wim_end = w->hdr.integrity.offset + w->hdr.integrity.size;
1694         else
1695                 old_wim_end = w->hdr.xml_res_entry.offset + w->hdr.xml_res_entry.size;
1696
1697         ret = for_lookup_table_entry(w->lookup_table, check_resource_offset,
1698                                      &old_wim_end);
1699         if (ret != 0)
1700                 return ret;
1701
1702         if (modified_image_idx == w->hdr.image_count && !w->deletion_occurred) {
1703                 /* If no images have been modified and no images have been
1704                  * deleted, a new lookup table does not need to be written. */
1705                 wimlib_assert(list_empty(&stream_list));
1706                 old_wim_end = w->hdr.lookup_table_res_entry.offset +
1707                               w->hdr.lookup_table_res_entry.size;
1708                 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
1709                                WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
1710         }
1711
1712         INIT_LIST_HEAD(&stream_list);
1713         for_lookup_table_entry(w->lookup_table, find_new_streams,
1714                                &stream_list);
1715
1716         ret = open_wim_writable(w, w->filename, false,
1717                                 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
1718         if (ret != 0)
1719                 return ret;
1720
1721         if (fseeko(w->out_fp, old_wim_end, SEEK_SET) != 0) {
1722                 ERROR_WITH_ERRNO("Can't seek to end of WIM");
1723                 return WIMLIB_ERR_WRITE;
1724         }
1725
1726         if (!list_empty(&stream_list)) {
1727                 DEBUG("Writing newly added streams (offset = %"PRIu64")",
1728                       old_wim_end);
1729                 ret = write_stream_list(&stream_list, w->out_fp,
1730                                         wimlib_get_compression_type(w),
1731                                         write_flags, num_threads,
1732                                         progress_func);
1733                 if (ret != 0)
1734                         goto out_ftruncate;
1735         } else {
1736                 DEBUG("No new streams were added");
1737         }
1738
1739         for (int i = modified_image_idx; i < w->hdr.image_count; i++) {
1740                 select_wim_image(w, i + 1);
1741                 ret = write_metadata_resource(w);
1742                 if (ret != 0)
1743                         goto out_ftruncate;
1744         }
1745         write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
1746         ret = finish_write(w, WIMLIB_ALL_IMAGES, write_flags,
1747                            progress_func);
1748 out_ftruncate:
1749         close_wim_writable(w);
1750         if (ret != 0) {
1751                 WARNING("Truncating `%s' to its original size (%"PRIu64" bytes)",
1752                         w->filename, old_wim_end);
1753                 truncate(w->filename, old_wim_end);
1754         }
1755         return ret;
1756 }
1757
1758 static int overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
1759                                      unsigned num_threads,
1760                                      wimlib_progress_func_t progress_func)
1761 {
1762         size_t wim_name_len;
1763         int ret;
1764
1765         DEBUG("Overwriting `%s' via a temporary file", w->filename);
1766
1767         /* Write the WIM to a temporary file in the same directory as the
1768          * original WIM. */
1769         wim_name_len = strlen(w->filename);
1770         char tmpfile[wim_name_len + 10];
1771         memcpy(tmpfile, w->filename, wim_name_len);
1772         randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
1773         tmpfile[wim_name_len + 9] = '\0';
1774
1775         ret = wimlib_write(w, tmpfile, WIMLIB_ALL_IMAGES,
1776                            write_flags | WIMLIB_WRITE_FLAG_FSYNC,
1777                            num_threads, progress_func);
1778         if (ret != 0) {
1779                 ERROR("Failed to write the WIM file `%s'", tmpfile);
1780                 goto err;
1781         }
1782
1783         /* Close the original WIM file that was opened for reading. */
1784         if (w->fp != NULL) {
1785                 fclose(w->fp);
1786                 w->fp = NULL;
1787         }
1788
1789         DEBUG("Renaming `%s' to `%s'", tmpfile, w->filename);
1790
1791         /* Rename the new file to the old file .*/
1792         if (rename(tmpfile, w->filename) != 0) {
1793                 ERROR_WITH_ERRNO("Failed to rename `%s' to `%s'",
1794                                  tmpfile, w->filename);
1795                 ret = WIMLIB_ERR_RENAME;
1796                 goto err;
1797         }
1798
1799         if (progress_func) {
1800                 union wimlib_progress_info progress;
1801                 progress.rename.from = tmpfile;
1802                 progress.rename.to = w->filename;
1803                 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
1804         }
1805
1806         /* Re-open the WIM read-only. */
1807         w->fp = fopen(w->filename, "rb");
1808         if (w->fp == NULL) {
1809                 ret = WIMLIB_ERR_REOPEN;
1810                 WARNING("Failed to re-open `%s' read-only: %s",
1811                         w->filename, strerror(errno));
1812         }
1813         return ret;
1814 err:
1815         /* Remove temporary file. */
1816         if (unlink(tmpfile) != 0)
1817                 WARNING("Failed to remove `%s': %s", tmpfile, strerror(errno));
1818         return ret;
1819 }
1820
1821 /*
1822  * Writes a WIM file to the original file that it was read from, overwriting it.
1823  */
1824 WIMLIBAPI int wimlib_overwrite(WIMStruct *w, int write_flags,
1825                                unsigned num_threads,
1826                                wimlib_progress_func_t progress_func)
1827 {
1828         if (!w)
1829                 return WIMLIB_ERR_INVALID_PARAM;
1830
1831         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1832
1833         if (!w->filename)
1834                 return WIMLIB_ERR_NO_FILENAME;
1835
1836         if (w->hdr.total_parts != 1) {
1837                 ERROR("Cannot modify a split WIM");
1838                 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1839         }
1840
1841         if ((!w->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
1842             && !(write_flags & WIMLIB_WRITE_FLAG_REBUILD))
1843         {
1844                 int i, modified_image_idx;
1845                 for (i = 0; i < w->hdr.image_count && !w->image_metadata[i].modified; i++)
1846                         ;
1847                 modified_image_idx = i;
1848                 for (; i < w->hdr.image_count && w->image_metadata[i].modified &&
1849                         !w->image_metadata[i].has_been_mounted_rw; i++)
1850                         ;
1851                 if (i == w->hdr.image_count) {
1852                         return overwrite_wim_inplace(w, write_flags, num_threads,
1853                                                      progress_func,
1854                                                      modified_image_idx);
1855                 }
1856         }
1857         return overwrite_wim_via_tmpfile(w, write_flags, num_threads,
1858                                          progress_func);
1859 }