Fix various typos
[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, 2014, 2015 Eric Biggers
10  *
11  * This file is free software; you can redistribute it and/or modify it under
12  * the terms of the GNU Lesser General Public License as published by the Free
13  * Software Foundation; either version 3 of the License, or (at your option) any
14  * later version.
15  *
16  * This file is distributed in the hope that it will be useful, but WITHOUT
17  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
18  * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
19  * details.
20  *
21  * You should have received a copy of the GNU Lesser General Public License
22  * along with this file; if not, see http://www.gnu.org/licenses/.
23  */
24
25 #ifdef HAVE_CONFIG_H
26 #  include "config.h"
27 #endif
28
29 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
30 /* On BSD, this should be included before "wimlib/list.h" so that "wimlib/list.h" can
31  * override the LIST_HEAD macro. */
32 #  include <sys/file.h>
33 #endif
34
35 #include <errno.h>
36 #include <fcntl.h>
37 #include <stdlib.h>
38 #include <unistd.h>
39
40 #include "wimlib/alloca.h"
41 #include "wimlib/assert.h"
42 #include "wimlib/blob_table.h"
43 #include "wimlib/chunk_compressor.h"
44 #include "wimlib/endianness.h"
45 #include "wimlib/error.h"
46 #include "wimlib/file_io.h"
47 #include "wimlib/header.h"
48 #include "wimlib/inode.h"
49 #include "wimlib/integrity.h"
50 #include "wimlib/metadata.h"
51 #include "wimlib/paths.h"
52 #include "wimlib/progress.h"
53 #include "wimlib/resource.h"
54 #include "wimlib/solid.h"
55 #include "wimlib/win32.h" /* win32_rename_replacement() */
56 #include "wimlib/write.h"
57 #include "wimlib/xml.h"
58
59
60 /* wimlib internal flags used when writing resources.  */
61 #define WRITE_RESOURCE_FLAG_RECOMPRESS          0x00000001
62 #define WRITE_RESOURCE_FLAG_PIPABLE             0x00000002
63 #define WRITE_RESOURCE_FLAG_SOLID               0x00000004
64 #define WRITE_RESOURCE_FLAG_SEND_DONE_WITH_FILE 0x00000008
65 #define WRITE_RESOURCE_FLAG_SOLID_SORT          0x00000010
66
67 static int
68 write_flags_to_resource_flags(int write_flags)
69 {
70         int write_resource_flags = 0;
71
72         if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
73                 write_resource_flags |= WRITE_RESOURCE_FLAG_RECOMPRESS;
74
75         if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
76                 write_resource_flags |= WRITE_RESOURCE_FLAG_PIPABLE;
77
78         if (write_flags & WIMLIB_WRITE_FLAG_SOLID)
79                 write_resource_flags |= WRITE_RESOURCE_FLAG_SOLID;
80
81         if (write_flags & WIMLIB_WRITE_FLAG_SEND_DONE_WITH_FILE_MESSAGES)
82                 write_resource_flags |= WRITE_RESOURCE_FLAG_SEND_DONE_WITH_FILE;
83
84         if ((write_flags & (WIMLIB_WRITE_FLAG_SOLID |
85                             WIMLIB_WRITE_FLAG_NO_SOLID_SORT)) ==
86             WIMLIB_WRITE_FLAG_SOLID)
87                 write_resource_flags |= WRITE_RESOURCE_FLAG_SOLID_SORT;
88
89         return write_resource_flags;
90 }
91
92 struct filter_context {
93         int write_flags;
94         WIMStruct *wim;
95 };
96
97 /*
98  * Determine whether the specified blob should be filtered out from the write.
99  *
100  * Return values:
101  *
102  *  < 0 : The blob should be hard-filtered; that is, not included in the output
103  *        WIM file at all.
104  *    0 : The blob should not be filtered out.
105  *  > 0 : The blob should be soft-filtered; that is, it already exists in the
106  *        WIM file and may not need to be written again.
107  */
108 static int
109 blob_filtered(const struct blob_descriptor *blob,
110               const struct filter_context *ctx)
111 {
112         int write_flags;
113         WIMStruct *wim;
114
115         if (ctx == NULL)
116                 return 0;
117
118         write_flags = ctx->write_flags;
119         wim = ctx->wim;
120
121         if (write_flags & WIMLIB_WRITE_FLAG_APPEND &&
122             blob->blob_location == BLOB_IN_WIM &&
123             blob->rdesc->wim == wim)
124                 return 1;
125
126         if (write_flags & WIMLIB_WRITE_FLAG_SKIP_EXTERNAL_WIMS &&
127             blob->blob_location == BLOB_IN_WIM &&
128             blob->rdesc->wim != wim)
129                 return -1;
130
131         return 0;
132 }
133
134 static bool
135 blob_hard_filtered(const struct blob_descriptor *blob,
136                    struct filter_context *ctx)
137 {
138         return blob_filtered(blob, ctx) < 0;
139 }
140
141 static inline bool
142 may_soft_filter_blobs(const struct filter_context *ctx)
143 {
144         return ctx && (ctx->write_flags & WIMLIB_WRITE_FLAG_APPEND);
145 }
146
147 static inline bool
148 may_hard_filter_blobs(const struct filter_context *ctx)
149 {
150         return ctx && (ctx->write_flags & WIMLIB_WRITE_FLAG_SKIP_EXTERNAL_WIMS);
151 }
152
153 static inline bool
154 may_filter_blobs(const struct filter_context *ctx)
155 {
156         return (may_soft_filter_blobs(ctx) || may_hard_filter_blobs(ctx));
157 }
158
159 /* Return true if the specified blob is located in a WIM resource which can be
160  * reused in the output WIM file, without being recompressed.  */
161 static bool
162 can_raw_copy(const struct blob_descriptor *blob, int write_resource_flags,
163              int out_ctype, u32 out_chunk_size)
164 {
165         const struct wim_resource_descriptor *rdesc;
166
167         /* Recompress everything if requested.  */
168         if (write_resource_flags & WRITE_RESOURCE_FLAG_RECOMPRESS)
169                 return false;
170
171         /* A blob not located in a WIM resource cannot be reused.  */
172         if (blob->blob_location != BLOB_IN_WIM)
173                 return false;
174
175         rdesc = blob->rdesc;
176
177         /* In the case of an in-place compaction, always reuse resources located
178          * in the WIM being compacted.  */
179         if (rdesc->wim->being_compacted)
180                 return true;
181
182         /* Otherwise, only reuse compressed resources.  */
183         if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
184             !(rdesc->flags & (WIM_RESHDR_FLAG_COMPRESSED |
185                               WIM_RESHDR_FLAG_SOLID)))
186                 return false;
187
188         /* When writing a pipable WIM, we can only reuse pipable resources; and
189          * when writing a non-pipable WIM, we can only reuse non-pipable
190          * resources.  */
191         if (rdesc->is_pipable !=
192             !!(write_resource_flags & WRITE_RESOURCE_FLAG_PIPABLE))
193                 return false;
194
195         /* When writing a solid WIM, we can only reuse solid resources; and when
196          * writing a non-solid WIM, we can only reuse non-solid resources.  */
197         if (!!(rdesc->flags & WIM_RESHDR_FLAG_SOLID) !=
198             !!(write_resource_flags & WRITE_RESOURCE_FLAG_SOLID))
199                 return false;
200
201         /* Note: it is theoretically possible to copy chunks of compressed data
202          * between non-solid, solid, and pipable resources.  However, we don't
203          * currently implement this optimization because it would be complex and
204          * would usually go unused.  */
205
206         if (rdesc->flags & WIM_RESHDR_FLAG_COMPRESSED) {
207                 /* To re-use a non-solid resource, it must use the desired
208                  * compression type and chunk size.  */
209                 return (rdesc->compression_type == out_ctype &&
210                         rdesc->chunk_size == out_chunk_size);
211         } else {
212                 /* Solid resource: Such resources may contain multiple blobs,
213                  * and in general only a subset of them need to be written.  As
214                  * a heuristic, re-use the raw data if more than two-thirds the
215                  * uncompressed size is being written.  */
216
217                 /* Note: solid resources contain a header that specifies the
218                  * compression type and chunk size; therefore we don't need to
219                  * check if they are compatible with @out_ctype and
220                  * @out_chunk_size.  */
221
222                 /* Did we already decide to reuse the resource?  */
223                 if (rdesc->raw_copy_ok)
224                         return true;
225
226                 struct blob_descriptor *res_blob;
227                 u64 write_size = 0;
228
229                 list_for_each_entry(res_blob, &rdesc->blob_list, rdesc_node)
230                         if (res_blob->will_be_in_output_wim)
231                                 write_size += res_blob->size;
232
233                 return (write_size > rdesc->uncompressed_size * 2 / 3);
234         }
235 }
236
237 static u32
238 reshdr_flags_for_blob(const struct blob_descriptor *blob)
239 {
240         u32 reshdr_flags = 0;
241         if (blob->is_metadata)
242                 reshdr_flags |= WIM_RESHDR_FLAG_METADATA;
243         return reshdr_flags;
244 }
245
246 static void
247 blob_set_out_reshdr_for_reuse(struct blob_descriptor *blob)
248 {
249         const struct wim_resource_descriptor *rdesc;
250
251         wimlib_assert(blob->blob_location == BLOB_IN_WIM);
252         rdesc = blob->rdesc;
253
254         if (rdesc->flags & WIM_RESHDR_FLAG_SOLID) {
255                 blob->out_reshdr.offset_in_wim = blob->offset_in_res;
256                 blob->out_reshdr.uncompressed_size = 0;
257                 blob->out_reshdr.size_in_wim = blob->size;
258
259                 blob->out_res_offset_in_wim = rdesc->offset_in_wim;
260                 blob->out_res_size_in_wim = rdesc->size_in_wim;
261                 blob->out_res_uncompressed_size = rdesc->uncompressed_size;
262         } else {
263                 blob->out_reshdr.offset_in_wim = rdesc->offset_in_wim;
264                 blob->out_reshdr.uncompressed_size = rdesc->uncompressed_size;
265                 blob->out_reshdr.size_in_wim = rdesc->size_in_wim;
266         }
267         blob->out_reshdr.flags = rdesc->flags;
268 }
269
270
271 /* Write the header for a blob in a pipable WIM.  */
272 static int
273 write_pwm_blob_header(const struct blob_descriptor *blob,
274                       struct filedes *out_fd, bool compressed)
275 {
276         struct pwm_blob_hdr blob_hdr;
277         u32 reshdr_flags;
278         int ret;
279
280         wimlib_assert(!blob->unhashed);
281
282         blob_hdr.magic = cpu_to_le64(PWM_BLOB_MAGIC);
283         blob_hdr.uncompressed_size = cpu_to_le64(blob->size);
284         copy_hash(blob_hdr.hash, blob->hash);
285         reshdr_flags = reshdr_flags_for_blob(blob);
286         if (compressed)
287                 reshdr_flags |= WIM_RESHDR_FLAG_COMPRESSED;
288         blob_hdr.flags = cpu_to_le32(reshdr_flags);
289         ret = full_write(out_fd, &blob_hdr, sizeof(blob_hdr));
290         if (ret)
291                 ERROR_WITH_ERRNO("Write error");
292         return ret;
293 }
294
295 struct write_blobs_progress_data {
296         wimlib_progress_func_t progfunc;
297         void *progctx;
298         union wimlib_progress_info progress;
299         u64 next_progress;
300 };
301
302 static int
303 do_write_blobs_progress(struct write_blobs_progress_data *progress_data,
304                         u64 complete_size, u32 complete_count, bool discarded)
305 {
306         union wimlib_progress_info *progress = &progress_data->progress;
307         int ret;
308
309         if (discarded) {
310                 progress->write_streams.total_bytes -= complete_size;
311                 progress->write_streams.total_streams -= complete_count;
312                 if (progress_data->next_progress != ~(u64)0 &&
313                     progress_data->next_progress > progress->write_streams.total_bytes)
314                 {
315                         progress_data->next_progress = progress->write_streams.total_bytes;
316                 }
317         } else {
318                 progress->write_streams.completed_bytes += complete_size;
319                 progress->write_streams.completed_streams += complete_count;
320         }
321
322         if (progress->write_streams.completed_bytes >= progress_data->next_progress) {
323
324                 ret = call_progress(progress_data->progfunc,
325                                     WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
326                                     progress,
327                                     progress_data->progctx);
328                 if (ret)
329                         return ret;
330
331                 set_next_progress(progress->write_streams.completed_bytes,
332                                   progress->write_streams.total_bytes,
333                                   &progress_data->next_progress);
334         }
335         return 0;
336 }
337
338 struct write_blobs_ctx {
339         /* File descriptor to which the blobs are being written.  */
340         struct filedes *out_fd;
341
342         /* Blob table for the WIMStruct on whose behalf the blobs are being
343          * written.  */
344         struct blob_table *blob_table;
345
346         /* Compression format to use.  */
347         int out_ctype;
348
349         /* Maximum uncompressed chunk size in compressed resources to use.  */
350         u32 out_chunk_size;
351
352         /* Flags that affect how the blobs will be written.  */
353         int write_resource_flags;
354
355         /* Data used for issuing WRITE_STREAMS progress.  */
356         struct write_blobs_progress_data progress_data;
357
358         struct filter_context *filter_ctx;
359
360         /* Pointer to the chunk_compressor implementation being used for
361          * compressing chunks of data, or NULL if chunks are being written
362          * uncompressed.  */
363         struct chunk_compressor *compressor;
364
365         /* A buffer of size @out_chunk_size that has been loaned out from the
366          * chunk compressor and is currently being filled with the uncompressed
367          * data of the next chunk.  */
368         u8 *cur_chunk_buf;
369
370         /* Number of bytes in @cur_chunk_buf that are currently filled.  */
371         size_t cur_chunk_buf_filled;
372
373         /* List of blobs that currently have chunks being compressed.  */
374         struct list_head blobs_being_compressed;
375
376         /* List of blobs in the solid resource.  Blobs are moved here after
377          * @blobs_being_compressed only when writing a solid resource.  */
378         struct list_head blobs_in_solid_resource;
379
380         /* Current uncompressed offset in the blob being read.  */
381         u64 cur_read_blob_offset;
382
383         /* Uncompressed size of the blob currently being read.  */
384         u64 cur_read_blob_size;
385
386         /* Current uncompressed offset in the blob being written.  */
387         u64 cur_write_blob_offset;
388
389         /* Uncompressed size of resource currently being written.  */
390         u64 cur_write_res_size;
391
392         /* Array that is filled in with compressed chunk sizes as a resource is
393          * being written.  */
394         u64 *chunk_csizes;
395
396         /* Index of next entry in @chunk_csizes to fill in.  */
397         size_t chunk_index;
398
399         /* Number of entries in @chunk_csizes currently allocated.  */
400         size_t num_alloc_chunks;
401
402         /* Offset in the output file of the start of the chunks of the resource
403          * currently being written.  */
404         u64 chunks_start_offset;
405 };
406
407 /* Reserve space for the chunk table and prepare to accumulate the chunk table
408  * in memory.  */
409 static int
410 begin_chunk_table(struct write_blobs_ctx *ctx, u64 res_expected_size)
411 {
412         u64 expected_num_chunks;
413         u64 expected_num_chunk_entries;
414         size_t reserve_size;
415         int ret;
416
417         /* Calculate the number of chunks and chunk entries that should be
418          * needed for the resource.  These normally will be the final values,
419          * but in SOLID mode some of the blobs we're planning to write into the
420          * resource may be duplicates, and therefore discarded, potentially
421          * decreasing the number of chunk entries needed.  */
422         expected_num_chunks = DIV_ROUND_UP(res_expected_size, ctx->out_chunk_size);
423         expected_num_chunk_entries = expected_num_chunks;
424         if (!(ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID))
425                 expected_num_chunk_entries--;
426
427         /* Make sure the chunk_csizes array is long enough to store the
428          * compressed size of each chunk.  */
429         if (expected_num_chunks > ctx->num_alloc_chunks) {
430                 u64 new_length = expected_num_chunks + 50;
431
432                 if ((size_t)new_length != new_length) {
433                         ERROR("Resource size too large (%"PRIu64" bytes!",
434                               res_expected_size);
435                         return WIMLIB_ERR_NOMEM;
436                 }
437
438                 FREE(ctx->chunk_csizes);
439                 ctx->chunk_csizes = MALLOC(new_length * sizeof(ctx->chunk_csizes[0]));
440                 if (ctx->chunk_csizes == NULL) {
441                         ctx->num_alloc_chunks = 0;
442                         return WIMLIB_ERR_NOMEM;
443                 }
444                 ctx->num_alloc_chunks = new_length;
445         }
446
447         ctx->chunk_index = 0;
448
449         if (!(ctx->write_resource_flags & WRITE_RESOURCE_FLAG_PIPABLE)) {
450                 /* Reserve space for the chunk table in the output file.  In the
451                  * case of solid resources this reserves the upper bound for the
452                  * needed space, not necessarily the exact space which will
453                  * prove to be needed.  At this point, we just use @chunk_csizes
454                  * for a buffer of 0's because the actual compressed chunk sizes
455                  * are unknown.  */
456                 reserve_size = expected_num_chunk_entries *
457                                get_chunk_entry_size(res_expected_size,
458                                                     0 != (ctx->write_resource_flags &
459                                                           WRITE_RESOURCE_FLAG_SOLID));
460                 if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID)
461                         reserve_size += sizeof(struct alt_chunk_table_header_disk);
462                 memset(ctx->chunk_csizes, 0, reserve_size);
463                 ret = full_write(ctx->out_fd, ctx->chunk_csizes, reserve_size);
464                 if (ret)
465                         return ret;
466         }
467         return 0;
468 }
469
470 static int
471 begin_write_resource(struct write_blobs_ctx *ctx, u64 res_expected_size)
472 {
473         int ret;
474
475         wimlib_assert(res_expected_size != 0);
476
477         if (ctx->compressor != NULL) {
478                 ret = begin_chunk_table(ctx, res_expected_size);
479                 if (ret)
480                         return ret;
481         }
482
483         /* Output file descriptor is now positioned at the offset at which to
484          * write the first chunk of the resource.  */
485         ctx->chunks_start_offset = ctx->out_fd->offset;
486         ctx->cur_write_blob_offset = 0;
487         ctx->cur_write_res_size = res_expected_size;
488         return 0;
489 }
490
491 static int
492 end_chunk_table(struct write_blobs_ctx *ctx, u64 res_actual_size,
493                 u64 *res_start_offset_ret, u64 *res_store_size_ret)
494 {
495         size_t actual_num_chunks;
496         size_t actual_num_chunk_entries;
497         size_t chunk_entry_size;
498         int ret;
499
500         actual_num_chunks = ctx->chunk_index;
501         actual_num_chunk_entries = actual_num_chunks;
502         if (!(ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID))
503                 actual_num_chunk_entries--;
504
505         chunk_entry_size = get_chunk_entry_size(res_actual_size,
506                                                 0 != (ctx->write_resource_flags &
507                                                       WRITE_RESOURCE_FLAG_SOLID));
508
509         typedef le64 _may_alias_attribute aliased_le64_t;
510         typedef le32 _may_alias_attribute aliased_le32_t;
511
512         if (chunk_entry_size == 4) {
513                 aliased_le32_t *entries = (aliased_le32_t*)ctx->chunk_csizes;
514
515                 if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID) {
516                         for (size_t i = 0; i < actual_num_chunk_entries; i++)
517                                 entries[i] = cpu_to_le32(ctx->chunk_csizes[i]);
518                 } else {
519                         u32 offset = ctx->chunk_csizes[0];
520                         for (size_t i = 0; i < actual_num_chunk_entries; i++) {
521                                 u32 next_size = ctx->chunk_csizes[i + 1];
522                                 entries[i] = cpu_to_le32(offset);
523                                 offset += next_size;
524                         }
525                 }
526         } else {
527                 aliased_le64_t *entries = (aliased_le64_t*)ctx->chunk_csizes;
528
529                 if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID) {
530                         for (size_t i = 0; i < actual_num_chunk_entries; i++)
531                                 entries[i] = cpu_to_le64(ctx->chunk_csizes[i]);
532                 } else {
533                         u64 offset = ctx->chunk_csizes[0];
534                         for (size_t i = 0; i < actual_num_chunk_entries; i++) {
535                                 u64 next_size = ctx->chunk_csizes[i + 1];
536                                 entries[i] = cpu_to_le64(offset);
537                                 offset += next_size;
538                         }
539                 }
540         }
541
542         size_t chunk_table_size = actual_num_chunk_entries * chunk_entry_size;
543         u64 res_start_offset;
544         u64 res_end_offset;
545
546         if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_PIPABLE) {
547                 ret = full_write(ctx->out_fd, ctx->chunk_csizes, chunk_table_size);
548                 if (ret)
549                         goto write_error;
550                 res_end_offset = ctx->out_fd->offset;
551                 res_start_offset = ctx->chunks_start_offset;
552         } else {
553                 res_end_offset = ctx->out_fd->offset;
554
555                 u64 chunk_table_offset;
556
557                 chunk_table_offset = ctx->chunks_start_offset - chunk_table_size;
558
559                 if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID) {
560                         struct alt_chunk_table_header_disk hdr;
561
562                         hdr.res_usize = cpu_to_le64(res_actual_size);
563                         hdr.chunk_size = cpu_to_le32(ctx->out_chunk_size);
564                         hdr.compression_format = cpu_to_le32(ctx->out_ctype);
565
566                         STATIC_ASSERT(WIMLIB_COMPRESSION_TYPE_XPRESS == 1);
567                         STATIC_ASSERT(WIMLIB_COMPRESSION_TYPE_LZX == 2);
568                         STATIC_ASSERT(WIMLIB_COMPRESSION_TYPE_LZMS == 3);
569
570                         ret = full_pwrite(ctx->out_fd, &hdr, sizeof(hdr),
571                                           chunk_table_offset - sizeof(hdr));
572                         if (ret)
573                                 goto write_error;
574                         res_start_offset = chunk_table_offset - sizeof(hdr);
575                 } else {
576                         res_start_offset = chunk_table_offset;
577                 }
578
579                 ret = full_pwrite(ctx->out_fd, ctx->chunk_csizes,
580                                   chunk_table_size, chunk_table_offset);
581                 if (ret)
582                         goto write_error;
583         }
584
585         *res_start_offset_ret = res_start_offset;
586         *res_store_size_ret = res_end_offset - res_start_offset;
587
588         return 0;
589
590 write_error:
591         ERROR_WITH_ERRNO("Write error");
592         return ret;
593 }
594
595 /* Finish writing a WIM resource by writing or updating the chunk table (if not
596  * writing the data uncompressed) and loading its metadata into @out_reshdr.  */
597 static int
598 end_write_resource(struct write_blobs_ctx *ctx, struct wim_reshdr *out_reshdr)
599 {
600         int ret;
601         u64 res_size_in_wim;
602         u64 res_uncompressed_size;
603         u64 res_offset_in_wim;
604
605         wimlib_assert(ctx->cur_write_blob_offset == ctx->cur_write_res_size ||
606                       (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID));
607         res_uncompressed_size = ctx->cur_write_res_size;
608
609         if (ctx->compressor) {
610                 ret = end_chunk_table(ctx, res_uncompressed_size,
611                                       &res_offset_in_wim, &res_size_in_wim);
612                 if (ret)
613                         return ret;
614         } else {
615                 res_offset_in_wim = ctx->chunks_start_offset;
616                 res_size_in_wim = ctx->out_fd->offset - res_offset_in_wim;
617         }
618         out_reshdr->uncompressed_size = res_uncompressed_size;
619         out_reshdr->size_in_wim = res_size_in_wim;
620         out_reshdr->offset_in_wim = res_offset_in_wim;
621         return 0;
622 }
623
624 /* Call when no more data from the file at @path is needed.  */
625 static int
626 done_with_file(const tchar *path, wimlib_progress_func_t progfunc, void *progctx)
627 {
628         union wimlib_progress_info info;
629
630         info.done_with_file.path_to_file = path;
631
632         return call_progress(progfunc, WIMLIB_PROGRESS_MSG_DONE_WITH_FILE,
633                              &info, progctx);
634 }
635
636 static int
637 do_done_with_blob(struct blob_descriptor *blob,
638                   wimlib_progress_func_t progfunc, void *progctx)
639 {
640         int ret;
641         struct wim_inode *inode;
642         tchar *cookie1;
643         tchar *cookie2;
644
645         if (!blob->may_send_done_with_file)
646                 return 0;
647
648         inode = blob->file_inode;
649
650         wimlib_assert(inode != NULL);
651         wimlib_assert(inode->i_num_remaining_streams > 0);
652         if (--inode->i_num_remaining_streams > 0)
653                 return 0;
654
655         cookie1 = progress_get_streamless_path(blob->file_on_disk);
656         cookie2 = progress_get_win32_path(blob->file_on_disk);
657
658         ret = done_with_file(blob->file_on_disk, progfunc, progctx);
659
660         progress_put_win32_path(cookie2);
661         progress_put_streamless_path(cookie1);
662
663         return ret;
664 }
665
666 /* Handle WIMLIB_WRITE_FLAG_SEND_DONE_WITH_FILE_MESSAGES mode.  */
667 static inline int
668 done_with_blob(struct blob_descriptor *blob, struct write_blobs_ctx *ctx)
669 {
670         if (likely(!(ctx->write_resource_flags &
671                      WRITE_RESOURCE_FLAG_SEND_DONE_WITH_FILE)))
672                 return 0;
673         return do_done_with_blob(blob, ctx->progress_data.progfunc,
674                                  ctx->progress_data.progctx);
675 }
676
677 /* Begin processing a blob for writing.  */
678 static int
679 write_blob_begin_read(struct blob_descriptor *blob, void *_ctx)
680 {
681         struct write_blobs_ctx *ctx = _ctx;
682         int ret;
683
684         wimlib_assert(blob->size > 0);
685
686         ctx->cur_read_blob_offset = 0;
687         ctx->cur_read_blob_size = blob->size;
688
689         /* As an optimization, we allow some blobs to be "unhashed", meaning
690          * their SHA-1 message digests are unknown.  This is the case with blobs
691          * that are added by scanning a directory tree with wimlib_add_image(),
692          * for example.  Since WIM uses single-instance blobs, we don't know
693          * whether such each such blob really need to written until it is
694          * actually checksummed, unless it has a unique size.  In such cases we
695          * read and checksum the blob in this function, thereby advancing ahead
696          * of read_blob_list(), which will still provide the data again to
697          * write_blob_process_chunk().  This is okay because an unhashed blob
698          * cannot be in a WIM resource, which might be costly to decompress.  */
699         if (ctx->blob_table != NULL && blob->unhashed && !blob->unique_size) {
700
701                 struct blob_descriptor *new_blob;
702
703                 ret = hash_unhashed_blob(blob, ctx->blob_table, &new_blob);
704                 if (ret)
705                         return ret;
706                 if (new_blob != blob) {
707                         /* Duplicate blob detected.  */
708
709                         if (new_blob->will_be_in_output_wim ||
710                             blob_filtered(new_blob, ctx->filter_ctx))
711                         {
712                                 /* The duplicate blob is already being included
713                                  * in the output WIM, or it would be filtered
714                                  * out if it had been.  Skip writing this blob
715                                  * (and reading it again) entirely, passing its
716                                  * output reference count to the duplicate blob
717                                  * in the former case.  */
718                                 ret = do_write_blobs_progress(&ctx->progress_data,
719                                                               blob->size, 1, true);
720                                 list_del(&blob->write_blobs_list);
721                                 list_del(&blob->blob_table_list);
722                                 if (new_blob->will_be_in_output_wim)
723                                         new_blob->out_refcnt += blob->out_refcnt;
724                                 if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID)
725                                         ctx->cur_write_res_size -= blob->size;
726                                 if (!ret)
727                                         ret = done_with_blob(blob, ctx);
728                                 free_blob_descriptor(blob);
729                                 if (ret)
730                                         return ret;
731                                 return BEGIN_BLOB_STATUS_SKIP_BLOB;
732                         } else {
733                                 /* The duplicate blob can validly be written,
734                                  * but was not marked as such.  Discard the
735                                  * current blob descriptor and use the
736                                  * duplicate, but actually freeing the current
737                                  * blob descriptor must wait until
738                                  * read_blob_list() has finished reading its
739                                  * data.  */
740                                 list_replace(&blob->write_blobs_list,
741                                              &new_blob->write_blobs_list);
742                                 list_replace(&blob->blob_table_list,
743                                              &new_blob->blob_table_list);
744                                 blob->will_be_in_output_wim = 0;
745                                 new_blob->out_refcnt = blob->out_refcnt;
746                                 new_blob->will_be_in_output_wim = 1;
747                                 new_blob->may_send_done_with_file = 0;
748                                 blob = new_blob;
749                         }
750                 }
751         }
752         list_move_tail(&blob->write_blobs_list, &ctx->blobs_being_compressed);
753         return 0;
754 }
755
756 /* Rewrite a blob that was just written compressed (as a non-solid WIM resource)
757  * as uncompressed instead.  */
758 static int
759 write_blob_uncompressed(struct blob_descriptor *blob, struct filedes *out_fd)
760 {
761         int ret;
762         u64 begin_offset = blob->out_reshdr.offset_in_wim;
763         u64 end_offset = out_fd->offset;
764
765         if (filedes_seek(out_fd, begin_offset) == -1)
766                 return 0;
767
768         ret = extract_blob_to_fd(blob, out_fd);
769         if (ret) {
770                 /* Error reading the uncompressed data.  */
771                 if (out_fd->offset == begin_offset &&
772                     filedes_seek(out_fd, end_offset) != -1)
773                 {
774                         /* Nothing was actually written yet, and we successfully
775                          * seeked to the end of the compressed resource, so
776                          * don't issue a hard error; just keep the compressed
777                          * resource instead.  */
778                         WARNING("Recovered compressed resource of "
779                                 "size %"PRIu64", continuing on.", blob->size);
780                         return 0;
781                 }
782                 return ret;
783         }
784
785         wimlib_assert(out_fd->offset - begin_offset == blob->size);
786
787         /* We could ftruncate() the file to 'out_fd->offset' here, but there
788          * isn't much point.  Usually we will only be truncating by a few bytes
789          * and will just overwrite the data immediately.  */
790
791         blob->out_reshdr.size_in_wim = blob->size;
792         blob->out_reshdr.flags &= ~(WIM_RESHDR_FLAG_COMPRESSED |
793                                     WIM_RESHDR_FLAG_SOLID);
794         return 0;
795 }
796
797 /* Returns true if the specified blob, which was written as a non-solid
798  * resource, should be truncated from the WIM file and re-written uncompressed.
799  * blob->out_reshdr must be filled in from the initial write of the blob.  */
800 static bool
801 should_rewrite_blob_uncompressed(const struct write_blobs_ctx *ctx,
802                                  const struct blob_descriptor *blob)
803 {
804         /* If the compressed data is smaller than the uncompressed data, prefer
805          * the compressed data.  */
806         if (blob->out_reshdr.size_in_wim < blob->out_reshdr.uncompressed_size)
807                 return false;
808
809         /* If we're not actually writing compressed data, then there's no need
810          * for re-writing.  */
811         if (!ctx->compressor)
812                 return false;
813
814         /* If writing a pipable WIM, everything we write to the output is final
815          * (it might actually be a pipe!).  */
816         if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_PIPABLE)
817                 return false;
818
819         /* If the blob that would need to be re-read is located in a solid
820          * resource in another WIM file, then re-reading it would be costly.  So
821          * don't do it.
822          *
823          * Exception: if the compressed size happens to be *exactly* the same as
824          * the uncompressed size, then the blob *must* be written uncompressed
825          * in order to remain compatible with the Windows Overlay Filesystem
826          * Filter Driver (WOF).
827          *
828          * TODO: we are currently assuming that the optimization for
829          * single-chunk resources in maybe_rewrite_blob_uncompressed() prevents
830          * this case from being triggered too often.  To fully prevent excessive
831          * decompressions in degenerate cases, we really should obtain the
832          * uncompressed data by decompressing the compressed data we wrote to
833          * the output file.
834          */
835         if (blob->blob_location == BLOB_IN_WIM &&
836             blob->size != blob->rdesc->uncompressed_size &&
837             blob->size != blob->out_reshdr.size_in_wim)
838                 return false;
839
840         return true;
841 }
842
843 static int
844 maybe_rewrite_blob_uncompressed(struct write_blobs_ctx *ctx,
845                                 struct blob_descriptor *blob)
846 {
847         if (!should_rewrite_blob_uncompressed(ctx, blob))
848                 return 0;
849
850         /* Regular (non-solid) WIM resources with exactly one chunk and
851          * compressed size equal to uncompressed size are exactly the same as
852          * the corresponding compressed data --- since there must be 0 entries
853          * in the chunk table and the only chunk must be stored uncompressed.
854          * In this case, there's no need to rewrite anything.  */
855         if (ctx->chunk_index == 1 &&
856             blob->out_reshdr.size_in_wim == blob->out_reshdr.uncompressed_size)
857         {
858                 blob->out_reshdr.flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
859                 return 0;
860         }
861
862         return write_blob_uncompressed(blob, ctx->out_fd);
863 }
864
865 /* Write the next chunk of (typically compressed) data to the output WIM,
866  * handling the writing of the chunk table.  */
867 static int
868 write_chunk(struct write_blobs_ctx *ctx, const void *cchunk,
869             size_t csize, size_t usize)
870 {
871         int ret;
872         struct blob_descriptor *blob;
873         u32 completed_blob_count;
874         u32 completed_size;
875
876         blob = list_entry(ctx->blobs_being_compressed.next,
877                           struct blob_descriptor, write_blobs_list);
878
879         if (ctx->cur_write_blob_offset == 0 &&
880             !(ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID))
881         {
882                 /* Starting to write a new blob in non-solid mode.  */
883
884                 if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_PIPABLE) {
885                         ret = write_pwm_blob_header(blob, ctx->out_fd,
886                                                     ctx->compressor != NULL);
887                         if (ret)
888                                 return ret;
889                 }
890
891                 ret = begin_write_resource(ctx, blob->size);
892                 if (ret)
893                         return ret;
894         }
895
896         if (ctx->compressor != NULL) {
897                 /* Record the compresed chunk size.  */
898                 wimlib_assert(ctx->chunk_index < ctx->num_alloc_chunks);
899                 ctx->chunk_csizes[ctx->chunk_index++] = csize;
900
901                /* If writing a pipable WIM, before the chunk data write a chunk
902                 * header that provides the compressed chunk size.  */
903                 if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_PIPABLE) {
904                         struct pwm_chunk_hdr chunk_hdr = {
905                                 .compressed_size = cpu_to_le32(csize),
906                         };
907                         ret = full_write(ctx->out_fd, &chunk_hdr,
908                                          sizeof(chunk_hdr));
909                         if (ret)
910                                 goto write_error;
911                 }
912         }
913
914         /* Write the chunk data.  */
915         ret = full_write(ctx->out_fd, cchunk, csize);
916         if (ret)
917                 goto write_error;
918
919         ctx->cur_write_blob_offset += usize;
920
921         completed_size = usize;
922         completed_blob_count = 0;
923         if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID) {
924                 /* Wrote chunk in solid mode.  It may have finished multiple
925                  * blobs.  */
926                 struct blob_descriptor *next_blob;
927
928                 while (blob && ctx->cur_write_blob_offset >= blob->size) {
929
930                         ctx->cur_write_blob_offset -= blob->size;
931
932                         if (ctx->cur_write_blob_offset)
933                                 next_blob = list_entry(blob->write_blobs_list.next,
934                                                       struct blob_descriptor,
935                                                       write_blobs_list);
936                         else
937                                 next_blob = NULL;
938
939                         ret = done_with_blob(blob, ctx);
940                         if (ret)
941                                 return ret;
942                         list_move_tail(&blob->write_blobs_list, &ctx->blobs_in_solid_resource);
943                         completed_blob_count++;
944
945                         blob = next_blob;
946                 }
947         } else {
948                 /* Wrote chunk in non-solid mode.  It may have finished a
949                  * blob.  */
950                 if (ctx->cur_write_blob_offset == blob->size) {
951
952                         wimlib_assert(ctx->cur_write_blob_offset ==
953                                       ctx->cur_write_res_size);
954
955                         ret = end_write_resource(ctx, &blob->out_reshdr);
956                         if (ret)
957                                 return ret;
958
959                         blob->out_reshdr.flags = reshdr_flags_for_blob(blob);
960                         if (ctx->compressor != NULL)
961                                 blob->out_reshdr.flags |= WIM_RESHDR_FLAG_COMPRESSED;
962
963                         ret = maybe_rewrite_blob_uncompressed(ctx, blob);
964                         if (ret)
965                                 return ret;
966
967                         wimlib_assert(blob->out_reshdr.uncompressed_size == blob->size);
968
969                         ctx->cur_write_blob_offset = 0;
970
971                         ret = done_with_blob(blob, ctx);
972                         if (ret)
973                                 return ret;
974                         list_del(&blob->write_blobs_list);
975                         completed_blob_count++;
976                 }
977         }
978
979         return do_write_blobs_progress(&ctx->progress_data, completed_size,
980                                        completed_blob_count, false);
981
982 write_error:
983         ERROR_WITH_ERRNO("Write error");
984         return ret;
985 }
986
987 static int
988 prepare_chunk_buffer(struct write_blobs_ctx *ctx)
989 {
990         /* While we are unable to get a new chunk buffer due to too many chunks
991          * already outstanding, retrieve and write the next compressed chunk. */
992         while (!(ctx->cur_chunk_buf =
993                  ctx->compressor->get_chunk_buffer(ctx->compressor)))
994         {
995                 const void *cchunk;
996                 u32 csize;
997                 u32 usize;
998                 bool bret;
999                 int ret;
1000
1001                 bret = ctx->compressor->get_compression_result(ctx->compressor,
1002                                                                &cchunk,
1003                                                                &csize,
1004                                                                &usize);
1005                 wimlib_assert(bret);
1006
1007                 ret = write_chunk(ctx, cchunk, csize, usize);
1008                 if (ret)
1009                         return ret;
1010         }
1011         return 0;
1012 }
1013
1014 /* Process the next chunk of data to be written to a WIM resource.  */
1015 static int
1016 write_blob_process_chunk(const void *chunk, size_t size, void *_ctx)
1017 {
1018         struct write_blobs_ctx *ctx = _ctx;
1019         int ret;
1020         const u8 *chunkptr, *chunkend;
1021
1022         wimlib_assert(size != 0);
1023
1024         if (ctx->compressor == NULL) {
1025                 /* Write chunk uncompressed.  */
1026                  ret = write_chunk(ctx, chunk, size, size);
1027                  if (ret)
1028                          return ret;
1029                  ctx->cur_read_blob_offset += size;
1030                  return 0;
1031         }
1032
1033         /* Submit the chunk for compression, but take into account that the
1034          * @size the chunk was provided in may not correspond to the
1035          * @out_chunk_size being used for compression.  */
1036         chunkptr = chunk;
1037         chunkend = chunkptr + size;
1038         do {
1039                 size_t needed_chunk_size;
1040                 size_t bytes_consumed;
1041
1042                 if (!ctx->cur_chunk_buf) {
1043                         ret = prepare_chunk_buffer(ctx);
1044                         if (ret)
1045                                 return ret;
1046                 }
1047
1048                 if (ctx->write_resource_flags & WRITE_RESOURCE_FLAG_SOLID) {
1049                         needed_chunk_size = ctx->out_chunk_size;
1050                 } else {
1051                         needed_chunk_size = min(ctx->out_chunk_size,
1052                                                 ctx->cur_chunk_buf_filled +
1053                                                         (ctx->cur_read_blob_size -
1054                                                          ctx->cur_read_blob_offset));
1055                 }
1056
1057                 bytes_consumed = min(chunkend - chunkptr,
1058                                      needed_chunk_size - ctx->cur_chunk_buf_filled);
1059
1060                 memcpy(&ctx->cur_chunk_buf[ctx->cur_chunk_buf_filled],
1061                        chunkptr, bytes_consumed);
1062
1063                 chunkptr += bytes_consumed;
1064                 ctx->cur_read_blob_offset += bytes_consumed;
1065                 ctx->cur_chunk_buf_filled += bytes_consumed;
1066
1067                 if (ctx->cur_chunk_buf_filled == needed_chunk_size) {
1068                         ctx->compressor->signal_chunk_filled(ctx->compressor,
1069                                                              ctx->cur_chunk_buf_filled);
1070                         ctx->cur_chunk_buf = NULL;
1071                         ctx->cur_chunk_buf_filled = 0;
1072                 }
1073         } while (chunkptr != chunkend);
1074         return 0;
1075 }
1076
1077 /* Finish processing a blob for writing.  It may not have been completely
1078  * written yet, as the chunk_compressor implementation may still have chunks
1079  * buffered or being compressed.  */
1080 static int
1081 write_blob_end_read(struct blob_descriptor *blob, int status, void *_ctx)
1082 {
1083         struct write_blobs_ctx *ctx = _ctx;
1084
1085         wimlib_assert(ctx->cur_read_blob_offset == ctx->cur_read_blob_size || status);
1086
1087         if (!blob->will_be_in_output_wim) {
1088                 /* The blob was a duplicate.  Now that its data has finished
1089                  * being read, it is being discarded in favor of the duplicate
1090                  * entry.  It therefore is no longer needed, and we can fire the
1091                  * DONE_WITH_FILE callback because the file will not be read
1092                  * again.
1093                  *
1094                  * Note: we can't yet fire DONE_WITH_FILE for non-duplicate
1095                  * blobs, since it needs to be possible to re-read the file if
1096                  * it does not compress to less than its original size.  */
1097                 if (!status)
1098                         status = done_with_blob(blob, ctx);
1099                 free_blob_descriptor(blob);
1100         } else if (!status && blob->unhashed && ctx->blob_table != NULL) {
1101                 /* The blob was not a duplicate and was previously unhashed.
1102                  * Since we passed COMPUTE_MISSING_BLOB_HASHES to
1103                  * read_blob_list(), blob->hash is now computed and valid.  So
1104                  * turn this blob into a "hashed" blob.  */
1105                 list_del(&blob->unhashed_list);
1106                 blob_table_insert(ctx->blob_table, blob);
1107                 blob->unhashed = 0;
1108         }
1109         return status;
1110 }
1111
1112 /*
1113  * Compute statistics about a list of blobs that will be written.
1114  *
1115  * Assumes the blobs are sorted such that all blobs located in each distinct WIM
1116  * (specified by WIMStruct) are together.
1117  *
1118  * For compactions, also verify that there are no overlapping resources.  This
1119  * really should be checked earlier, but for now it's easiest to check here.
1120  */
1121 static int
1122 compute_blob_list_stats(struct list_head *blob_list,
1123                         struct write_blobs_ctx *ctx)
1124 {
1125         struct blob_descriptor *blob;
1126         u64 total_bytes = 0;
1127         u64 num_blobs = 0;
1128         u64 total_parts = 0;
1129         WIMStruct *prev_wim_part = NULL;
1130         const struct wim_resource_descriptor *prev_rdesc = NULL;
1131
1132         list_for_each_entry(blob, blob_list, write_blobs_list) {
1133                 num_blobs++;
1134                 total_bytes += blob->size;
1135                 if (blob->blob_location == BLOB_IN_WIM) {
1136                         const struct wim_resource_descriptor *rdesc = blob->rdesc;
1137                         WIMStruct *wim = rdesc->wim;
1138
1139                         if (prev_wim_part != wim) {
1140                                 prev_wim_part = wim;
1141                                 total_parts++;
1142                         }
1143                         if (unlikely(wim->being_compacted) && rdesc != prev_rdesc) {
1144                                 if (prev_rdesc != NULL &&
1145                                     rdesc->offset_in_wim <
1146                                                 prev_rdesc->offset_in_wim +
1147                                                 prev_rdesc->size_in_wim)
1148                                 {
1149                                         WARNING("WIM file contains overlapping "
1150                                                 "resources!  Compaction is not "
1151                                                 "possible.");
1152                                         return WIMLIB_ERR_RESOURCE_ORDER;
1153                                 }
1154                                 prev_rdesc = rdesc;
1155                         }
1156                 }
1157         }
1158         ctx->progress_data.progress.write_streams.total_bytes       = total_bytes;
1159         ctx->progress_data.progress.write_streams.total_streams     = num_blobs;
1160         ctx->progress_data.progress.write_streams.completed_bytes   = 0;
1161         ctx->progress_data.progress.write_streams.completed_streams = 0;
1162         ctx->progress_data.progress.write_streams.compression_type  = ctx->out_ctype;
1163         ctx->progress_data.progress.write_streams.total_parts       = total_parts;
1164         ctx->progress_data.progress.write_streams.completed_parts   = 0;
1165         ctx->progress_data.next_progress = 0;
1166         return 0;
1167 }
1168
1169 /* Find blobs in @blob_list that can be copied to the output WIM in raw form
1170  * rather than compressed.  Delete these blobs from @blob_list and move them to
1171  * @raw_copy_blobs.  Return the total uncompressed size of the blobs that need
1172  * to be compressed.  */
1173 static u64
1174 find_raw_copy_blobs(struct list_head *blob_list, int write_resource_flags,
1175                     int out_ctype, u32 out_chunk_size,
1176                     struct list_head *raw_copy_blobs)
1177 {
1178         struct blob_descriptor *blob, *tmp;
1179         u64 num_nonraw_bytes = 0;
1180
1181         INIT_LIST_HEAD(raw_copy_blobs);
1182
1183         /* Initialize temporary raw_copy_ok flag.  */
1184         list_for_each_entry(blob, blob_list, write_blobs_list)
1185                 if (blob->blob_location == BLOB_IN_WIM)
1186                         blob->rdesc->raw_copy_ok = 0;
1187
1188         list_for_each_entry_safe(blob, tmp, blob_list, write_blobs_list) {
1189                 if (can_raw_copy(blob, write_resource_flags,
1190                                  out_ctype, out_chunk_size))
1191                 {
1192                         blob->rdesc->raw_copy_ok = 1;
1193                         list_move_tail(&blob->write_blobs_list, raw_copy_blobs);
1194                 } else {
1195                         num_nonraw_bytes += blob->size;
1196                 }
1197         }
1198
1199         return num_nonraw_bytes;
1200 }
1201
1202 /* Copy a raw compressed resource located in another WIM file to the WIM file
1203  * being written.  */
1204 static int
1205 write_raw_copy_resource(struct wim_resource_descriptor *in_rdesc,
1206                         struct filedes *out_fd)
1207 {
1208         u64 cur_read_offset;
1209         u64 end_read_offset;
1210         u8 buf[BUFFER_SIZE];
1211         size_t bytes_to_read;
1212         int ret;
1213         struct filedes *in_fd;
1214         struct blob_descriptor *blob;
1215         u64 out_offset_in_wim;
1216
1217         /* Copy the raw data.  */
1218         cur_read_offset = in_rdesc->offset_in_wim;
1219         end_read_offset = cur_read_offset + in_rdesc->size_in_wim;
1220
1221         out_offset_in_wim = out_fd->offset;
1222
1223         if (in_rdesc->is_pipable) {
1224                 if (cur_read_offset < sizeof(struct pwm_blob_hdr))
1225                         return WIMLIB_ERR_INVALID_PIPABLE_WIM;
1226                 cur_read_offset -= sizeof(struct pwm_blob_hdr);
1227                 out_offset_in_wim += sizeof(struct pwm_blob_hdr);
1228         }
1229         in_fd = &in_rdesc->wim->in_fd;
1230         wimlib_assert(cur_read_offset != end_read_offset);
1231
1232         if (likely(!in_rdesc->wim->being_compacted) ||
1233             in_rdesc->offset_in_wim > out_fd->offset) {
1234                 do {
1235                         bytes_to_read = min(sizeof(buf),
1236                                             end_read_offset - cur_read_offset);
1237
1238                         ret = full_pread(in_fd, buf, bytes_to_read,
1239                                          cur_read_offset);
1240                         if (ret)
1241                                 return ret;
1242
1243                         ret = full_write(out_fd, buf, bytes_to_read);
1244                         if (ret)
1245                                 return ret;
1246
1247                         cur_read_offset += bytes_to_read;
1248
1249                 } while (cur_read_offset != end_read_offset);
1250         } else {
1251                 /* Optimization: the WIM file is being compacted and the
1252                  * resource being written is already in the desired location.
1253                  * Skip over the data instead of re-writing it.  */
1254
1255                 /* Due the earlier check for overlapping resources, it should
1256                  * never be the case that we already overwrote the resource.  */
1257                 wimlib_assert(!(in_rdesc->offset_in_wim < out_fd->offset));
1258
1259                 if (-1 == filedes_seek(out_fd, out_fd->offset + in_rdesc->size_in_wim))
1260                         return WIMLIB_ERR_WRITE;
1261         }
1262
1263         list_for_each_entry(blob, &in_rdesc->blob_list, rdesc_node) {
1264                 if (blob->will_be_in_output_wim) {
1265                         blob_set_out_reshdr_for_reuse(blob);
1266                         if (in_rdesc->flags & WIM_RESHDR_FLAG_SOLID)
1267                                 blob->out_res_offset_in_wim = out_offset_in_wim;
1268                         else
1269                                 blob->out_reshdr.offset_in_wim = out_offset_in_wim;
1270
1271                 }
1272         }
1273         return 0;
1274 }
1275
1276 /* Copy a list of raw compressed resources located in other WIM file(s) to the
1277  * WIM file being written.  */
1278 static int
1279 write_raw_copy_resources(struct list_head *raw_copy_blobs,
1280                          struct filedes *out_fd,
1281                          struct write_blobs_progress_data *progress_data)
1282 {
1283         struct blob_descriptor *blob;
1284         int ret;
1285
1286         list_for_each_entry(blob, raw_copy_blobs, write_blobs_list)
1287                 blob->rdesc->raw_copy_ok = 1;
1288
1289         list_for_each_entry(blob, raw_copy_blobs, write_blobs_list) {
1290                 if (blob->rdesc->raw_copy_ok) {
1291                         /* Write each solid resource only one time.  */
1292                         ret = write_raw_copy_resource(blob->rdesc, out_fd);
1293                         if (ret)
1294                                 return ret;
1295                         blob->rdesc->raw_copy_ok = 0;
1296                 }
1297                 ret = do_write_blobs_progress(progress_data, blob->size,
1298                                               1, false);
1299                 if (ret)
1300                         return ret;
1301         }
1302         return 0;
1303 }
1304
1305 /* Wait for and write all chunks pending in the compressor.  */
1306 static int
1307 finish_remaining_chunks(struct write_blobs_ctx *ctx)
1308 {
1309         const void *cdata;
1310         u32 csize;
1311         u32 usize;
1312         int ret;
1313
1314         if (ctx->compressor == NULL)
1315                 return 0;
1316
1317         if (ctx->cur_chunk_buf_filled != 0) {
1318                 ctx->compressor->signal_chunk_filled(ctx->compressor,
1319                                                      ctx->cur_chunk_buf_filled);
1320         }
1321
1322         while (ctx->compressor->get_compression_result(ctx->compressor, &cdata,
1323                                                        &csize, &usize))
1324         {
1325                 ret = write_chunk(ctx, cdata, csize, usize);
1326                 if (ret)
1327                         return ret;
1328         }
1329         return 0;
1330 }
1331
1332 static void
1333 validate_blob_list(struct list_head *blob_list)
1334 {
1335         struct blob_descriptor *blob;
1336
1337         list_for_each_entry(blob, blob_list, write_blobs_list) {
1338                 wimlib_assert(blob->will_be_in_output_wim);
1339                 wimlib_assert(blob->size != 0);
1340         }
1341 }
1342
1343 static inline bool
1344 blob_is_in_file(const struct blob_descriptor *blob)
1345 {
1346         return blob->blob_location == BLOB_IN_FILE_ON_DISK
1347 #ifdef __WIN32__
1348             || blob->blob_location == BLOB_IN_WINNT_FILE_ON_DISK
1349             || blob->blob_location == BLOB_WIN32_ENCRYPTED
1350 #endif
1351            ;
1352 }
1353
1354 static void
1355 init_done_with_file_info(struct list_head *blob_list)
1356 {
1357         struct blob_descriptor *blob;
1358
1359         list_for_each_entry(blob, blob_list, write_blobs_list) {
1360                 if (blob_is_in_file(blob)) {
1361                         blob->file_inode->i_num_remaining_streams = 0;
1362                         blob->may_send_done_with_file = 1;
1363                 } else {
1364                         blob->may_send_done_with_file = 0;
1365                 }
1366         }
1367
1368         list_for_each_entry(blob, blob_list, write_blobs_list)
1369                 if (blob->may_send_done_with_file)
1370                         blob->file_inode->i_num_remaining_streams++;
1371 }
1372
1373 /*
1374  * Write a list of blobs to the output WIM file.
1375  *
1376  * @blob_list
1377  *      The list of blobs to write, specified by a list of 'struct blob_descriptor' linked
1378  *      by the 'write_blobs_list' member.
1379  *
1380  * @out_fd
1381  *      The file descriptor, opened for writing, to which to write the blobs.
1382  *
1383  * @write_resource_flags
1384  *      Flags to modify how the blobs are written:
1385  *
1386  *      WRITE_RESOURCE_FLAG_RECOMPRESS:
1387  *              Force compression of all resources, even if they could otherwise
1388  *              be re-used by copying the raw data, due to being located in a WIM
1389  *              file with compatible compression parameters.
1390  *
1391  *      WRITE_RESOURCE_FLAG_PIPABLE:
1392  *              Write the resources in the wimlib-specific pipable format, and
1393  *              furthermore do so in such a way that no seeking backwards in
1394  *              @out_fd will be performed (so it may be a pipe).
1395  *
1396  *      WRITE_RESOURCE_FLAG_SOLID:
1397  *              Combine all the blobs into a single resource rather than writing
1398  *              them in separate resources.  This flag is only valid if the WIM
1399  *              version number has been, or will be, set to WIM_VERSION_SOLID.
1400  *              This flag may not be combined with WRITE_RESOURCE_FLAG_PIPABLE.
1401  *
1402  * @out_ctype
1403  *      Compression format to use in the output resources, specified as one of
1404  *      the WIMLIB_COMPRESSION_TYPE_* constants.  WIMLIB_COMPRESSION_TYPE_NONE
1405  *      is allowed.
1406  *
1407  * @out_chunk_size
1408  *      Compression chunk size to use in the output resources.  It must be a
1409  *      valid chunk size for the specified compression format @out_ctype, unless
1410  *      @out_ctype is WIMLIB_COMPRESSION_TYPE_NONE, in which case this parameter
1411  *      is ignored.
1412  *
1413  * @num_threads
1414  *      Number of threads to use to compress data.  If 0, a default number of
1415  *      threads will be chosen.  The number of threads still may be decreased
1416  *      from the specified value if insufficient memory is detected.
1417  *
1418  * @blob_table
1419  *      If on-the-fly deduplication of unhashed blobs is desired, this parameter
1420  *      must be pointer to the blob table for the WIMStruct on whose behalf the
1421  *      blobs are being written.  Otherwise, this parameter can be NULL.
1422  *
1423  * @filter_ctx
1424  *      If on-the-fly deduplication of unhashed blobs is desired, this parameter
1425  *      can be a pointer to a context for blob filtering used to detect whether
1426  *      the duplicate blob has been hard-filtered or not.  If no blobs are
1427  *      hard-filtered or no blobs are unhashed, this parameter can be NULL.
1428  *
1429  * This function will write the blobs in @blob_list to resources in
1430  * consecutive positions in the output WIM file, or to a single solid resource
1431  * if WRITE_RESOURCE_FLAG_SOLID was specified in @write_resource_flags.  In both
1432  * cases, the @out_reshdr of the `struct blob_descriptor' for each blob written will be
1433  * updated to specify its location, size, and flags in the output WIM.  In the
1434  * solid resource case, WIM_RESHDR_FLAG_SOLID will be set in the @flags field of
1435  * each @out_reshdr, and furthermore @out_res_offset_in_wim and
1436  * @out_res_size_in_wim of each @out_reshdr will be set to the offset and size,
1437  * respectively, in the output WIM of the solid resource containing the
1438  * corresponding blob.
1439  *
1440  * Each of the blobs to write may be in any location supported by the
1441  * resource-handling code (specifically, read_blob_list()), such as the contents
1442  * of external file that has been logically added to the output WIM, or a blob
1443  * in another WIM file that has been imported, or even a blob in the "same" WIM
1444  * file of which a modified copy is being written.  In the case that a blob is
1445  * already in a WIM file and uses compatible compression parameters, by default
1446  * this function will re-use the raw data instead of decompressing it, then
1447  * recompressing it; however, with WRITE_RESOURCE_FLAG_RECOMPRESS
1448  * specified in @write_resource_flags, this is not done.
1449  *
1450  * As a further requirement, this function requires that the
1451  * @will_be_in_output_wim member be set to 1 on all blobs in @blob_list as well
1452  * as any other blobs not in @blob_list that will be in the output WIM file, but
1453  * set to 0 on any other blobs in the output WIM's blob table or sharing a solid
1454  * resource with a blob in @blob_list.  Still furthermore, if on-the-fly
1455  * deduplication of blobs is possible, then all blobs in @blob_list must also be
1456  * linked by @blob_table_list along with any other blobs that have
1457  * @will_be_in_output_wim set.
1458  *
1459  * This function handles on-the-fly deduplication of blobs for which SHA-1
1460  * message digests have not yet been calculated.  Such blobs may or may not need
1461  * to be written.  If @blob_table is non-NULL, then each blob in @blob_list that
1462  * has @unhashed set but not @unique_size set is checksummed immediately before
1463  * it would otherwise be read for writing in order to determine if it is
1464  * identical to another blob already being written or one that would be filtered
1465  * out of the output WIM using blob_filtered() with the context @filter_ctx.
1466  * Each such duplicate blob will be removed from @blob_list, its reference count
1467  * transferred to the pre-existing duplicate blob, its memory freed, and will
1468  * not be written.  Alternatively, if a blob in @blob_list is a duplicate with
1469  * any blob in @blob_table that has not been marked for writing or would not be
1470  * hard-filtered, it is freed and the pre-existing duplicate is written instead,
1471  * taking ownership of the reference count and slot in the @blob_table_list.
1472  *
1473  * Returns 0 if every blob was either written successfully or did not need to be
1474  * written; otherwise returns a non-zero error code.
1475  */
1476 static int
1477 write_blob_list(struct list_head *blob_list,
1478                 struct filedes *out_fd,
1479                 int write_resource_flags,
1480                 int out_ctype,
1481                 u32 out_chunk_size,
1482                 unsigned num_threads,
1483                 struct blob_table *blob_table,
1484                 struct filter_context *filter_ctx,
1485                 wimlib_progress_func_t progfunc,
1486                 void *progctx)
1487 {
1488         int ret;
1489         struct write_blobs_ctx ctx;
1490         struct list_head raw_copy_blobs;
1491         u64 num_nonraw_bytes;
1492
1493         wimlib_assert((write_resource_flags &
1494                        (WRITE_RESOURCE_FLAG_SOLID |
1495                         WRITE_RESOURCE_FLAG_PIPABLE)) !=
1496                                 (WRITE_RESOURCE_FLAG_SOLID |
1497                                  WRITE_RESOURCE_FLAG_PIPABLE));
1498
1499         validate_blob_list(blob_list);
1500
1501         if (list_empty(blob_list))
1502                 return 0;
1503
1504         /* If needed, set auxiliary information so that we can detect when the
1505          * library has finished using each external file.  */
1506         if (unlikely(write_resource_flags & WRITE_RESOURCE_FLAG_SEND_DONE_WITH_FILE))
1507                 init_done_with_file_info(blob_list);
1508
1509         memset(&ctx, 0, sizeof(ctx));
1510
1511         ctx.out_fd = out_fd;
1512         ctx.blob_table = blob_table;
1513         ctx.out_ctype = out_ctype;
1514         ctx.out_chunk_size = out_chunk_size;
1515         ctx.write_resource_flags = write_resource_flags;
1516         ctx.filter_ctx = filter_ctx;
1517
1518         /*
1519          * We normally sort the blobs to write by a "sequential" order that is
1520          * optimized for reading.  But when using solid compression, we instead
1521          * sort the blobs by file extension and file name (when applicable; and
1522          * we don't do this for blobs from solid resources) so that similar
1523          * files are grouped together, which improves the compression ratio.
1524          * This is somewhat of a hack since a blob does not necessarily
1525          * correspond one-to-one with a filename, nor is there any guarantee
1526          * that two files with similar names or extensions are actually similar
1527          * in content.  A potential TODO is to sort the blobs based on some
1528          * measure of similarity of their actual contents.
1529          */
1530
1531         ret = sort_blob_list_by_sequential_order(blob_list,
1532                                                  offsetof(struct blob_descriptor,
1533                                                           write_blobs_list));
1534         if (ret)
1535                 return ret;
1536
1537         ret = compute_blob_list_stats(blob_list, &ctx);
1538         if (ret)
1539                 return ret;
1540
1541         if (write_resource_flags & WRITE_RESOURCE_FLAG_SOLID_SORT) {
1542                 ret = sort_blob_list_for_solid_compression(blob_list);
1543                 if (unlikely(ret))
1544                         WARNING("Failed to sort blobs for solid compression. Continuing anyways.");
1545         }
1546
1547         ctx.progress_data.progfunc = progfunc;
1548         ctx.progress_data.progctx = progctx;
1549
1550         num_nonraw_bytes = find_raw_copy_blobs(blob_list, write_resource_flags,
1551                                                out_ctype, out_chunk_size,
1552                                                &raw_copy_blobs);
1553
1554         /* Copy any compressed resources for which the raw data can be reused
1555          * without decompression.  */
1556         ret = write_raw_copy_resources(&raw_copy_blobs, ctx.out_fd,
1557                                        &ctx.progress_data);
1558
1559         if (ret || num_nonraw_bytes == 0)
1560                 goto out_destroy_context;
1561
1562         /* Unless uncompressed output was required, allocate a chunk_compressor
1563          * to do compression.  There are serial and parallel implementations of
1564          * the chunk_compressor interface.  We default to parallel using the
1565          * specified number of threads, unless the upper bound on the number
1566          * bytes needing to be compressed is less than a heuristic value.  */
1567         if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
1568
1569         #ifdef ENABLE_MULTITHREADED_COMPRESSION
1570                 if (num_nonraw_bytes > max(2000000, out_chunk_size)) {
1571                         ret = new_parallel_chunk_compressor(out_ctype,
1572                                                             out_chunk_size,
1573                                                             num_threads, 0,
1574                                                             &ctx.compressor);
1575                         if (ret > 0) {
1576                                 WARNING("Couldn't create parallel chunk compressor: %"TS".\n"
1577                                         "          Falling back to single-threaded compression.",
1578                                         wimlib_get_error_string(ret));
1579                         }
1580                 }
1581         #endif
1582
1583                 if (ctx.compressor == NULL) {
1584                         ret = new_serial_chunk_compressor(out_ctype, out_chunk_size,
1585                                                           &ctx.compressor);
1586                         if (ret)
1587                                 goto out_destroy_context;
1588                 }
1589         }
1590
1591         if (ctx.compressor)
1592                 ctx.progress_data.progress.write_streams.num_threads = ctx.compressor->num_threads;
1593         else
1594                 ctx.progress_data.progress.write_streams.num_threads = 1;
1595
1596         INIT_LIST_HEAD(&ctx.blobs_being_compressed);
1597         INIT_LIST_HEAD(&ctx.blobs_in_solid_resource);
1598
1599         ret = call_progress(ctx.progress_data.progfunc,
1600                             WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
1601                             &ctx.progress_data.progress,
1602                             ctx.progress_data.progctx);
1603         if (ret)
1604                 goto out_destroy_context;
1605
1606         if (write_resource_flags & WRITE_RESOURCE_FLAG_SOLID) {
1607                 ret = begin_write_resource(&ctx, num_nonraw_bytes);
1608                 if (ret)
1609                         goto out_destroy_context;
1610         }
1611
1612         /* Read the list of blobs needing to be compressed, using the specified
1613          * callbacks to execute processing of the data.  */
1614
1615         struct read_blob_callbacks cbs = {
1616                 .begin_blob     = write_blob_begin_read,
1617                 .consume_chunk  = write_blob_process_chunk,
1618                 .end_blob       = write_blob_end_read,
1619                 .ctx            = &ctx,
1620         };
1621
1622         ret = read_blob_list(blob_list,
1623                              offsetof(struct blob_descriptor, write_blobs_list),
1624                              &cbs,
1625                              BLOB_LIST_ALREADY_SORTED |
1626                                 VERIFY_BLOB_HASHES |
1627                                 COMPUTE_MISSING_BLOB_HASHES);
1628
1629         if (ret)
1630                 goto out_destroy_context;
1631
1632         ret = finish_remaining_chunks(&ctx);
1633         if (ret)
1634                 goto out_destroy_context;
1635
1636         if (write_resource_flags & WRITE_RESOURCE_FLAG_SOLID) {
1637                 struct wim_reshdr reshdr;
1638                 struct blob_descriptor *blob;
1639                 u64 offset_in_res;
1640
1641                 ret = end_write_resource(&ctx, &reshdr);
1642                 if (ret)
1643                         goto out_destroy_context;
1644
1645                 offset_in_res = 0;
1646                 list_for_each_entry(blob, &ctx.blobs_in_solid_resource, write_blobs_list) {
1647                         blob->out_reshdr.size_in_wim = blob->size;
1648                         blob->out_reshdr.flags = reshdr_flags_for_blob(blob) |
1649                                                  WIM_RESHDR_FLAG_SOLID;
1650                         blob->out_reshdr.uncompressed_size = 0;
1651                         blob->out_reshdr.offset_in_wim = offset_in_res;
1652                         blob->out_res_offset_in_wim = reshdr.offset_in_wim;
1653                         blob->out_res_size_in_wim = reshdr.size_in_wim;
1654                         blob->out_res_uncompressed_size = reshdr.uncompressed_size;
1655                         offset_in_res += blob->size;
1656                 }
1657                 wimlib_assert(offset_in_res == reshdr.uncompressed_size);
1658         }
1659
1660 out_destroy_context:
1661         FREE(ctx.chunk_csizes);
1662         if (ctx.compressor)
1663                 ctx.compressor->destroy(ctx.compressor);
1664         return ret;
1665 }
1666
1667
1668 static int
1669 write_file_data_blobs(WIMStruct *wim,
1670                       struct list_head *blob_list,
1671                       int write_flags,
1672                       unsigned num_threads,
1673                       struct filter_context *filter_ctx)
1674 {
1675         int out_ctype;
1676         u32 out_chunk_size;
1677         int write_resource_flags;
1678
1679         write_resource_flags = write_flags_to_resource_flags(write_flags);
1680
1681         if (write_resource_flags & WRITE_RESOURCE_FLAG_SOLID) {
1682                 out_chunk_size = wim->out_solid_chunk_size;
1683                 out_ctype = wim->out_solid_compression_type;
1684         } else {
1685                 out_chunk_size = wim->out_chunk_size;
1686                 out_ctype = wim->out_compression_type;
1687         }
1688
1689         return write_blob_list(blob_list,
1690                                &wim->out_fd,
1691                                write_resource_flags,
1692                                out_ctype,
1693                                out_chunk_size,
1694                                num_threads,
1695                                wim->blob_table,
1696                                filter_ctx,
1697                                wim->progfunc,
1698                                wim->progctx);
1699 }
1700
1701 /* Write the contents of the specified blob as a WIM resource.  */
1702 static int
1703 write_wim_resource(struct blob_descriptor *blob,
1704                    struct filedes *out_fd,
1705                    int out_ctype,
1706                    u32 out_chunk_size,
1707                    int write_resource_flags)
1708 {
1709         LIST_HEAD(blob_list);
1710         list_add(&blob->write_blobs_list, &blob_list);
1711         blob->will_be_in_output_wim = 1;
1712         return write_blob_list(&blob_list,
1713                                out_fd,
1714                                write_resource_flags & ~WRITE_RESOURCE_FLAG_SOLID,
1715                                out_ctype,
1716                                out_chunk_size,
1717                                1,
1718                                NULL,
1719                                NULL,
1720                                NULL,
1721                                NULL);
1722 }
1723
1724 /* Write the contents of the specified buffer as a WIM resource.  */
1725 int
1726 write_wim_resource_from_buffer(const void *buf,
1727                                size_t buf_size,
1728                                bool is_metadata,
1729                                struct filedes *out_fd,
1730                                int out_ctype,
1731                                u32 out_chunk_size,
1732                                struct wim_reshdr *out_reshdr,
1733                                u8 *hash_ret,
1734                                int write_resource_flags)
1735 {
1736         int ret;
1737         struct blob_descriptor blob;
1738
1739         if (unlikely(buf_size == 0)) {
1740                 zero_reshdr(out_reshdr);
1741                 if (hash_ret)
1742                         copy_hash(hash_ret, zero_hash);
1743                 return 0;
1744         }
1745
1746         blob_set_is_located_in_attached_buffer(&blob, (void *)buf, buf_size);
1747         sha1_buffer(buf, buf_size, blob.hash);
1748         blob.unhashed = 0;
1749         blob.is_metadata = is_metadata;
1750
1751         ret = write_wim_resource(&blob, out_fd, out_ctype, out_chunk_size,
1752                                  write_resource_flags);
1753         if (ret)
1754                 return ret;
1755
1756         copy_reshdr(out_reshdr, &blob.out_reshdr);
1757
1758         if (hash_ret)
1759                 copy_hash(hash_ret, blob.hash);
1760         return 0;
1761 }
1762
1763 struct blob_size_table {
1764         struct hlist_head *array;
1765         size_t num_entries;
1766         size_t capacity;
1767 };
1768
1769 static int
1770 init_blob_size_table(struct blob_size_table *tab, size_t capacity)
1771 {
1772         tab->array = CALLOC(capacity, sizeof(tab->array[0]));
1773         if (tab->array == NULL)
1774                 return WIMLIB_ERR_NOMEM;
1775         tab->num_entries = 0;
1776         tab->capacity = capacity;
1777         return 0;
1778 }
1779
1780 static void
1781 destroy_blob_size_table(struct blob_size_table *tab)
1782 {
1783         FREE(tab->array);
1784 }
1785
1786 static int
1787 blob_size_table_insert(struct blob_descriptor *blob, void *_tab)
1788 {
1789         struct blob_size_table *tab = _tab;
1790         size_t pos;
1791         struct blob_descriptor *same_size_blob;
1792
1793         pos = hash_u64(blob->size) % tab->capacity;
1794         blob->unique_size = 1;
1795         hlist_for_each_entry(same_size_blob, &tab->array[pos], hash_list_2) {
1796                 if (same_size_blob->size == blob->size) {
1797                         blob->unique_size = 0;
1798                         same_size_blob->unique_size = 0;
1799                         break;
1800                 }
1801         }
1802
1803         hlist_add_head(&blob->hash_list_2, &tab->array[pos]);
1804         tab->num_entries++;
1805         return 0;
1806 }
1807
1808 struct find_blobs_ctx {
1809         WIMStruct *wim;
1810         int write_flags;
1811         struct list_head blob_list;
1812         struct blob_size_table blob_size_tab;
1813 };
1814
1815 static void
1816 reference_blob_for_write(struct blob_descriptor *blob,
1817                          struct list_head *blob_list, u32 nref)
1818 {
1819         if (!blob->will_be_in_output_wim) {
1820                 blob->out_refcnt = 0;
1821                 list_add_tail(&blob->write_blobs_list, blob_list);
1822                 blob->will_be_in_output_wim = 1;
1823         }
1824         blob->out_refcnt += nref;
1825 }
1826
1827 static int
1828 fully_reference_blob_for_write(struct blob_descriptor *blob, void *_blob_list)
1829 {
1830         struct list_head *blob_list = _blob_list;
1831         blob->will_be_in_output_wim = 0;
1832         reference_blob_for_write(blob, blob_list, blob->refcnt);
1833         return 0;
1834 }
1835
1836 static int
1837 inode_find_blobs_to_reference(const struct wim_inode *inode,
1838                               const struct blob_table *table,
1839                               struct list_head *blob_list)
1840 {
1841         wimlib_assert(inode->i_nlink > 0);
1842
1843         for (unsigned i = 0; i < inode->i_num_streams; i++) {
1844                 struct blob_descriptor *blob;
1845                 const u8 *hash;
1846
1847                 blob = stream_blob(&inode->i_streams[i], table);
1848                 if (blob) {
1849                         reference_blob_for_write(blob, blob_list, inode->i_nlink);
1850                 } else {
1851                         hash = stream_hash(&inode->i_streams[i]);
1852                         if (!is_zero_hash(hash))
1853                                 return blob_not_found_error(inode, hash);
1854                 }
1855         }
1856         return 0;
1857 }
1858
1859 static int
1860 do_blob_set_not_in_output_wim(struct blob_descriptor *blob, void *_ignore)
1861 {
1862         blob->will_be_in_output_wim = 0;
1863         return 0;
1864 }
1865
1866 static int
1867 image_find_blobs_to_reference(WIMStruct *wim)
1868 {
1869         struct wim_image_metadata *imd;
1870         struct wim_inode *inode;
1871         struct blob_descriptor *blob;
1872         struct list_head *blob_list;
1873         int ret;
1874
1875         imd = wim_get_current_image_metadata(wim);
1876
1877         image_for_each_unhashed_blob(blob, imd)
1878                 blob->will_be_in_output_wim = 0;
1879
1880         blob_list = wim->private;
1881         image_for_each_inode(inode, imd) {
1882                 ret = inode_find_blobs_to_reference(inode,
1883                                                     wim->blob_table,
1884                                                     blob_list);
1885                 if (ret)
1886                         return ret;
1887         }
1888         return 0;
1889 }
1890
1891 static int
1892 prepare_unfiltered_list_of_blobs_in_output_wim(WIMStruct *wim,
1893                                                int image,
1894                                                int blobs_ok,
1895                                                struct list_head *blob_list_ret)
1896 {
1897         int ret;
1898
1899         INIT_LIST_HEAD(blob_list_ret);
1900
1901         if (blobs_ok && (image == WIMLIB_ALL_IMAGES ||
1902                          (image == 1 && wim->hdr.image_count == 1)))
1903         {
1904                 /* Fast case:  Assume that all blobs are being written and that
1905                  * the reference counts are correct.  */
1906                 struct blob_descriptor *blob;
1907                 struct wim_image_metadata *imd;
1908                 unsigned i;
1909
1910                 for_blob_in_table(wim->blob_table,
1911                                   fully_reference_blob_for_write,
1912                                   blob_list_ret);
1913
1914                 for (i = 0; i < wim->hdr.image_count; i++) {
1915                         imd = wim->image_metadata[i];
1916                         image_for_each_unhashed_blob(blob, imd)
1917                                 fully_reference_blob_for_write(blob, blob_list_ret);
1918                 }
1919         } else {
1920                 /* Slow case:  Walk through the images being written and
1921                  * determine the blobs referenced.  */
1922                 for_blob_in_table(wim->blob_table,
1923                                   do_blob_set_not_in_output_wim, NULL);
1924                 wim->private = blob_list_ret;
1925                 ret = for_image(wim, image, image_find_blobs_to_reference);
1926                 if (ret)
1927                         return ret;
1928         }
1929
1930         return 0;
1931 }
1932
1933 struct insert_other_if_hard_filtered_ctx {
1934         struct blob_size_table *tab;
1935         struct filter_context *filter_ctx;
1936 };
1937
1938 static int
1939 insert_other_if_hard_filtered(struct blob_descriptor *blob, void *_ctx)
1940 {
1941         struct insert_other_if_hard_filtered_ctx *ctx = _ctx;
1942
1943         if (!blob->will_be_in_output_wim &&
1944             blob_hard_filtered(blob, ctx->filter_ctx))
1945                 blob_size_table_insert(blob, ctx->tab);
1946         return 0;
1947 }
1948
1949 static int
1950 determine_blob_size_uniquity(struct list_head *blob_list,
1951                              struct blob_table *lt,
1952                              struct filter_context *filter_ctx)
1953 {
1954         int ret;
1955         struct blob_size_table tab;
1956         struct blob_descriptor *blob;
1957
1958         ret = init_blob_size_table(&tab, 9001);
1959         if (ret)
1960                 return ret;
1961
1962         if (may_hard_filter_blobs(filter_ctx)) {
1963                 struct insert_other_if_hard_filtered_ctx ctx = {
1964                         .tab = &tab,
1965                         .filter_ctx = filter_ctx,
1966                 };
1967                 for_blob_in_table(lt, insert_other_if_hard_filtered, &ctx);
1968         }
1969
1970         list_for_each_entry(blob, blob_list, write_blobs_list)
1971                 blob_size_table_insert(blob, &tab);
1972
1973         destroy_blob_size_table(&tab);
1974         return 0;
1975 }
1976
1977 static void
1978 filter_blob_list_for_write(struct list_head *blob_list,
1979                            struct filter_context *filter_ctx)
1980 {
1981         struct blob_descriptor *blob, *tmp;
1982
1983         list_for_each_entry_safe(blob, tmp, blob_list, write_blobs_list) {
1984                 int status = blob_filtered(blob, filter_ctx);
1985
1986                 if (status == 0) {
1987                         /* Not filtered.  */
1988                         continue;
1989                 } else {
1990                         if (status > 0) {
1991                                 /* Soft filtered.  */
1992                         } else {
1993                                 /* Hard filtered.  */
1994                                 blob->will_be_in_output_wim = 0;
1995                                 list_del(&blob->blob_table_list);
1996                         }
1997                         list_del(&blob->write_blobs_list);
1998                 }
1999         }
2000 }
2001
2002 /*
2003  * prepare_blob_list_for_write() -
2004  *
2005  * Prepare the list of blobs to write for writing a WIM containing the specified
2006  * image(s) with the specified write flags.
2007  *
2008  * @wim
2009  *      The WIMStruct on whose behalf the write is occurring.
2010  *
2011  * @image
2012  *      Image(s) from the WIM to write; may be WIMLIB_ALL_IMAGES.
2013  *
2014  * @write_flags
2015  *      WIMLIB_WRITE_FLAG_* flags for the write operation:
2016  *
2017  *      STREAMS_OK:  For writes of all images, assume that all blobs in the blob
2018  *      table of @wim and the per-image lists of unhashed blobs should be taken
2019  *      as-is, and image metadata should not be searched for references.  This
2020  *      does not exclude filtering with APPEND and SKIP_EXTERNAL_WIMS, below.
2021  *
2022  *      APPEND:  Blobs already present in @wim shall not be returned in
2023  *      @blob_list_ret.
2024  *
2025  *      SKIP_EXTERNAL_WIMS:  Blobs already present in a WIM file, but not @wim,
2026  *      shall be returned in neither @blob_list_ret nor @blob_table_list_ret.
2027  *
2028  * @blob_list_ret
2029  *      List of blobs, linked by write_blobs_list, that need to be written will
2030  *      be returned here.
2031  *
2032  *      Note that this function assumes that unhashed blobs will be written; it
2033  *      does not take into account that they may become duplicates when actually
2034  *      hashed.
2035  *
2036  * @blob_table_list_ret
2037  *      List of blobs, linked by blob_table_list, that need to be included in
2038  *      the WIM's blob table will be returned here.  This will be a superset of
2039  *      the blobs in @blob_list_ret.
2040  *
2041  *      This list will be a proper superset of @blob_list_ret if and only if
2042  *      WIMLIB_WRITE_FLAG_APPEND was specified in @write_flags and some of the
2043  *      blobs that would otherwise need to be written were already located in
2044  *      the WIM file.
2045  *
2046  *      All blobs in this list will have @out_refcnt set to the number of
2047  *      references to the blob in the output WIM.  If
2048  *      WIMLIB_WRITE_FLAG_STREAMS_OK was specified in @write_flags, @out_refcnt
2049  *      may be as low as 0.
2050  *
2051  * @filter_ctx_ret
2052  *      A context for queries of blob filter status with blob_filtered() is
2053  *      returned in this location.
2054  *
2055  * In addition, @will_be_in_output_wim will be set to 1 in all blobs inserted
2056  * into @blob_table_list_ret and to 0 in all blobs in the blob table of @wim not
2057  * inserted into @blob_table_list_ret.
2058  *
2059  * Still furthermore, @unique_size will be set to 1 on all blobs in
2060  * @blob_list_ret that have unique size among all blobs in @blob_list_ret and
2061  * among all blobs in the blob table of @wim that are ineligible for being
2062  * written due to filtering.
2063  *
2064  * Returns 0 on success; nonzero on read error, memory allocation error, or
2065  * otherwise.
2066  */
2067 static int
2068 prepare_blob_list_for_write(WIMStruct *wim, int image,
2069                             int write_flags,
2070                             struct list_head *blob_list_ret,
2071                             struct list_head *blob_table_list_ret,
2072                             struct filter_context *filter_ctx_ret)
2073 {
2074         int ret;
2075         struct blob_descriptor *blob;
2076
2077         filter_ctx_ret->write_flags = write_flags;
2078         filter_ctx_ret->wim = wim;
2079
2080         ret = prepare_unfiltered_list_of_blobs_in_output_wim(
2081                                 wim,
2082                                 image,
2083                                 write_flags & WIMLIB_WRITE_FLAG_STREAMS_OK,
2084                                 blob_list_ret);
2085         if (ret)
2086                 return ret;
2087
2088         INIT_LIST_HEAD(blob_table_list_ret);
2089         list_for_each_entry(blob, blob_list_ret, write_blobs_list)
2090                 list_add_tail(&blob->blob_table_list, blob_table_list_ret);
2091
2092         ret = determine_blob_size_uniquity(blob_list_ret, wim->blob_table,
2093                                            filter_ctx_ret);
2094         if (ret)
2095                 return ret;
2096
2097         if (may_filter_blobs(filter_ctx_ret))
2098                 filter_blob_list_for_write(blob_list_ret, filter_ctx_ret);
2099
2100         return 0;
2101 }
2102
2103 static int
2104 write_file_data(WIMStruct *wim, int image, int write_flags,
2105                 unsigned num_threads,
2106                 struct list_head *blob_list_override,
2107                 struct list_head *blob_table_list_ret)
2108 {
2109         int ret;
2110         struct list_head _blob_list;
2111         struct list_head *blob_list;
2112         struct blob_descriptor *blob;
2113         struct filter_context _filter_ctx;
2114         struct filter_context *filter_ctx;
2115
2116         if (blob_list_override == NULL) {
2117                 /* Normal case: prepare blob list from image(s) being written.
2118                  */
2119                 blob_list = &_blob_list;
2120                 filter_ctx = &_filter_ctx;
2121                 ret = prepare_blob_list_for_write(wim, image, write_flags,
2122                                                   blob_list,
2123                                                   blob_table_list_ret,
2124                                                   filter_ctx);
2125                 if (ret)
2126                         return ret;
2127         } else {
2128                 /* Currently only as a result of wimlib_split() being called:
2129                  * use blob list already explicitly provided.  Use existing
2130                  * reference counts.  */
2131                 blob_list = blob_list_override;
2132                 filter_ctx = NULL;
2133                 INIT_LIST_HEAD(blob_table_list_ret);
2134                 list_for_each_entry(blob, blob_list, write_blobs_list) {
2135                         blob->out_refcnt = blob->refcnt;
2136                         blob->will_be_in_output_wim = 1;
2137                         blob->unique_size = 0;
2138                         list_add_tail(&blob->blob_table_list, blob_table_list_ret);
2139                 }
2140         }
2141
2142         return write_file_data_blobs(wim,
2143                                      blob_list,
2144                                      write_flags,
2145                                      num_threads,
2146                                      filter_ctx);
2147 }
2148
2149 static int
2150 write_metadata_resources(WIMStruct *wim, int image, int write_flags)
2151 {
2152         int ret;
2153         int start_image;
2154         int end_image;
2155         int write_resource_flags;
2156
2157         if (write_flags & WIMLIB_WRITE_FLAG_NO_METADATA)
2158                 return 0;
2159
2160         write_resource_flags = write_flags_to_resource_flags(write_flags);
2161
2162         write_resource_flags &= ~WRITE_RESOURCE_FLAG_SOLID;
2163
2164         ret = call_progress(wim->progfunc,
2165                             WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN,
2166                             NULL, wim->progctx);
2167         if (ret)
2168                 return ret;
2169
2170         if (image == WIMLIB_ALL_IMAGES) {
2171                 start_image = 1;
2172                 end_image = wim->hdr.image_count;
2173         } else {
2174                 start_image = image;
2175                 end_image = image;
2176         }
2177
2178         for (int i = start_image; i <= end_image; i++) {
2179                 struct wim_image_metadata *imd;
2180
2181                 imd = wim->image_metadata[i - 1];
2182                 /* Build a new metadata resource only if image was modified from
2183                  * the original (or was newly added).  Otherwise just copy the
2184                  * existing one.  */
2185                 if (imd->modified) {
2186                         ret = write_metadata_resource(wim, i,
2187                                                       write_resource_flags);
2188                 } else if (write_flags & WIMLIB_WRITE_FLAG_UNSAFE_COMPACT) {
2189                         /* For compactions, existing metadata resources are
2190                          * written along with the existing file resources.  */
2191                         ret = 0;
2192                 } else if (write_flags & WIMLIB_WRITE_FLAG_APPEND) {
2193                         blob_set_out_reshdr_for_reuse(imd->metadata_blob);
2194                         ret = 0;
2195                 } else {
2196                         ret = write_wim_resource(imd->metadata_blob,
2197                                                  &wim->out_fd,
2198                                                  wim->out_compression_type,
2199                                                  wim->out_chunk_size,
2200                                                  write_resource_flags);
2201                 }
2202                 if (ret)
2203                         return ret;
2204         }
2205
2206         return call_progress(wim->progfunc,
2207                              WIMLIB_PROGRESS_MSG_WRITE_METADATA_END,
2208                              NULL, wim->progctx);
2209 }
2210
2211 static int
2212 open_wim_writable(WIMStruct *wim, const tchar *path, int open_flags)
2213 {
2214         int raw_fd = topen(path, open_flags | O_BINARY, 0644);
2215         if (raw_fd < 0) {
2216                 ERROR_WITH_ERRNO("Failed to open \"%"TS"\" for writing", path);
2217                 return WIMLIB_ERR_OPEN;
2218         }
2219         filedes_init(&wim->out_fd, raw_fd);
2220         return 0;
2221 }
2222
2223 static int
2224 close_wim_writable(WIMStruct *wim, int write_flags)
2225 {
2226         int ret = 0;
2227
2228         if (!(write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR))
2229                 if (filedes_valid(&wim->out_fd))
2230                         if (filedes_close(&wim->out_fd))
2231                                 ret = WIMLIB_ERR_WRITE;
2232         filedes_invalidate(&wim->out_fd);
2233         return ret;
2234 }
2235
2236 static int
2237 cmp_blobs_by_out_rdesc(const void *p1, const void *p2)
2238 {
2239         const struct blob_descriptor *blob1, *blob2;
2240
2241         blob1 = *(const struct blob_descriptor**)p1;
2242         blob2 = *(const struct blob_descriptor**)p2;
2243
2244         if (blob1->out_reshdr.flags & WIM_RESHDR_FLAG_SOLID) {
2245                 if (blob2->out_reshdr.flags & WIM_RESHDR_FLAG_SOLID) {
2246                         if (blob1->out_res_offset_in_wim != blob2->out_res_offset_in_wim)
2247                                 return cmp_u64(blob1->out_res_offset_in_wim,
2248                                                blob2->out_res_offset_in_wim);
2249                 } else {
2250                         return 1;
2251                 }
2252         } else {
2253                 if (blob2->out_reshdr.flags & WIM_RESHDR_FLAG_SOLID)
2254                         return -1;
2255         }
2256         return cmp_u64(blob1->out_reshdr.offset_in_wim,
2257                        blob2->out_reshdr.offset_in_wim);
2258 }
2259
2260 static int
2261 write_blob_table(WIMStruct *wim, int image, int write_flags,
2262                  struct list_head *blob_table_list)
2263 {
2264         int ret;
2265
2266         /* Set output resource metadata for blobs already present in WIM.  */
2267         if (write_flags & WIMLIB_WRITE_FLAG_APPEND) {
2268                 struct blob_descriptor *blob;
2269                 list_for_each_entry(blob, blob_table_list, blob_table_list) {
2270                         if (blob->blob_location == BLOB_IN_WIM &&
2271                             blob->rdesc->wim == wim)
2272                         {
2273                                 blob_set_out_reshdr_for_reuse(blob);
2274                         }
2275                 }
2276         }
2277
2278         ret = sort_blob_list(blob_table_list,
2279                              offsetof(struct blob_descriptor, blob_table_list),
2280                              cmp_blobs_by_out_rdesc);
2281         if (ret)
2282                 return ret;
2283
2284         /* Add entries for metadata resources.  */
2285         if (!(write_flags & WIMLIB_WRITE_FLAG_NO_METADATA)) {
2286                 int start_image;
2287                 int end_image;
2288
2289                 if (image == WIMLIB_ALL_IMAGES) {
2290                         start_image = 1;
2291                         end_image = wim->hdr.image_count;
2292                 } else {
2293                         start_image = image;
2294                         end_image = image;
2295                 }
2296
2297                 /* Push metadata blob table entries onto the front of the list
2298                  * in reverse order, so that they're written in order.
2299                  */
2300                 for (int i = end_image; i >= start_image; i--) {
2301                         struct blob_descriptor *metadata_blob;
2302
2303                         metadata_blob = wim->image_metadata[i - 1]->metadata_blob;
2304                         wimlib_assert(metadata_blob->out_reshdr.flags & WIM_RESHDR_FLAG_METADATA);
2305                         metadata_blob->out_refcnt = 1;
2306                         list_add(&metadata_blob->blob_table_list, blob_table_list);
2307                 }
2308         }
2309
2310         return write_blob_table_from_blob_list(blob_table_list,
2311                                                &wim->out_fd,
2312                                                wim->out_hdr.part_number,
2313                                                &wim->out_hdr.blob_table_reshdr,
2314                                                write_flags_to_resource_flags(write_flags));
2315 }
2316
2317 /*
2318  * Finish writing a WIM file: write the blob table, xml data, and integrity
2319  * table, then overwrite the WIM header.
2320  *
2321  * The output file descriptor is closed on success, except when writing to a
2322  * user-specified file descriptor (WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR set).
2323  */
2324 static int
2325 finish_write(WIMStruct *wim, int image, int write_flags,
2326              struct list_head *blob_table_list)
2327 {
2328         int write_resource_flags;
2329         off_t old_blob_table_end = 0;
2330         struct integrity_table *old_integrity_table = NULL;
2331         off_t new_blob_table_end;
2332         u64 xml_totalbytes;
2333         int ret;
2334
2335         write_resource_flags = write_flags_to_resource_flags(write_flags);
2336
2337         /* In the WIM header, there is room for the resource entry for a
2338          * metadata resource labeled as the "boot metadata".  This entry should
2339          * be zeroed out if there is no bootable image (boot_idx 0).  Otherwise,
2340          * it should be a copy of the resource entry for the image that is
2341          * marked as bootable.  */
2342         if (wim->out_hdr.boot_idx == 0) {
2343                 zero_reshdr(&wim->out_hdr.boot_metadata_reshdr);
2344         } else {
2345                 copy_reshdr(&wim->out_hdr.boot_metadata_reshdr,
2346                             &wim->image_metadata[
2347                                 wim->out_hdr.boot_idx - 1]->metadata_blob->out_reshdr);
2348         }
2349
2350         /* If appending to a WIM file containing an integrity table, we'd like
2351          * to re-use the information in the old integrity table instead of
2352          * recalculating it.  But we might overwrite the old integrity table
2353          * when we expand the XML data.  Read it into memory just in case.  */
2354         if ((write_flags & (WIMLIB_WRITE_FLAG_APPEND |
2355                             WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)) ==
2356                 (WIMLIB_WRITE_FLAG_APPEND |
2357                  WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
2358             && wim_has_integrity_table(wim))
2359         {
2360                 old_blob_table_end = wim->hdr.blob_table_reshdr.offset_in_wim +
2361                                      wim->hdr.blob_table_reshdr.size_in_wim;
2362                 (void)read_integrity_table(wim,
2363                                            old_blob_table_end - WIM_HEADER_DISK_SIZE,
2364                                            &old_integrity_table);
2365                 /* If we couldn't read the old integrity table, we can still
2366                  * re-calculate the full integrity table ourselves.  Hence the
2367                  * ignoring of the return value.  */
2368         }
2369
2370         /* Write blob table if needed.  */
2371         if (!(write_flags & WIMLIB_WRITE_FLAG_NO_NEW_BLOBS)) {
2372                 ret = write_blob_table(wim, image, write_flags,
2373                                        blob_table_list);
2374                 if (ret) {
2375                         free_integrity_table(old_integrity_table);
2376                         return ret;
2377                 }
2378         }
2379
2380         /* Write XML data.  */
2381         xml_totalbytes = wim->out_fd.offset;
2382         if (write_flags & WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES)
2383                 xml_totalbytes = WIM_TOTALBYTES_USE_EXISTING;
2384         ret = write_wim_xml_data(wim, image, xml_totalbytes,
2385                                  &wim->out_hdr.xml_data_reshdr,
2386                                  write_resource_flags);
2387         if (ret) {
2388                 free_integrity_table(old_integrity_table);
2389                 return ret;
2390         }
2391
2392         /* Write integrity table if needed.  */
2393         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
2394                 if (write_flags & WIMLIB_WRITE_FLAG_NO_NEW_BLOBS) {
2395                         /* The XML data we wrote may have overwritten part of
2396                          * the old integrity table, so while calculating the new
2397                          * integrity table we should temporarily update the WIM
2398                          * header to remove the integrity table reference.   */
2399                         struct wim_header checkpoint_hdr;
2400                         memcpy(&checkpoint_hdr, &wim->out_hdr, sizeof(struct wim_header));
2401                         zero_reshdr(&checkpoint_hdr.integrity_table_reshdr);
2402                         checkpoint_hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2403                         ret = write_wim_header(&checkpoint_hdr, &wim->out_fd, 0);
2404                         if (ret) {
2405                                 free_integrity_table(old_integrity_table);
2406                                 return ret;
2407                         }
2408                 }
2409
2410                 new_blob_table_end = wim->out_hdr.blob_table_reshdr.offset_in_wim +
2411                                      wim->out_hdr.blob_table_reshdr.size_in_wim;
2412
2413                 ret = write_integrity_table(wim,
2414                                             new_blob_table_end,
2415                                             old_blob_table_end,
2416                                             old_integrity_table);
2417                 free_integrity_table(old_integrity_table);
2418                 if (ret)
2419                         return ret;
2420         } else {
2421                 /* No integrity table.  */
2422                 zero_reshdr(&wim->out_hdr.integrity_table_reshdr);
2423         }
2424
2425         /* Now that all information in the WIM header has been determined, the
2426          * preliminary header written earlier can be overwritten, the header of
2427          * the existing WIM file can be overwritten, or the final header can be
2428          * written to the end of the pipable WIM.  */
2429         wim->out_hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2430         if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
2431                 ret = write_wim_header(&wim->out_hdr, &wim->out_fd, wim->out_fd.offset);
2432         else
2433                 ret = write_wim_header(&wim->out_hdr, &wim->out_fd, 0);
2434         if (ret)
2435                 return ret;
2436
2437         if (unlikely(write_flags & WIMLIB_WRITE_FLAG_UNSAFE_COMPACT)) {
2438                 /* Truncate any data the compaction freed up.  */
2439                 if (ftruncate(wim->out_fd.fd, wim->out_fd.offset)) {
2440                         ERROR_WITH_ERRNO("Failed to truncate the output WIM file");
2441                         return WIMLIB_ERR_WRITE;
2442                 }
2443         }
2444
2445         /* Possibly sync file data to disk before closing.  On POSIX systems, it
2446          * is necessary to do this before using rename() to overwrite an
2447          * existing file with a new file.  Otherwise, data loss would occur if
2448          * the system is abruptly terminated when the metadata for the rename
2449          * operation has been written to disk, but the new file data has not.
2450          */
2451         if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
2452                 if (fsync(wim->out_fd.fd)) {
2453                         ERROR_WITH_ERRNO("Error syncing data to WIM file");
2454                         return WIMLIB_ERR_WRITE;
2455                 }
2456         }
2457
2458         if (close_wim_writable(wim, write_flags)) {
2459                 ERROR_WITH_ERRNO("Failed to close the output WIM file");
2460                 return WIMLIB_ERR_WRITE;
2461         }
2462
2463         return 0;
2464 }
2465
2466 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
2467
2468 /* Set advisory lock on WIM file (if not already done so)  */
2469 int
2470 lock_wim_for_append(WIMStruct *wim)
2471 {
2472         if (wim->locked_for_append)
2473                 return 0;
2474         if (!flock(wim->in_fd.fd, LOCK_EX | LOCK_NB)) {
2475                 wim->locked_for_append = 1;
2476                 return 0;
2477         }
2478         if (errno != EWOULDBLOCK)
2479                 return 0;
2480         return WIMLIB_ERR_ALREADY_LOCKED;
2481 }
2482
2483 /* Remove advisory lock on WIM file (if present)  */
2484 void
2485 unlock_wim_for_append(WIMStruct *wim)
2486 {
2487         if (wim->locked_for_append) {
2488                 flock(wim->in_fd.fd, LOCK_UN);
2489                 wim->locked_for_append = 0;
2490         }
2491 }
2492 #endif
2493
2494 /*
2495  * write_pipable_wim():
2496  *
2497  * Perform the intermediate stages of creating a "pipable" WIM (i.e. a WIM
2498  * capable of being applied from a pipe).
2499  *
2500  * Pipable WIMs are a wimlib-specific modification of the WIM format such that
2501  * images can be applied from them sequentially when the file data is sent over
2502  * a pipe.  In addition, a pipable WIM can be written sequentially to a pipe.
2503  * The modifications made to the WIM format for pipable WIMs are:
2504  *
2505  * - Magic characters in header are "WLPWM\0\0\0" (wimlib pipable WIM) instead
2506  *   of "MSWIM\0\0\0".  This lets wimlib know that the WIM is pipable and also
2507  *   stops other software from trying to read the file as a normal WIM.
2508  *
2509  * - The header at the beginning of the file does not contain all the normal
2510  *   information; in particular it will have all 0's for the blob table and XML
2511  *   data resource entries.  This is because this information cannot be
2512  *   determined until the blob table and XML data have been written.
2513  *   Consequently, wimlib will write the full header at the very end of the
2514  *   file.  The header at the end, however, is only used when reading the WIM
2515  *   from a seekable file (not a pipe).
2516  *
2517  * - An extra copy of the XML data is placed directly after the header.  This
2518  *   allows image names and sizes to be determined at an appropriate time when
2519  *   reading the WIM from a pipe.  This copy of the XML data is ignored if the
2520  *   WIM is read from a seekable file (not a pipe).
2521  *
2522  * - Solid resources are not allowed.  Each blob is always stored in its own
2523  *   resource.
2524  *
2525  * - The format of resources, or blobs, has been modified to allow them to be
2526  *   used before the "blob table" has been read.  Each blob is prefixed with a
2527  *   `struct pwm_blob_hdr' that is basically an abbreviated form of `struct
2528  *   blob_descriptor_disk' that only contains the SHA-1 message digest,
2529  *   uncompressed blob size, and flags that indicate whether the blob is
2530  *   compressed.  The data of uncompressed blobs then follows literally, while
2531  *   the data of compressed blobs follows in a modified format.  Compressed
2532  *   blobs do not begin with a chunk table, since the chunk table cannot be
2533  *   written until all chunks have been compressed.  Instead, each compressed
2534  *   chunk is prefixed by a `struct pwm_chunk_hdr' that gives its size.
2535  *   Furthermore, the chunk table is written at the end of the resource instead
2536  *   of the start.  Note: chunk offsets are given in the chunk table as if the
2537  *   `struct pwm_chunk_hdr's were not present; also, the chunk table is only
2538  *   used if the WIM is being read from a seekable file (not a pipe).
2539  *
2540  * - Metadata blobs always come before non-metadata blobs.  (This does not by
2541  *   itself constitute an incompatibility with normal WIMs, since this is valid
2542  *   in normal WIMs.)
2543  *
2544  * - At least up to the end of the blobs, all components must be packed as
2545  *   tightly as possible; there cannot be any "holes" in the WIM.  (This does
2546  *   not by itself consititute an incompatibility with normal WIMs, since this
2547  *   is valid in normal WIMs.)
2548  *
2549  * Note: the blob table, XML data, and header at the end are not used when
2550  * applying from a pipe.  They exist to support functionality such as image
2551  * application and export when the WIM is *not* read from a pipe.
2552  *
2553  *   Layout of pipable WIM:
2554  *
2555  * ---------+----------+--------------------+----------------+--------------+-----------+--------+
2556  * | Header | XML data | Metadata resources | File resources |  Blob table  | XML data  | Header |
2557  * ---------+----------+--------------------+----------------+--------------+-----------+--------+
2558  *
2559  *   Layout of normal WIM:
2560  *
2561  * +--------+-----------------------------+-------------------------+
2562  * | Header | File and metadata resources |  Blob table  | XML data |
2563  * +--------+-----------------------------+-------------------------+
2564  *
2565  * An optional integrity table can follow the final XML data in both normal and
2566  * pipable WIMs.  However, due to implementation details, wimlib currently can
2567  * only include an integrity table in a pipable WIM when writing it to a
2568  * seekable file (not a pipe).
2569  *
2570  * Do note that since pipable WIMs are not supported by Microsoft's software,
2571  * wimlib does not create them unless explicitly requested (with
2572  * WIMLIB_WRITE_FLAG_PIPABLE) and as stated above they use different magic
2573  * characters to identify the file.
2574  */
2575 static int
2576 write_pipable_wim(WIMStruct *wim, int image, int write_flags,
2577                   unsigned num_threads,
2578                   struct list_head *blob_list_override,
2579                   struct list_head *blob_table_list_ret)
2580 {
2581         int ret;
2582         struct wim_reshdr xml_reshdr;
2583
2584         WARNING("Creating a pipable WIM, which will "
2585                 "be incompatible\n"
2586                 "          with Microsoft's software (WIMGAPI/ImageX/DISM).");
2587
2588         /* At this point, the header at the beginning of the file has already
2589          * been written.  */
2590
2591         /* For efficiency, when wimlib adds an image to the WIM with
2592          * wimlib_add_image(), the SHA-1 message digests of files are not
2593          * calculated; instead, they are calculated while the files are being
2594          * written.  However, this does not work when writing a pipable WIM,
2595          * since when writing a blob to a pipable WIM, its SHA-1 message digest
2596          * needs to be known before the blob data is written.  Therefore, before
2597          * getting much farther, we need to pre-calculate the SHA-1 message
2598          * digests of all blobs that will be written.  */
2599         ret = wim_checksum_unhashed_blobs(wim);
2600         if (ret)
2601                 return ret;
2602
2603         /* Write extra copy of the XML data.  */
2604         ret = write_wim_xml_data(wim, image, WIM_TOTALBYTES_OMIT,
2605                                  &xml_reshdr, WRITE_RESOURCE_FLAG_PIPABLE);
2606         if (ret)
2607                 return ret;
2608
2609         /* Write metadata resources for the image(s) being included in the
2610          * output WIM.  */
2611         ret = write_metadata_resources(wim, image, write_flags);
2612         if (ret)
2613                 return ret;
2614
2615         /* Write file data needed for the image(s) being included in the output
2616          * WIM, or file data needed for the split WIM part.  */
2617         return write_file_data(wim, image, write_flags,
2618                                num_threads, blob_list_override,
2619                                blob_table_list_ret);
2620
2621         /* The blob table, XML data, and header at end are handled by
2622          * finish_write().  */
2623 }
2624
2625 static bool
2626 should_default_to_solid_compression(WIMStruct *wim, int write_flags)
2627 {
2628         return wim->out_hdr.wim_version == WIM_VERSION_SOLID &&
2629                 !(write_flags & (WIMLIB_WRITE_FLAG_SOLID |
2630                                  WIMLIB_WRITE_FLAG_PIPABLE)) &&
2631                 wim_has_solid_resources(wim);
2632 }
2633
2634 /* Write a standalone WIM or split WIM (SWM) part to a new file or to a file
2635  * descriptor.  */
2636 int
2637 write_wim_part(WIMStruct *wim,
2638                const void *path_or_fd,
2639                int image,
2640                int write_flags,
2641                unsigned num_threads,
2642                unsigned part_number,
2643                unsigned total_parts,
2644                struct list_head *blob_list_override,
2645                const u8 *guid)
2646 {
2647         int ret;
2648         struct list_head blob_table_list;
2649
2650         /* Internally, this is always called with a valid part number and total
2651          * parts.  */
2652         wimlib_assert(total_parts >= 1);
2653         wimlib_assert(part_number >= 1 && part_number <= total_parts);
2654
2655         /* A valid image (or all images) must be specified.  */
2656         if (image != WIMLIB_ALL_IMAGES &&
2657              (image < 1 || image > wim->hdr.image_count))
2658                 return WIMLIB_ERR_INVALID_IMAGE;
2659
2660         /* If we need to write metadata resources, make sure the ::WIMStruct has
2661          * the needed information attached (e.g. is not a resource-only WIM,
2662          * such as a non-first part of a split WIM).  */
2663         if (!wim_has_metadata(wim) &&
2664             !(write_flags & WIMLIB_WRITE_FLAG_NO_METADATA))
2665                 return WIMLIB_ERR_METADATA_NOT_FOUND;
2666
2667         /* Check for contradictory flags.  */
2668         if ((write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2669                             WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY))
2670                                 == (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2671                                     WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY))
2672                 return WIMLIB_ERR_INVALID_PARAM;
2673
2674         if ((write_flags & (WIMLIB_WRITE_FLAG_PIPABLE |
2675                             WIMLIB_WRITE_FLAG_NOT_PIPABLE))
2676                                 == (WIMLIB_WRITE_FLAG_PIPABLE |
2677                                     WIMLIB_WRITE_FLAG_NOT_PIPABLE))
2678                 return WIMLIB_ERR_INVALID_PARAM;
2679
2680         /* Only wimlib_overwrite() accepts UNSAFE_COMPACT.  */
2681         if (write_flags & WIMLIB_WRITE_FLAG_UNSAFE_COMPACT)
2682                 return WIMLIB_ERR_INVALID_PARAM;
2683
2684         /* Include an integrity table by default if no preference was given and
2685          * the WIM already had an integrity table.  */
2686         if (!(write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2687                              WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY))) {
2688                 if (wim_has_integrity_table(wim))
2689                         write_flags |= WIMLIB_WRITE_FLAG_CHECK_INTEGRITY;
2690         }
2691
2692         /* Write a pipable WIM by default if no preference was given and the WIM
2693          * was already pipable.  */
2694         if (!(write_flags & (WIMLIB_WRITE_FLAG_PIPABLE |
2695                              WIMLIB_WRITE_FLAG_NOT_PIPABLE))) {
2696                 if (wim_is_pipable(wim))
2697                         write_flags |= WIMLIB_WRITE_FLAG_PIPABLE;
2698         }
2699
2700         if ((write_flags & (WIMLIB_WRITE_FLAG_PIPABLE |
2701                             WIMLIB_WRITE_FLAG_SOLID))
2702                                     == (WIMLIB_WRITE_FLAG_PIPABLE |
2703                                         WIMLIB_WRITE_FLAG_SOLID))
2704         {
2705                 ERROR("Solid compression is unsupported in pipable WIMs");
2706                 return WIMLIB_ERR_INVALID_PARAM;
2707         }
2708
2709         /* Start initializing the new file header.  */
2710         memset(&wim->out_hdr, 0, sizeof(wim->out_hdr));
2711
2712         /* Set the magic number.  */
2713         if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
2714                 wim->out_hdr.magic = PWM_MAGIC;
2715         else
2716                 wim->out_hdr.magic = WIM_MAGIC;
2717
2718         /* Set the version number.  */
2719         if ((write_flags & WIMLIB_WRITE_FLAG_SOLID) ||
2720             wim->out_compression_type == WIMLIB_COMPRESSION_TYPE_LZMS)
2721                 wim->out_hdr.wim_version = WIM_VERSION_SOLID;
2722         else
2723                 wim->out_hdr.wim_version = WIM_VERSION_DEFAULT;
2724
2725         /* Default to solid compression if it is valid in the chosen WIM file
2726          * format and the WIMStruct references any solid resources.  This is
2727          * useful when exporting an image from a solid WIM.  */
2728         if (should_default_to_solid_compression(wim, write_flags))
2729                 write_flags |= WIMLIB_WRITE_FLAG_SOLID;
2730
2731         /* Set the header flags.  */
2732         wim->out_hdr.flags = (wim->hdr.flags & (WIM_HDR_FLAG_RP_FIX |
2733                                                 WIM_HDR_FLAG_READONLY));
2734         if (total_parts != 1)
2735                 wim->out_hdr.flags |= WIM_HDR_FLAG_SPANNED;
2736         if (wim->out_compression_type != WIMLIB_COMPRESSION_TYPE_NONE) {
2737                 wim->out_hdr.flags |= WIM_HDR_FLAG_COMPRESSION;
2738                 switch (wim->out_compression_type) {
2739                 case WIMLIB_COMPRESSION_TYPE_XPRESS:
2740                         wim->out_hdr.flags |= WIM_HDR_FLAG_COMPRESS_XPRESS;
2741                         break;
2742                 case WIMLIB_COMPRESSION_TYPE_LZX:
2743                         wim->out_hdr.flags |= WIM_HDR_FLAG_COMPRESS_LZX;
2744                         break;
2745                 case WIMLIB_COMPRESSION_TYPE_LZMS:
2746                         wim->out_hdr.flags |= WIM_HDR_FLAG_COMPRESS_LZMS;
2747                         break;
2748                 }
2749         }
2750
2751         /* Set the chunk size.  */
2752         wim->out_hdr.chunk_size = wim->out_chunk_size;
2753
2754         /* Set the GUID.  */
2755         if (write_flags & WIMLIB_WRITE_FLAG_RETAIN_GUID)
2756                 guid = wim->hdr.guid;
2757         if (guid)
2758                 copy_guid(wim->out_hdr.guid, guid);
2759         else
2760                 generate_guid(wim->out_hdr.guid);
2761
2762         /* Set the part number and total parts.  */
2763         wim->out_hdr.part_number = part_number;
2764         wim->out_hdr.total_parts = total_parts;
2765
2766         /* Set the image count.  */
2767         if (image == WIMLIB_ALL_IMAGES)
2768                 wim->out_hdr.image_count = wim->hdr.image_count;
2769         else
2770                 wim->out_hdr.image_count = 1;
2771
2772         /* Set the boot index.  */
2773         wim->out_hdr.boot_idx = 0;
2774         if (total_parts == 1) {
2775                 if (image == WIMLIB_ALL_IMAGES)
2776                         wim->out_hdr.boot_idx = wim->hdr.boot_idx;
2777                 else if (image == wim->hdr.boot_idx)
2778                         wim->out_hdr.boot_idx = 1;
2779         }
2780
2781         /* Set up the output file descriptor.  */
2782         if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR) {
2783                 /* File descriptor was explicitly provided.  */
2784                 filedes_init(&wim->out_fd, *(const int *)path_or_fd);
2785                 if (!filedes_is_seekable(&wim->out_fd)) {
2786                         /* The file descriptor is a pipe.  */
2787                         ret = WIMLIB_ERR_INVALID_PARAM;
2788                         if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE))
2789                                 goto out_cleanup;
2790                         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
2791                                 ERROR("Can't include integrity check when "
2792                                       "writing pipable WIM to pipe!");
2793                                 goto out_cleanup;
2794                         }
2795                 }
2796         } else {
2797                 /* Filename of WIM to write was provided; open file descriptor
2798                  * to it.  */
2799                 ret = open_wim_writable(wim, (const tchar*)path_or_fd,
2800                                         O_TRUNC | O_CREAT | O_RDWR);
2801                 if (ret)
2802                         goto out_cleanup;
2803         }
2804
2805         /* Write initial header.  This is merely a "dummy" header since it
2806          * doesn't have resource entries filled in yet, so it will be
2807          * overwritten later (unless writing a pipable WIM).  */
2808         if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE))
2809                 wim->out_hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2810         ret = write_wim_header(&wim->out_hdr, &wim->out_fd, wim->out_fd.offset);
2811         wim->out_hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2812         if (ret)
2813                 goto out_cleanup;
2814
2815         /* Write file data and metadata resources.  */
2816         if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE)) {
2817                 /* Default case: create a normal (non-pipable) WIM.  */
2818                 ret = write_file_data(wim, image, write_flags,
2819                                       num_threads,
2820                                       blob_list_override,
2821                                       &blob_table_list);
2822                 if (ret)
2823                         goto out_cleanup;
2824
2825                 ret = write_metadata_resources(wim, image, write_flags);
2826                 if (ret)
2827                         goto out_cleanup;
2828         } else {
2829                 /* Non-default case: create pipable WIM.  */
2830                 ret = write_pipable_wim(wim, image, write_flags, num_threads,
2831                                         blob_list_override,
2832                                         &blob_table_list);
2833                 if (ret)
2834                         goto out_cleanup;
2835         }
2836
2837         /* Write blob table, XML data, and (optional) integrity table.  */
2838         ret = finish_write(wim, image, write_flags, &blob_table_list);
2839 out_cleanup:
2840         (void)close_wim_writable(wim, write_flags);
2841         return ret;
2842 }
2843
2844 /* Write a standalone WIM to a file or file descriptor.  */
2845 static int
2846 write_standalone_wim(WIMStruct *wim, const void *path_or_fd,
2847                      int image, int write_flags, unsigned num_threads)
2848 {
2849         return write_wim_part(wim, path_or_fd, image, write_flags,
2850                               num_threads, 1, 1, NULL, NULL);
2851 }
2852
2853 /* API function documented in wimlib.h  */
2854 WIMLIBAPI int
2855 wimlib_write(WIMStruct *wim, const tchar *path,
2856              int image, int write_flags, unsigned num_threads)
2857 {
2858         if (write_flags & ~WIMLIB_WRITE_MASK_PUBLIC)
2859                 return WIMLIB_ERR_INVALID_PARAM;
2860
2861         if (path == NULL || path[0] == T('\0'))
2862                 return WIMLIB_ERR_INVALID_PARAM;
2863
2864         return write_standalone_wim(wim, path, image, write_flags, num_threads);
2865 }
2866
2867 /* API function documented in wimlib.h  */
2868 WIMLIBAPI int
2869 wimlib_write_to_fd(WIMStruct *wim, int fd,
2870                    int image, int write_flags, unsigned num_threads)
2871 {
2872         if (write_flags & ~WIMLIB_WRITE_MASK_PUBLIC)
2873                 return WIMLIB_ERR_INVALID_PARAM;
2874
2875         if (fd < 0)
2876                 return WIMLIB_ERR_INVALID_PARAM;
2877
2878         write_flags |= WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR;
2879
2880         return write_standalone_wim(wim, &fd, image, write_flags, num_threads);
2881 }
2882
2883 static bool
2884 any_images_modified(WIMStruct *wim)
2885 {
2886         for (int i = 0; i < wim->hdr.image_count; i++)
2887                 if (wim->image_metadata[i]->modified)
2888                         return true;
2889         return false;
2890 }
2891
2892 static int
2893 check_resource_offset(struct blob_descriptor *blob, void *_wim)
2894 {
2895         const WIMStruct *wim = _wim;
2896         off_t end_offset = *(const off_t*)wim->private;
2897
2898         if (blob->blob_location == BLOB_IN_WIM &&
2899             blob->rdesc->wim == wim &&
2900             blob->rdesc->offset_in_wim + blob->rdesc->size_in_wim > end_offset)
2901                 return WIMLIB_ERR_RESOURCE_ORDER;
2902         return 0;
2903 }
2904
2905 /* Make sure no file or metadata resources are located after the XML data (or
2906  * integrity table if present)--- otherwise we can't safely append to the WIM
2907  * file and we return WIMLIB_ERR_RESOURCE_ORDER.  */
2908 static int
2909 check_resource_offsets(WIMStruct *wim, off_t end_offset)
2910 {
2911         int ret;
2912         unsigned i;
2913
2914         wim->private = &end_offset;
2915         ret = for_blob_in_table(wim->blob_table, check_resource_offset, wim);
2916         if (ret)
2917                 return ret;
2918
2919         for (i = 0; i < wim->hdr.image_count; i++) {
2920                 ret = check_resource_offset(wim->image_metadata[i]->metadata_blob, wim);
2921                 if (ret)
2922                         return ret;
2923         }
2924         return 0;
2925 }
2926
2927 /*
2928  * Overwrite a WIM, possibly appending new resources to it.
2929  *
2930  * A WIM looks like (or is supposed to look like) the following:
2931  *
2932  *                   Header (212 bytes)
2933  *                   Resources for metadata and files (variable size)
2934  *                   Blob table (variable size)
2935  *                   XML data (variable size)
2936  *                   Integrity table (optional) (variable size)
2937  *
2938  * If we are not adding any new files or metadata, then the blob table is
2939  * unchanged--- so we only need to overwrite the XML data, integrity table, and
2940  * header.  This operation is potentially unsafe if the program is abruptly
2941  * terminated while the XML data or integrity table are being overwritten, but
2942  * before the new header has been written.  To partially alleviate this problem,
2943  * we write a temporary header after the XML data has been written.  This may
2944  * prevent the WIM from becoming corrupted if the program is terminated while
2945  * the integrity table is being calculated (but no guarantees, due to write
2946  * re-ordering...).
2947  *
2948  * If we are adding new blobs, including new file data as well as any metadata
2949  * for any new images, then the blob table needs to be changed, and those blobs
2950  * need to be written.  In this case, we try to perform a safe update of the WIM
2951  * file by writing the blobs *after* the end of the previous WIM, then writing
2952  * the new blob table, XML data, and (optionally) integrity table following the
2953  * new blobs.  This will produce a layout like the following:
2954  *
2955  *                   Header (212 bytes)
2956  *                   (OLD) Resources for metadata and files (variable size)
2957  *                   (OLD) Blob table (variable size)
2958  *                   (OLD) XML data (variable size)
2959  *                   (OLD) Integrity table (optional) (variable size)
2960  *                   (NEW) Resources for metadata and files (variable size)
2961  *                   (NEW) Blob table (variable size)
2962  *                   (NEW) XML data (variable size)
2963  *                   (NEW) Integrity table (optional) (variable size)
2964  *
2965  * At all points, the WIM is valid as nothing points to the new data yet.  Then,
2966  * the header is overwritten to point to the new blob table, XML data, and
2967  * integrity table, to produce the following layout:
2968  *
2969  *                   Header (212 bytes)
2970  *                   Resources for metadata and files (variable size)
2971  *                   Nothing (variable size)
2972  *                   Resources for metadata and files (variable size)
2973  *                   Blob table (variable size)
2974  *                   XML data (variable size)
2975  *                   Integrity table (optional) (variable size)
2976  *
2977  * This function allows an image to be appended to a large WIM very quickly, and
2978  * is crash-safe except in the case of write re-ordering, but the disadvantage
2979  * is that a small hole is left in the WIM where the old blob table, xml data,
2980  * and integrity table were.  (These usually only take up a small amount of
2981  * space compared to the blobs, however.)
2982  *
2983  * Finally, this function also supports "compaction" overwrites as an
2984  * alternative to the normal "append" overwrites described above.  In a
2985  * compaction, data is written starting immediately from the end of the header.
2986  * All existing resources are written first, in order by file offset.  New
2987  * resources are written afterwards, and at the end any extra data is truncated
2988  * from the file.  The advantage of this approach is that is that the WIM file
2989  * ends up fully optimized, without any holes remaining.  The main disadavantage
2990  * is that this operation is fundamentally unsafe and cannot be interrupted
2991  * without data corruption.  Consequently, compactions are only ever done when
2992  * explicitly requested by the library user with the flag
2993  * WIMLIB_WRITE_FLAG_UNSAFE_COMPACT.  (Another disadvantage is that a compaction
2994  * can be much slower than an append.)
2995  */
2996 static int
2997 overwrite_wim_inplace(WIMStruct *wim, int write_flags, unsigned num_threads)
2998 {
2999         int ret;
3000         off_t old_wim_end;
3001         struct list_head blob_list;
3002         struct list_head blob_table_list;
3003         struct filter_context filter_ctx;
3004
3005         /* Include an integrity table by default if no preference was given and
3006          * the WIM already had an integrity table.  */
3007         if (!(write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
3008                              WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY)))
3009                 if (wim_has_integrity_table(wim))
3010                         write_flags |= WIMLIB_WRITE_FLAG_CHECK_INTEGRITY;
3011
3012         /* Start preparing the updated file header.  */
3013         memcpy(&wim->out_hdr, &wim->hdr, sizeof(wim->out_hdr));
3014
3015         /* If using solid compression, the version number must be set to
3016          * WIM_VERSION_SOLID.  */
3017         if (write_flags & WIMLIB_WRITE_FLAG_SOLID)
3018                 wim->out_hdr.wim_version = WIM_VERSION_SOLID;
3019
3020         /* Default to solid compression if it is valid in the chosen WIM file
3021          * format and the WIMStruct references any solid resources.  This is
3022          * useful when updating a solid WIM.  */
3023         if (should_default_to_solid_compression(wim, write_flags))
3024                 write_flags |= WIMLIB_WRITE_FLAG_SOLID;
3025
3026         if (unlikely(write_flags & WIMLIB_WRITE_FLAG_UNSAFE_COMPACT)) {
3027
3028                 /* In-place compaction  */
3029
3030                 WARNING("The WIM file \"%"TS"\" is being compacted in place.\n"
3031                         "          Do *not* interrupt the operation, or else "
3032                         "the WIM file will be\n"
3033                         "          corrupted!", wim->filename);
3034                 wim->being_compacted = 1;
3035                 old_wim_end = WIM_HEADER_DISK_SIZE;
3036
3037                 ret = prepare_blob_list_for_write(wim, WIMLIB_ALL_IMAGES,
3038                                                   write_flags, &blob_list,
3039                                                   &blob_table_list, &filter_ctx);
3040                 if (ret)
3041                         goto out;
3042
3043                 if (wim_has_metadata(wim)) {
3044                         /* Add existing metadata resources to be compacted along
3045                          * with the file resources.  */
3046                         for (int i = 0; i < wim->hdr.image_count; i++) {
3047                                 struct wim_image_metadata *imd = wim->image_metadata[i];
3048                                 if (!imd->modified) {
3049                                         fully_reference_blob_for_write(imd->metadata_blob,
3050                                                                        &blob_list);
3051                                 }
3052                         }
3053                 }
3054         } else {
3055                 u64 old_blob_table_end, old_xml_begin, old_xml_end;
3056
3057                 /* Set additional flags for append.  */
3058                 write_flags |= WIMLIB_WRITE_FLAG_APPEND |
3059                                WIMLIB_WRITE_FLAG_STREAMS_OK;
3060
3061                 /* Make sure there is no data after the XML data, except
3062                  * possibily an integrity table.  If this were the case, then
3063                  * this data would be overwritten.  */
3064                 old_xml_begin = wim->hdr.xml_data_reshdr.offset_in_wim;
3065                 old_xml_end = old_xml_begin + wim->hdr.xml_data_reshdr.size_in_wim;
3066                 old_blob_table_end = wim->hdr.blob_table_reshdr.offset_in_wim +
3067                                      wim->hdr.blob_table_reshdr.size_in_wim;
3068                 if (wim_has_integrity_table(wim) &&
3069                     wim->hdr.integrity_table_reshdr.offset_in_wim < old_xml_end) {
3070                         WARNING("Didn't expect the integrity table to be "
3071                                 "before the XML data");
3072                         ret = WIMLIB_ERR_RESOURCE_ORDER;
3073                         goto out;
3074                 }
3075
3076                 if (old_blob_table_end > old_xml_begin) {
3077                         WARNING("Didn't expect the blob table to be after "
3078                                 "the XML data");
3079                         ret = WIMLIB_ERR_RESOURCE_ORDER;
3080                         goto out;
3081                 }
3082                 /* Set @old_wim_end, which indicates the point beyond which we
3083                  * don't allow any file and metadata resources to appear without
3084                  * returning WIMLIB_ERR_RESOURCE_ORDER (due to the fact that we
3085                  * would otherwise overwrite these resources). */
3086                 if (!wim->image_deletion_occurred && !any_images_modified(wim)) {
3087                         /* If no images have been modified and no images have
3088                          * been deleted, a new blob table does not need to be
3089                          * written.  We shall write the new XML data and
3090                          * optional integrity table immediately after the blob
3091                          * table.  Note that this may overwrite an existing
3092                          * integrity table. */
3093                         old_wim_end = old_blob_table_end;
3094                         write_flags |= WIMLIB_WRITE_FLAG_NO_NEW_BLOBS;
3095                 } else if (wim_has_integrity_table(wim)) {
3096                         /* Old WIM has an integrity table; begin writing new
3097                          * blobs after it. */
3098                         old_wim_end = wim->hdr.integrity_table_reshdr.offset_in_wim +
3099                                       wim->hdr.integrity_table_reshdr.size_in_wim;
3100                 } else {
3101                         /* No existing integrity table; begin writing new blobs
3102                          * after the old XML data. */
3103                         old_wim_end = old_xml_end;
3104                 }
3105
3106                 ret = check_resource_offsets(wim, old_wim_end);
3107                 if (ret)
3108                         goto out;
3109
3110                 ret = prepare_blob_list_for_write(wim, WIMLIB_ALL_IMAGES,
3111                                                   write_flags, &blob_list,
3112                                                   &blob_table_list, &filter_ctx);
3113                 if (ret)
3114                         goto out;
3115
3116                 if (write_flags & WIMLIB_WRITE_FLAG_NO_NEW_BLOBS)
3117                         wimlib_assert(list_empty(&blob_list));
3118         }
3119
3120         ret = open_wim_writable(wim, wim->filename, O_RDWR);
3121         if (ret)
3122                 goto out;
3123
3124         ret = lock_wim_for_append(wim);
3125         if (ret)
3126                 goto out_close_wim;
3127
3128         /* Set WIM_HDR_FLAG_WRITE_IN_PROGRESS flag in header. */
3129         wim->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
3130         ret = write_wim_header_flags(wim->hdr.flags, &wim->out_fd);
3131         wim->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
3132         if (ret) {
3133                 ERROR_WITH_ERRNO("Error updating WIM header flags");
3134                 goto out_unlock_wim;
3135         }
3136
3137         if (filedes_seek(&wim->out_fd, old_wim_end) == -1) {
3138                 ERROR_WITH_ERRNO("Can't seek to end of WIM");
3139                 ret = WIMLIB_ERR_WRITE;
3140                 goto out_restore_hdr;
3141         }
3142
3143         ret = write_file_data_blobs(wim, &blob_list, write_flags,
3144                                     num_threads, &filter_ctx);
3145         if (ret)
3146                 goto out_truncate;
3147
3148         ret = write_metadata_resources(wim, WIMLIB_ALL_IMAGES, write_flags);
3149         if (ret)
3150                 goto out_truncate;
3151
3152         ret = finish_write(wim, WIMLIB_ALL_IMAGES, write_flags,
3153                            &blob_table_list);
3154         if (ret)
3155                 goto out_truncate;
3156
3157         unlock_wim_for_append(wim);
3158         return 0;
3159
3160 out_truncate:
3161         if (!(write_flags & (WIMLIB_WRITE_FLAG_NO_NEW_BLOBS |
3162                              WIMLIB_WRITE_FLAG_UNSAFE_COMPACT))) {
3163                 WARNING("Truncating \"%"TS"\" to its original size "
3164                         "(%"PRIu64" bytes)", wim->filename, old_wim_end);
3165                 /* Return value of ftruncate() is ignored because this is
3166                  * already an error path.  */
3167                 (void)ftruncate(wim->out_fd.fd, old_wim_end);
3168         }
3169 out_restore_hdr:
3170         (void)write_wim_header_flags(wim->hdr.flags, &wim->out_fd);
3171 out_unlock_wim:
3172         unlock_wim_for_append(wim);
3173 out_close_wim:
3174         (void)close_wim_writable(wim, write_flags);
3175 out:
3176         wim->being_compacted = 0;
3177         return ret;
3178 }
3179
3180 static int
3181 overwrite_wim_via_tmpfile(WIMStruct *wim, int write_flags, unsigned num_threads)
3182 {
3183         size_t wim_name_len;
3184         int ret;
3185
3186         /* Write the WIM to a temporary file in the same directory as the
3187          * original WIM. */
3188         wim_name_len = tstrlen(wim->filename);
3189         tchar tmpfile[wim_name_len + 10];
3190         tmemcpy(tmpfile, wim->filename, wim_name_len);
3191         randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
3192         tmpfile[wim_name_len + 9] = T('\0');
3193
3194         ret = wimlib_write(wim, tmpfile, WIMLIB_ALL_IMAGES,
3195                            write_flags |
3196                                 WIMLIB_WRITE_FLAG_FSYNC |
3197                                 WIMLIB_WRITE_FLAG_RETAIN_GUID,
3198                            num_threads);
3199         if (ret) {
3200                 tunlink(tmpfile);
3201                 return ret;
3202         }
3203
3204         if (filedes_valid(&wim->in_fd)) {
3205                 filedes_close(&wim->in_fd);
3206                 filedes_invalidate(&wim->in_fd);
3207         }
3208
3209         /* Rename the new WIM file to the original WIM file.  Note: on Windows
3210          * this actually calls win32_rename_replacement(), not _wrename(), so
3211          * that removing the existing destination file can be handled.  */
3212         ret = trename(tmpfile, wim->filename);
3213         if (ret) {
3214                 ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
3215                                  tmpfile, wim->filename);
3216         #ifdef __WIN32__
3217                 if (ret < 0)
3218         #endif
3219                 {
3220                         tunlink(tmpfile);
3221                 }
3222                 return WIMLIB_ERR_RENAME;
3223         }
3224
3225         union wimlib_progress_info progress;
3226         progress.rename.from = tmpfile;
3227         progress.rename.to = wim->filename;
3228         return call_progress(wim->progfunc, WIMLIB_PROGRESS_MSG_RENAME,
3229                              &progress, wim->progctx);
3230 }
3231
3232 /* Determine if the specified WIM file may be updated in-place rather than by
3233  * writing and replacing it with an entirely new file.  */
3234 static bool
3235 can_overwrite_wim_inplace(const WIMStruct *wim, int write_flags)
3236 {
3237         /* REBUILD flag forces full rebuild.  */
3238         if (write_flags & WIMLIB_WRITE_FLAG_REBUILD)
3239                 return false;
3240
3241         /* Image deletions cause full rebuild by default.  */
3242         if (wim->image_deletion_occurred &&
3243             !(write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
3244                 return false;
3245
3246         /* Pipable WIMs cannot be updated in place, nor can a non-pipable WIM be
3247          * turned into a pipable WIM in-place.  */
3248         if (wim_is_pipable(wim) || (write_flags & WIMLIB_WRITE_FLAG_PIPABLE))
3249                 return false;
3250
3251         /* The default compression type and compression chunk size selected for
3252          * the output WIM must be the same as those currently used for the WIM.
3253          */
3254         if (wim->compression_type != wim->out_compression_type)
3255                 return false;
3256         if (wim->chunk_size != wim->out_chunk_size)
3257                 return false;
3258
3259         return true;
3260 }
3261
3262 /* API function documented in wimlib.h  */
3263 WIMLIBAPI int
3264 wimlib_overwrite(WIMStruct *wim, int write_flags, unsigned num_threads)
3265 {
3266         int ret;
3267         u32 orig_hdr_flags;
3268
3269         if (write_flags & ~WIMLIB_WRITE_MASK_PUBLIC)
3270                 return WIMLIB_ERR_INVALID_PARAM;
3271
3272         if (!wim->filename)
3273                 return WIMLIB_ERR_NO_FILENAME;
3274
3275         if (unlikely(write_flags & WIMLIB_WRITE_FLAG_UNSAFE_COMPACT)) {
3276                 /*
3277                  * In UNSAFE_COMPACT mode:
3278                  *      - RECOMPRESS is forbidden
3279                  *      - REBUILD is ignored
3280                  *      - SOFT_DELETE and NO_SOLID_SORT are implied
3281                  */
3282                 if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
3283                         return WIMLIB_ERR_COMPACTION_NOT_POSSIBLE;
3284                 write_flags &= ~WIMLIB_WRITE_FLAG_REBUILD;
3285                 write_flags |= WIMLIB_WRITE_FLAG_SOFT_DELETE;
3286                 write_flags |= WIMLIB_WRITE_FLAG_NO_SOLID_SORT;
3287         }
3288
3289         orig_hdr_flags = wim->hdr.flags;
3290         if (write_flags & WIMLIB_WRITE_FLAG_IGNORE_READONLY_FLAG)
3291                 wim->hdr.flags &= ~WIM_HDR_FLAG_READONLY;
3292         ret = can_modify_wim(wim);
3293         wim->hdr.flags = orig_hdr_flags;
3294         if (ret)
3295                 return ret;
3296
3297         if (can_overwrite_wim_inplace(wim, write_flags)) {
3298                 ret = overwrite_wim_inplace(wim, write_flags, num_threads);
3299                 if (ret != WIMLIB_ERR_RESOURCE_ORDER)
3300                         return ret;
3301                 WARNING("Falling back to re-building entire WIM");
3302         }
3303         if (write_flags & WIMLIB_WRITE_FLAG_UNSAFE_COMPACT)
3304                 return WIMLIB_ERR_COMPACTION_NOT_POSSIBLE;
3305         return overwrite_wim_via_tmpfile(wim, write_flags, num_threads);
3306 }