wimlib.net Git - wimlib/blob - src/write.c

   1 /*
   2  * write.c
   3  *
   4  * Support for writing WIM files; write a WIM file, overwrite a WIM file, write
   5  * compressed file resources, etc.
   6  */
   7
   8 /*
   9  * Copyright (C) 2012, 2013 Eric Biggers
  10  *
  11  * This file is part of wimlib, a library for working with WIM files.
  12  *
  13  * wimlib is free software; you can redistribute it and/or modify it under the
  14  * terms of the GNU General Public License as published by the Free
  15  * Software Foundation; either version 3 of the License, or (at your option)
  16  * any later version.
  17  *
  18  * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
  19  * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
  20  * A PARTICULAR PURPOSE. See the GNU General Public License for more
  21  * details.
  22  *
  23  * You should have received a copy of the GNU General Public License
  24  * along with wimlib; if not, see http://www.gnu.org/licenses/.
  25  */
  26
  27 #include "config.h"
  28
  29 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
  30 /* On BSD, this should be included before "list.h" so that "list.h" can
  31  * overwrite the LIST_HEAD macro. */
  32 #include <sys/file.h>
  33 #endif
  34
  35 #include "list.h"
  36 #include "wimlib_internal.h"
  37 #include "buffer_io.h"
  38 #include "dentry.h"
  39 #include "lookup_table.h"
  40 #include "xml.h"
  41 #include "lzx.h"
  42 #include "xpress.h"
  43
  44 #ifdef ENABLE_MULTITHREADED_COMPRESSION
  45 #include <pthread.h>
  46 #endif
  47
  48 #include <unistd.h>
  49 #include <errno.h>
  50
  51 #ifdef WITH_NTFS_3G
  52 #include <time.h>
  53 #include <ntfs-3g/attrib.h>
  54 #include <ntfs-3g/inode.h>
  55 #include <ntfs-3g/dir.h>
  56 #endif
  57
  58 #ifdef HAVE_ALLOCA_H
  59 #include <alloca.h>
  60 #else
  61 #include <stdlib.h>
  62 #endif
  63
  64 static int fflush_and_ftruncate(FILE *fp, off_t size)
  65 {
  66         int ret;
  67
  68         ret = fflush(fp);
  69         if (ret != 0) {
  70                 ERROR_WITH_ERRNO("Failed to flush data to output WIM file");
  71                 return WIMLIB_ERR_WRITE;
  72         }
  73         ret = ftruncate(fileno(fp), size);
  74         if (ret != 0) {
  75                 ERROR_WITH_ERRNO("Failed to truncate output WIM file to "
  76                                  "%"PRIu64" bytes", size);
  77                 return WIMLIB_ERR_WRITE;
  78         }
  79         return 0;
  80 }
  81
  82 /* Chunk table that's located at the beginning of each compressed resource in
  83  * the WIM.  (This is not the on-disk format; the on-disk format just has an
  84  * array of offsets.) */
  85 struct chunk_table {
  86         off_t file_offset;
  87         u64 num_chunks;
  88         u64 original_resource_size;
  89         u64 bytes_per_chunk_entry;
  90         u64 table_disk_size;
  91         u64 cur_offset;
  92         u64 *cur_offset_p;
  93         u64 offsets[0];
  94 };
  95
  96 /*
  97  * Allocates and initializes a chunk table, and reserves space for it in the
  98  * output file.
  99  */
 100 static int
 101 begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
 102                              FILE *out_fp,
 103                              off_t file_offset,
 104                              struct chunk_table **chunk_tab_ret)
 105 {
 106         u64 size = wim_resource_size(lte);
 107         u64 num_chunks = (size + WIM_CHUNK_SIZE - 1) / WIM_CHUNK_SIZE;
 108         size_t alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
 109         struct chunk_table *chunk_tab = CALLOC(1, alloc_size);
 110         int ret;
 111
 112         if (!chunk_tab) {
 113                 ERROR("Failed to allocate chunk table for %"PRIu64" byte "
 114                       "resource", size);
 115                 ret = WIMLIB_ERR_NOMEM;
 116                 goto out;
 117         }
 118         chunk_tab->file_offset = file_offset;
 119         chunk_tab->num_chunks = num_chunks;
 120         chunk_tab->original_resource_size = size;
 121         chunk_tab->bytes_per_chunk_entry = (size >= (1ULL << 32)) ? 8 : 4;
 122         chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
 123                                      (num_chunks - 1);
 124         chunk_tab->cur_offset = 0;
 125         chunk_tab->cur_offset_p = chunk_tab->offsets;
 126
 127         if (fwrite(chunk_tab, 1, chunk_tab->table_disk_size, out_fp) !=
 128                    chunk_tab->table_disk_size) {
 129                 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
 130                                  "file resource");
 131                 ret = WIMLIB_ERR_WRITE;
 132                 goto out;
 133         }
 134
 135         ret = 0;
 136 out:
 137         *chunk_tab_ret = chunk_tab;
 138         return ret;
 139 }
 140
 141 /*
 142  * Pointer to function to compresses a chunk of a WIM resource.
 143  *
 144  * @chunk:              Uncompressed data of the chunk.
 145  * @chunk_size:         Size of the uncompressed chunk in bytes.
 146  * @compressed_chunk:   Pointer to output buffer of size at least
 147  *                              (@chunk_size - 1) bytes.
 148  * @compressed_chunk_len_ret:   Pointer to an unsigned int into which the size
 149  *                                      of the compressed chunk will be
 150  *                                      returned.
 151  *
 152  * Returns zero if compressed succeeded, and nonzero if the chunk could not be
 153  * compressed to any smaller than @chunk_size.  This function cannot fail for
 154  * any other reasons.
 155  */
 156 typedef int (*compress_func_t)(const void *, unsigned, void *, unsigned *);
 157
 158 compress_func_t get_compress_func(int out_ctype)
 159 {
 160         if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX)
 161                 return lzx_compress;
 162         else
 163                 return xpress_compress;
 164 }
 165
 166 /*
 167  * Writes a chunk of a WIM resource to an output file.
 168  *
 169  * @chunk:        Uncompressed data of the chunk.
 170  * @chunk_size:   Size of the chunk (<= WIM_CHUNK_SIZE)
 171  * @out_fp:       FILE * to write tho chunk to.
 172  * @out_ctype:    Compression type to use when writing the chunk (ignored if no
 173  *                      chunk table provided)
 174  * @chunk_tab:    Pointer to chunk table being created.  It is updated with the
 175  *                      offset of the chunk we write.
 176  *
 177  * Returns 0 on success; nonzero on failure.
 178  */
 179 static int write_wim_resource_chunk(const u8 chunk[], unsigned chunk_size,
 180                                     FILE *out_fp, compress_func_t compress,
 181                                     struct chunk_table *chunk_tab)
 182 {
 183         const u8 *out_chunk;
 184         unsigned out_chunk_size;
 185         if (chunk_tab) {
 186                 u8 *compressed_chunk = alloca(chunk_size);
 187                 int ret;
 188
 189                 ret = compress(chunk, chunk_size, compressed_chunk,
 190                                &out_chunk_size);
 191                 if (ret == 0) {
 192                         out_chunk = compressed_chunk;
 193                 } else {
 194                         out_chunk = chunk;
 195                         out_chunk_size = chunk_size;
 196                 }
 197                 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
 198                 chunk_tab->cur_offset += out_chunk_size;
 199         } else {
 200                 out_chunk = chunk;
 201                 out_chunk_size = chunk_size;
 202         }
 203         if (fwrite(out_chunk, 1, out_chunk_size, out_fp) != out_chunk_size) {
 204                 ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
 205                 return WIMLIB_ERR_WRITE;
 206         }
 207         return 0;
 208 }
 209
 210 /*
 211  * Finishes a WIM chunk table and writes it to the output file at the correct
 212  * offset.
 213  *
 214  * The final size of the full compressed resource is returned in the
 215  * @compressed_size_p.
 216  */
 217 static int
 218 finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
 219                               FILE *out_fp, u64 *compressed_size_p)
 220 {
 221         size_t bytes_written;
 222         if (fseeko(out_fp, chunk_tab->file_offset, SEEK_SET) != 0) {
 223                 ERROR_WITH_ERRNO("Failed to seek to byte %"PRIu64" of output "
 224                                  "WIM file", chunk_tab->file_offset);
 225                 return WIMLIB_ERR_WRITE;
 226         }
 227
 228         if (chunk_tab->bytes_per_chunk_entry == 8) {
 229                 array_cpu_to_le64(chunk_tab->offsets, chunk_tab->num_chunks);
 230         } else {
 231                 for (u64 i = 0; i < chunk_tab->num_chunks; i++)
 232                         ((u32*)chunk_tab->offsets)[i] =
 233                                 cpu_to_le32(chunk_tab->offsets[i]);
 234         }
 235         bytes_written = fwrite((u8*)chunk_tab->offsets +
 236                                         chunk_tab->bytes_per_chunk_entry,
 237                                1, chunk_tab->table_disk_size, out_fp);
 238         if (bytes_written != chunk_tab->table_disk_size) {
 239                 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
 240                                  "file resource");
 241                 return WIMLIB_ERR_WRITE;
 242         }
 243         if (fseeko(out_fp, 0, SEEK_END) != 0) {
 244                 ERROR_WITH_ERRNO("Failed to seek to end of output WIM file");
 245                 return WIMLIB_ERR_WRITE;
 246         }
 247         *compressed_size_p = chunk_tab->cur_offset + chunk_tab->table_disk_size;
 248         return 0;
 249 }
 250
 251 /* Prepare for multiple reads to a resource by caching a FILE * or NTFS
 252  * attribute pointer in the lookup table entry. */
 253 static int prepare_resource_for_read(struct wim_lookup_table_entry *lte
 254
 255                                         #ifdef WITH_NTFS_3G
 256                                         , ntfs_inode **ni_ret
 257                                         #endif
 258                 )
 259 {
 260         if (lte->resource_location == RESOURCE_IN_FILE_ON_DISK
 261              && !lte->file_on_disk_fp)
 262         {
 263                 wimlib_assert(lte->file_on_disk);
 264                 lte->file_on_disk_fp = fopen(lte->file_on_disk, "rb");
 265                 if (!lte->file_on_disk_fp) {
 266                         ERROR_WITH_ERRNO("Failed to open the file `%s' for "
 267                                          "reading", lte->file_on_disk);
 268                         return WIMLIB_ERR_OPEN;
 269                 }
 270         }
 271 #ifdef WITH_NTFS_3G
 272         else if (lte->resource_location == RESOURCE_IN_NTFS_VOLUME
 273                   && !lte->attr)
 274         {
 275                 struct ntfs_location *loc = lte->ntfs_loc;
 276                 ntfs_inode *ni;
 277                 wimlib_assert(loc);
 278                 ni = ntfs_pathname_to_inode(*loc->ntfs_vol_p, NULL, loc->path_utf8);
 279                 if (!ni) {
 280                         ERROR_WITH_ERRNO("Failed to open inode `%s' in NTFS "
 281                                          "volume", loc->path_utf8);
 282                         return WIMLIB_ERR_NTFS_3G;
 283                 }
 284                 lte->attr = ntfs_attr_open(ni,
 285                                            loc->is_reparse_point ? AT_REPARSE_POINT : AT_DATA,
 286                                            (ntfschar*)loc->stream_name_utf16,
 287                                            loc->stream_name_utf16_num_chars);
 288                 if (!lte->attr) {
 289                         ERROR_WITH_ERRNO("Failed to open attribute of `%s' in "
 290                                          "NTFS volume", loc->path_utf8);
 291                         ntfs_inode_close(ni);
 292                         return WIMLIB_ERR_NTFS_3G;
 293                 }
 294                 *ni_ret = ni;
 295         }
 296 #endif
 297         return 0;
 298 }
 299
 300 /* Undo prepare_resource_for_read() by closing the cached FILE * or NTFS
 301  * attribute. */
 302 static void end_wim_resource_read(struct wim_lookup_table_entry *lte
 303                                 #ifdef WITH_NTFS_3G
 304                                         , ntfs_inode *ni
 305                                 #endif
 306                                         )
 307 {
 308         if (lte->resource_location == RESOURCE_IN_FILE_ON_DISK
 309             && lte->file_on_disk_fp) {
 310                 fclose(lte->file_on_disk_fp);
 311                 lte->file_on_disk_fp = NULL;
 312         }
 313 #ifdef WITH_NTFS_3G
 314         else if (lte->resource_location == RESOURCE_IN_NTFS_VOLUME) {
 315                 if (lte->attr) {
 316                         ntfs_attr_close(lte->attr);
 317                         lte->attr = NULL;
 318                 }
 319                 if (ni)
 320                         ntfs_inode_close(ni);
 321         }
 322 #endif
 323 }
 324
 325 static int
 326 write_uncompressed_resource_and_truncate(struct wim_lookup_table_entry *lte,
 327                                          FILE *out_fp,
 328                                          off_t file_offset,
 329                                          struct resource_entry *out_res_entry)
 330 {
 331         int ret;
 332         if (fseeko(out_fp, file_offset, SEEK_SET) != 0) {
 333                 ERROR_WITH_ERRNO("Failed to seek to byte %"PRIu64" of "
 334                                  "output WIM file", file_offset);
 335                 return WIMLIB_ERR_WRITE;
 336         }
 337         ret = write_wim_resource(lte, out_fp, WIMLIB_COMPRESSION_TYPE_NONE,
 338                                  out_res_entry, 0);
 339         if (ret != 0)
 340                 return ret;
 341
 342         return fflush_and_ftruncate(out_fp,
 343                                     file_offset + wim_resource_size(lte));
 344 }
 345
 346 /*
 347  * Writes a WIM resource to a FILE * opened for writing.  The resource may be
 348  * written uncompressed or compressed depending on the @out_ctype parameter.
 349  *
 350  * If by chance the resource compresses to more than the original size (this may
 351  * happen with random data or files than are pre-compressed), the resource is
 352  * instead written uncompressed (and this is reflected in the @out_res_entry by
 353  * removing the WIM_RESHDR_FLAG_COMPRESSED flag).
 354  *
 355  * @lte:        The lookup table entry for the WIM resource.
 356  * @out_fp:     The FILE * to write the resource to.
 357  * @out_ctype:  The compression type of the resource to write.  Note: if this is
 358  *                      the same as the compression type of the WIM resource we
 359  *                      need to read, we simply copy the data (i.e. we do not
 360  *                      uncompress it, then compress it again).
 361  * @out_res_entry:  If non-NULL, a resource entry that is filled in with the
 362  *                  offset, original size, compressed size, and compression flag
 363  *                  of the output resource.
 364  *
 365  * Returns 0 on success; nonzero on failure.
 366  */
 367 int write_wim_resource(struct wim_lookup_table_entry *lte,
 368                        FILE *out_fp, int out_ctype,
 369                        struct resource_entry *out_res_entry,
 370                        int flags)
 371 {
 372         u64 bytes_remaining;
 373         u64 original_size;
 374         u64 old_compressed_size;
 375         u64 new_compressed_size;
 376         u64 offset;
 377         int ret;
 378         struct chunk_table *chunk_tab = NULL;
 379         bool raw;
 380         off_t file_offset;
 381         compress_func_t compress = NULL;
 382 #ifdef WITH_NTFS_3G
 383         ntfs_inode *ni = NULL;
 384 #endif
 385
 386         wimlib_assert(lte);
 387
 388         /* Original size of the resource */
 389         original_size = wim_resource_size(lte);
 390
 391         /* Compressed size of the resource (as it exists now) */
 392         old_compressed_size = wim_resource_compressed_size(lte);
 393
 394         /* Current offset in output file */
 395         file_offset = ftello(out_fp);
 396         if (file_offset == -1) {
 397                 ERROR_WITH_ERRNO("Failed to get offset in output "
 398                                  "stream");
 399                 return WIMLIB_ERR_WRITE;
 400         }
 401
 402         /* Are the compression types the same?  If so, do a raw copy (copy
 403          * without decompressing and recompressing the data). */
 404         raw = (wim_resource_compression_type(lte) == out_ctype
 405                && out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
 406                && !(flags & WIMLIB_RESOURCE_FLAG_RECOMPRESS));
 407
 408         if (raw) {
 409                 flags |= WIMLIB_RESOURCE_FLAG_RAW;
 410                 bytes_remaining = old_compressed_size;
 411         } else {
 412                 flags &= ~WIMLIB_RESOURCE_FLAG_RAW;
 413                 bytes_remaining = original_size;
 414         }
 415
 416         /* Empty resource; nothing needs to be done, so just return success. */
 417         if (bytes_remaining == 0)
 418                 return 0;
 419
 420         /* Buffer for reading chunks for the resource */
 421         u8 buf[min(WIM_CHUNK_SIZE, bytes_remaining)];
 422
 423         /* If we are writing a compressed resource and not doing a raw copy, we
 424          * need to initialize the chunk table */
 425         if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE && !raw) {
 426                 ret = begin_wim_resource_chunk_tab(lte, out_fp, file_offset,
 427                                                    &chunk_tab);
 428                 if (ret != 0)
 429                         goto out;
 430         }
 431
 432         /* If the WIM resource is in an external file, open a FILE * to it so we
 433          * don't have to open a temporary one in read_wim_resource() for each
 434          * chunk. */
 435 #ifdef WITH_NTFS_3G
 436         ret = prepare_resource_for_read(lte, &ni);
 437 #else
 438         ret = prepare_resource_for_read(lte);
 439 #endif
 440         if (ret != 0)
 441                 goto out;
 442
 443         /* If we aren't doing a raw copy, we will compute the SHA1 message
 444          * digest of the resource as we read it, and verify it's the same as the
 445          * hash given in the lookup table entry once we've finished reading the
 446          * resource. */
 447         SHA_CTX ctx;
 448         if (!raw) {
 449                 sha1_init(&ctx);
 450                 compress = get_compress_func(out_ctype);
 451         }
 452         offset = 0;
 453
 454         /* While there are still bytes remaining in the WIM resource, read a
 455          * chunk of the resource, update SHA1, then write that chunk using the
 456          * desired compression type. */
 457         do {
 458                 u64 to_read = min(bytes_remaining, WIM_CHUNK_SIZE);
 459                 ret = read_wim_resource(lte, buf, to_read, offset, flags);
 460                 if (ret != 0)
 461                         goto out_fclose;
 462                 if (!raw)
 463                         sha1_update(&ctx, buf, to_read);
 464                 ret = write_wim_resource_chunk(buf, to_read, out_fp,
 465                                                compress, chunk_tab);
 466                 if (ret != 0)
 467                         goto out_fclose;
 468                 bytes_remaining -= to_read;
 469                 offset += to_read;
 470         } while (bytes_remaining);
 471
 472         /* Raw copy:  The new compressed size is the same as the old compressed
 473          * size
 474          *
 475          * Using WIMLIB_COMPRESSION_TYPE_NONE:  The new compressed size is the
 476          * original size
 477          *
 478          * Using a different compression type:  Call
 479          * finish_wim_resource_chunk_tab() and it will provide the new
 480          * compressed size.
 481          */
 482         if (raw) {
 483                 new_compressed_size = old_compressed_size;
 484         } else {
 485                 if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE)
 486                         new_compressed_size = original_size;
 487                 else {
 488                         ret = finish_wim_resource_chunk_tab(chunk_tab, out_fp,
 489                                                             &new_compressed_size);
 490                         if (ret != 0)
 491                                 goto out_fclose;
 492                 }
 493         }
 494
 495         /* Verify SHA1 message digest of the resource, unless we are doing a raw
 496          * write (in which case we never even saw the uncompressed data).  Or,
 497          * if the hash we had before is all 0's, just re-set it to be the new
 498          * hash. */
 499         if (!raw) {
 500                 u8 md[SHA1_HASH_SIZE];
 501                 sha1_final(md, &ctx);
 502                 if (is_zero_hash(lte->hash)) {
 503                         copy_hash(lte->hash, md);
 504                 } else if (!hashes_equal(md, lte->hash)) {
 505                         ERROR("WIM resource has incorrect hash!");
 506                         if (lte->resource_location == RESOURCE_IN_FILE_ON_DISK) {
 507                                 ERROR("We were reading it from `%s'; maybe it changed "
 508                                       "while we were reading it.",
 509                                       lte->file_on_disk);
 510                         }
 511                         ret = WIMLIB_ERR_INVALID_RESOURCE_HASH;
 512                         goto out_fclose;
 513                 }
 514         }
 515
 516         if (!raw && new_compressed_size >= original_size &&
 517             out_ctype != WIMLIB_COMPRESSION_TYPE_NONE)
 518         {
 519                 /* Oops!  We compressed the resource to larger than the original
 520                  * size.  Write the resource uncompressed instead. */
 521                 ret = write_uncompressed_resource_and_truncate(lte,
 522                                                                out_fp,
 523                                                                file_offset,
 524                                                                out_res_entry);
 525                 if (ret != 0)
 526                         goto out_fclose;
 527         } else {
 528                 if (out_res_entry) {
 529                         out_res_entry->size          = new_compressed_size;
 530                         out_res_entry->original_size = original_size;
 531                         out_res_entry->offset        = file_offset;
 532                         out_res_entry->flags         = lte->resource_entry.flags
 533                                                         & ~WIM_RESHDR_FLAG_COMPRESSED;
 534                         if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE)
 535                                 out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
 536                 }
 537         }
 538         ret = 0;
 539 out_fclose:
 540 #ifdef WITH_NTFS_3G
 541         end_wim_resource_read(lte, ni);
 542 #else
 543         end_wim_resource_read(lte);
 544 #endif
 545 out:
 546         FREE(chunk_tab);
 547         return ret;
 548 }
 549
 550 #ifdef ENABLE_MULTITHREADED_COMPRESSION
 551
 552 /* Blocking shared queue (solves the producer-consumer problem) */
 553 struct shared_queue {
 554         unsigned size;
 555         unsigned front;
 556         unsigned back;
 557         unsigned filled_slots;
 558         void **array;
 559         pthread_mutex_t lock;
 560         pthread_cond_t msg_avail_cond;
 561         pthread_cond_t space_avail_cond;
 562 };
 563
 564 static int shared_queue_init(struct shared_queue *q, unsigned size)
 565 {
 566         wimlib_assert(size != 0);
 567         q->array = CALLOC(sizeof(q->array[0]), size);
 568         if (!q->array)
 569                 return WIMLIB_ERR_NOMEM;
 570         q->filled_slots = 0;
 571         q->front = 0;
 572         q->back = size - 1;
 573         q->size = size;
 574         pthread_mutex_init(&q->lock, NULL);
 575         pthread_cond_init(&q->msg_avail_cond, NULL);
 576         pthread_cond_init(&q->space_avail_cond, NULL);
 577         return 0;
 578 }
 579
 580 static void shared_queue_destroy(struct shared_queue *q)
 581 {
 582         FREE(q->array);
 583         pthread_mutex_destroy(&q->lock);
 584         pthread_cond_destroy(&q->msg_avail_cond);
 585         pthread_cond_destroy(&q->space_avail_cond);
 586 }
 587
 588 static void shared_queue_put(struct shared_queue *q, void *obj)
 589 {
 590         pthread_mutex_lock(&q->lock);
 591         while (q->filled_slots == q->size)
 592                 pthread_cond_wait(&q->space_avail_cond, &q->lock);
 593
 594         q->back = (q->back + 1) % q->size;
 595         q->array[q->back] = obj;
 596         q->filled_slots++;
 597
 598         pthread_cond_broadcast(&q->msg_avail_cond);
 599         pthread_mutex_unlock(&q->lock);
 600 }
 601
 602 static void *shared_queue_get(struct shared_queue *q)
 603 {
 604         void *obj;
 605
 606         pthread_mutex_lock(&q->lock);
 607         while (q->filled_slots == 0)
 608                 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
 609
 610         obj = q->array[q->front];
 611         q->array[q->front] = NULL;
 612         q->front = (q->front + 1) % q->size;
 613         q->filled_slots--;
 614
 615         pthread_cond_broadcast(&q->space_avail_cond);
 616         pthread_mutex_unlock(&q->lock);
 617         return obj;
 618 }
 619
 620 struct compressor_thread_params {
 621         struct shared_queue *res_to_compress_queue;
 622         struct shared_queue *compressed_res_queue;
 623         compress_func_t compress;
 624 };
 625
 626 #define MAX_CHUNKS_PER_MSG 2
 627
 628 struct message {
 629         struct wim_lookup_table_entry *lte;
 630         u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
 631         u8 *out_compressed_chunks[MAX_CHUNKS_PER_MSG];
 632         u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
 633         unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
 634         unsigned compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
 635         unsigned num_chunks;
 636         struct list_head list;
 637         bool complete;
 638         u64 begin_chunk;
 639 };
 640
 641 static void compress_chunks(struct message *msg, compress_func_t compress)
 642 {
 643         for (unsigned i = 0; i < msg->num_chunks; i++) {
 644                 DEBUG2("compress chunk %u of %u", i, msg->num_chunks);
 645                 int ret = compress(msg->uncompressed_chunks[i],
 646                                    msg->uncompressed_chunk_sizes[i],
 647                                    msg->compressed_chunks[i],
 648                                    &msg->compressed_chunk_sizes[i]);
 649                 if (ret == 0) {
 650                         msg->out_compressed_chunks[i] = msg->compressed_chunks[i];
 651                 } else {
 652                         msg->out_compressed_chunks[i] = msg->uncompressed_chunks[i];
 653                         msg->compressed_chunk_sizes[i] = msg->uncompressed_chunk_sizes[i];
 654                 }
 655         }
 656 }
 657
 658 /* Compressor thread routine.  This is a lot simpler than the main thread
 659  * routine: just repeatedly get a group of chunks from the
 660  * res_to_compress_queue, compress them, and put them in the
 661  * compressed_res_queue.  A NULL pointer indicates that the thread should stop.
 662  * */
 663 static void *compressor_thread_proc(void *arg)
 664 {
 665         struct compressor_thread_params *params = arg;
 666         struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
 667         struct shared_queue *compressed_res_queue = params->compressed_res_queue;
 668         compress_func_t compress = params->compress;
 669         struct message *msg;
 670
 671         DEBUG("Compressor thread ready");
 672         while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
 673                 compress_chunks(msg, compress);
 674                 shared_queue_put(compressed_res_queue, msg);
 675         }
 676         DEBUG("Compressor thread terminating");
 677         return NULL;
 678 }
 679 #endif
 680
 681 static int do_write_stream_list(struct list_head *my_resources,
 682                                 FILE *out_fp,
 683                                 int out_ctype,
 684                                 wimlib_progress_func_t progress_func,
 685                                 union wimlib_progress_info *progress,
 686                                 int write_resource_flags)
 687 {
 688         int ret;
 689         struct wim_lookup_table_entry *lte, *tmp;
 690
 691         list_for_each_entry_safe(lte, tmp, my_resources, staging_list) {
 692                 ret = write_wim_resource(lte,
 693                                          out_fp,
 694                                          out_ctype,
 695                                          &lte->output_resource_entry,
 696                                          write_resource_flags);
 697                 if (ret != 0)
 698                         return ret;
 699                 list_del(&lte->staging_list);
 700                 progress->write_streams.completed_bytes +=
 701                         wim_resource_size(lte);
 702                 progress->write_streams.completed_streams++;
 703                 if (progress_func) {
 704                         progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
 705                                       progress);
 706                 }
 707         }
 708         return 0;
 709 }
 710
 711 static int write_stream_list_serial(struct list_head *stream_list,
 712                                     FILE *out_fp,
 713                                     int out_ctype,
 714                                     int write_flags,
 715                                     wimlib_progress_func_t progress_func,
 716                                     union wimlib_progress_info *progress)
 717 {
 718         int write_resource_flags;
 719
 720         if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
 721                 write_resource_flags = WIMLIB_RESOURCE_FLAG_RECOMPRESS;
 722         else
 723                 write_resource_flags = 0;
 724         progress->write_streams.num_threads = 1;
 725         if (progress_func)
 726                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
 727         return do_write_stream_list(stream_list, out_fp,
 728                                     out_ctype, progress_func,
 729                                     progress, write_resource_flags);
 730 }
 731
 732 #ifdef ENABLE_MULTITHREADED_COMPRESSION
 733 static int write_wim_chunks(struct message *msg, FILE *out_fp,
 734                             struct chunk_table *chunk_tab)
 735 {
 736         for (unsigned i = 0; i < msg->num_chunks; i++) {
 737                 unsigned chunk_csize = msg->compressed_chunk_sizes[i];
 738
 739                 DEBUG2("Write wim chunk %u of %u (csize = %u)",
 740                       i, msg->num_chunks, chunk_csize);
 741
 742                 if (fwrite(msg->out_compressed_chunks[i], 1, chunk_csize, out_fp)
 743                     != chunk_csize)
 744                 {
 745                         ERROR_WITH_ERRNO("Failed to write WIM chunk");
 746                         return WIMLIB_ERR_WRITE;
 747                 }
 748
 749                 *chunk_tab->cur_offset_p++ = chunk_tab->cur_offset;
 750                 chunk_tab->cur_offset += chunk_csize;
 751         }
 752         return 0;
 753 }
 754
 755 /*
 756  * This function is executed by the main thread when the resources are being
 757  * compressed in parallel.  The main thread is in change of all reading of the
 758  * uncompressed data and writing of the compressed data.  The compressor threads
 759  * *only* do compression from/to in-memory buffers.
 760  *
 761  * Each unit of work given to a compressor thread is up to MAX_CHUNKS_PER_MSG
 762  * chunks of compressed data to compress, represented in a `struct message'.
 763  * Each message is passed from the main thread to a worker thread through the
 764  * res_to_compress_queue, and it is passed back through the
 765  * compressed_res_queue.
 766  */
 767 static int main_writer_thread_proc(struct list_head *stream_list,
 768                                    FILE *out_fp,
 769                                    int out_ctype,
 770                                    struct shared_queue *res_to_compress_queue,
 771                                    struct shared_queue *compressed_res_queue,
 772                                    size_t num_messages,
 773                                    int write_flags,
 774                                    wimlib_progress_func_t progress_func,
 775                                    union wimlib_progress_info *progress)
 776 {
 777         int ret;
 778         struct chunk_table *cur_chunk_tab = NULL;
 779         struct message *msgs = CALLOC(num_messages, sizeof(struct message));
 780         struct wim_lookup_table_entry *next_lte = NULL;
 781
 782         // Initially, all the messages are available to use.
 783         LIST_HEAD(available_msgs);
 784
 785         if (!msgs) {
 786                 ret = WIMLIB_ERR_NOMEM;
 787                 goto out;
 788         }
 789
 790         for (size_t i = 0; i < num_messages; i++)
 791                 list_add(&msgs[i].list, &available_msgs);
 792
 793         // outstanding_resources is the list of resources that currently have
 794         // had chunks sent off for compression.
 795         //
 796         // The first stream in outstanding_resources is the stream that is
 797         // currently being written (cur_lte).
 798         //
 799         // The last stream in outstanding_resources is the stream that is
 800         // currently being read and chunks fed to the compressor threads
 801         // (next_lte).
 802         //
 803         // Depending on the number of threads and the sizes of the resource,
 804         // the outstanding streams list may contain streams between cur_lte and
 805         // next_lte that have all their chunks compressed or being compressed,
 806         // but haven't been written yet.
 807         //
 808         LIST_HEAD(outstanding_resources);
 809         struct list_head *next_resource = stream_list->next;
 810         u64 next_chunk = 0;
 811         u64 next_num_chunks = 0;
 812
 813         // As in write_wim_resource(), each resource we read is checksummed.
 814         SHA_CTX next_sha_ctx;
 815         u8 next_hash[SHA1_HASH_SIZE];
 816
 817         // Resources that don't need any chunks compressed are added to this
 818         // list and written directly by the main thread.
 819         LIST_HEAD(my_resources);
 820
 821         struct wim_lookup_table_entry *cur_lte = NULL;
 822         struct message *msg;
 823
 824 #ifdef WITH_NTFS_3G
 825         ntfs_inode *ni = NULL;
 826 #endif
 827
 828         DEBUG("Initializing buffers for uncompressed "
 829               "and compressed data (%zu bytes needed)",
 830               num_messages * MAX_CHUNKS_PER_MSG * WIM_CHUNK_SIZE * 2);
 831
 832         // Pre-allocate all the buffers that will be needed to do the chunk
 833         // compression.
 834         for (size_t i = 0; i < num_messages; i++) {
 835                 for (size_t j = 0; j < MAX_CHUNKS_PER_MSG; j++) {
 836                         msgs[i].compressed_chunks[j] = MALLOC(WIM_CHUNK_SIZE);
 837
 838                         // The extra 8 bytes is because longest_match() in lz.c
 839                         // may read a little bit off the end of the uncompressed
 840                         // data.  It doesn't need to be initialized--- we really
 841                         // just need to avoid accessing an unmapped page.
 842                         msgs[i].uncompressed_chunks[j] = MALLOC(WIM_CHUNK_SIZE + 8);
 843                         if (msgs[i].compressed_chunks[j] == NULL ||
 844                             msgs[i].uncompressed_chunks[j] == NULL)
 845                         {
 846                                 ret = WIMLIB_ERR_NOMEM;
 847                                 goto out;
 848                         }
 849                 }
 850         }
 851
 852         // This loop is executed until all resources have been written, except
 853         // possibly a few that have been added to the @my_resources list for
 854         // writing later.
 855         while (1) {
 856                 // Send chunks to the compressor threads until either (a) there
 857                 // are no more messages available since they were all sent off,
 858                 // or (b) there are no more resources that need to be
 859                 // compressed.
 860                 while (!list_empty(&available_msgs)) {
 861                         if (next_chunk == next_num_chunks) {
 862                                 // If next_chunk == next_num_chunks, there are
 863                                 // no more chunks to write in the current
 864                                 // stream.  So, check the SHA1 message digest of
 865                                 // the stream that was just finished (unless
 866                                 // next_lte == NULL, which is the case the very
 867                                 // first time this loop is entered, and also
 868                                 // near the very end of the compression when
 869                                 // there are no more streams.)  Then, advance to
 870                                 // the next stream (if there is one).
 871                                 if (next_lte != NULL) {
 872                                 #ifdef WITH_NTFS_3G
 873                                         end_wim_resource_read(next_lte, ni);
 874                                         ni = NULL;
 875                                 #else
 876                                         end_wim_resource_read(next_lte);
 877                                 #endif
 878                                         DEBUG2("Finalize SHA1 md (next_num_chunks=%zu)",
 879                                                next_num_chunks);
 880                                         sha1_final(next_hash, &next_sha_ctx);
 881                                         if (!hashes_equal(next_lte->hash, next_hash)) {
 882                                                 ERROR("WIM resource has incorrect hash!");
 883                                                 if (next_lte->resource_location ==
 884                                                     RESOURCE_IN_FILE_ON_DISK)
 885                                                 {
 886                                                         ERROR("We were reading it from `%s'; "
 887                                                               "maybe it changed while we were "
 888                                                               "reading it.",
 889                                                               next_lte->file_on_disk);
 890                                                 }
 891                                                 ret = WIMLIB_ERR_INVALID_RESOURCE_HASH;
 892                                                 goto out;
 893                                         }
 894                                 }
 895
 896                                 // Advance to the next resource.
 897                                 //
 898                                 // If the next resource needs no compression, just write
 899                                 // it with this thread (not now though--- we could be in
 900                                 // the middle of writing another resource.)  Keep doing
 901                                 // this until we either get to the end of the resources
 902                                 // list, or we get to a resource that needs compression.
 903                                 while (1) {
 904                                         if (next_resource == stream_list) {
 905                                                 // No more resources to send for
 906                                                 // compression
 907                                                 next_lte = NULL;
 908                                                 break;
 909                                         }
 910                                         next_lte = container_of(next_resource,
 911                                                                 struct wim_lookup_table_entry,
 912                                                                 staging_list);
 913                                         next_resource = next_resource->next;
 914                                         if ((!(write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
 915                                                && wim_resource_compression_type(next_lte) == out_ctype)
 916                                             || wim_resource_size(next_lte) == 0)
 917                                         {
 918                                                 list_add_tail(&next_lte->staging_list,
 919                                                               &my_resources);
 920                                         } else {
 921                                                 list_add_tail(&next_lte->staging_list,
 922                                                               &outstanding_resources);
 923                                                 next_chunk = 0;
 924                                                 next_num_chunks = wim_resource_chunks(next_lte);
 925                                                 sha1_init(&next_sha_ctx);
 926                                                 INIT_LIST_HEAD(&next_lte->msg_list);
 927                                         #ifdef WITH_NTFS_3G
 928                                                 ret = prepare_resource_for_read(next_lte, &ni);
 929                                         #else
 930                                                 ret = prepare_resource_for_read(next_lte);
 931                                         #endif
 932
 933                                                 if (ret != 0)
 934                                                         goto out;
 935                                                 if (cur_lte == NULL) {
 936                                                         // Set cur_lte for the
 937                                                         // first time
 938                                                         cur_lte = next_lte;
 939                                                 }
 940                                                 break;
 941                                         }
 942                                 }
 943                         }
 944
 945                         if (next_lte == NULL) {
 946                                 // No more resources to send for compression
 947                                 break;
 948                         }
 949
 950                         // Get a message from the available messages
 951                         // list
 952                         msg = container_of(available_msgs.next,
 953                                            struct message,
 954                                            list);
 955
 956                         // ... and delete it from the available messages
 957                         // list
 958                         list_del(&msg->list);
 959
 960                         // Initialize the message with the chunks to
 961                         // compress.
 962                         msg->num_chunks = min(next_num_chunks - next_chunk,
 963                                               MAX_CHUNKS_PER_MSG);
 964                         msg->lte = next_lte;
 965                         msg->complete = false;
 966                         msg->begin_chunk = next_chunk;
 967
 968                         unsigned size = WIM_CHUNK_SIZE;
 969                         for (unsigned i = 0; i < msg->num_chunks; i++) {
 970
 971                                 // Read chunk @next_chunk of the stream into the
 972                                 // message so that a compressor thread can
 973                                 // compress it.
 974
 975                                 if (next_chunk == next_num_chunks - 1) {
 976                                         size = MODULO_NONZERO(wim_resource_size(next_lte),
 977                                                               WIM_CHUNK_SIZE);
 978                                 }
 979
 980                                 DEBUG2("Read resource (size=%u, offset=%zu)",
 981                                       size, next_chunk * WIM_CHUNK_SIZE);
 982
 983                                 msg->uncompressed_chunk_sizes[i] = size;
 984
 985                                 ret = read_wim_resource(next_lte,
 986                                                         msg->uncompressed_chunks[i],
 987                                                         size,
 988                                                         next_chunk * WIM_CHUNK_SIZE,
 989                                                         0);
 990                                 if (ret != 0)
 991                                         goto out;
 992                                 sha1_update(&next_sha_ctx,
 993                                             msg->uncompressed_chunks[i], size);
 994                                 next_chunk++;
 995                         }
 996
 997                         // Send the compression request
 998                         list_add_tail(&msg->list, &next_lte->msg_list);
 999                         shared_queue_put(res_to_compress_queue, msg);
1000                         DEBUG2("Compression request sent");
1001                 }
1002
1003                 // If there are no outstanding resources, there are no more
1004                 // resources that need to be written.
1005                 if (list_empty(&outstanding_resources)) {
1006                         ret = 0;
1007                         goto out;
1008                 }
1009
1010                 // Get the next message from the queue and process it.
1011                 // The message will contain 1 or more data chunks that have been
1012                 // compressed.
1013                 msg = shared_queue_get(compressed_res_queue);
1014                 msg->complete = true;
1015
1016                 // Is this the next chunk in the current resource?  If it's not
1017                 // (i.e., an earlier chunk in a same or different resource
1018                 // hasn't been compressed yet), do nothing, and keep this
1019                 // message around until all earlier chunks are received.
1020                 //
1021                 // Otherwise, write all the chunks we can.
1022                 while (cur_lte != NULL &&
1023                        !list_empty(&cur_lte->msg_list) &&
1024                        (msg = container_of(cur_lte->msg_list.next,
1025                                            struct message,
1026                                            list))->complete)
1027                 {
1028                         DEBUG2("Complete msg (begin_chunk=%"PRIu64")", msg->begin_chunk);
1029                         if (msg->begin_chunk == 0) {
1030                                 DEBUG2("Begin chunk tab");
1031
1032                                 // This is the first set of chunks.  Leave space
1033                                 // for the chunk table in the output file.
1034                                 off_t cur_offset = ftello(out_fp);
1035                                 if (cur_offset == -1) {
1036                                         ret = WIMLIB_ERR_WRITE;
1037                                         goto out;
1038                                 }
1039                                 ret = begin_wim_resource_chunk_tab(cur_lte,
1040                                                                    out_fp,
1041                                                                    cur_offset,
1042                                                                    &cur_chunk_tab);
1043                                 if (ret != 0)
1044                                         goto out;
1045                         }
1046
1047                         // Write the compressed chunks from the message.
1048                         ret = write_wim_chunks(msg, out_fp, cur_chunk_tab);
1049                         if (ret != 0)
1050                                 goto out;
1051
1052                         list_del(&msg->list);
1053
1054                         // This message is available to use for different chunks
1055                         // now.
1056                         list_add(&msg->list, &available_msgs);
1057
1058                         // Was this the last chunk of the stream?  If so, finish
1059                         // it.
1060                         if (list_empty(&cur_lte->msg_list) &&
1061                             msg->begin_chunk + msg->num_chunks == cur_chunk_tab->num_chunks)
1062                         {
1063                                 DEBUG2("Finish wim chunk tab");
1064                                 u64 res_csize;
1065                                 ret = finish_wim_resource_chunk_tab(cur_chunk_tab,
1066                                                                     out_fp,
1067                                                                     &res_csize);
1068                                 if (ret != 0)
1069                                         goto out;
1070
1071                                 if (res_csize >= wim_resource_size(cur_lte)) {
1072                                         /* Oops!  We compressed the resource to
1073                                          * larger than the original size.  Write
1074                                          * the resource uncompressed instead. */
1075                                         ret = write_uncompressed_resource_and_truncate(
1076                                                          cur_lte,
1077                                                          out_fp,
1078                                                          cur_chunk_tab->file_offset,
1079                                                          &cur_lte->output_resource_entry);
1080                                         if (ret != 0)
1081                                                 goto out;
1082                                 } else {
1083                                         cur_lte->output_resource_entry.size =
1084                                                 res_csize;
1085
1086                                         cur_lte->output_resource_entry.original_size =
1087                                                 cur_lte->resource_entry.original_size;
1088
1089                                         cur_lte->output_resource_entry.offset =
1090                                                 cur_chunk_tab->file_offset;
1091
1092                                         cur_lte->output_resource_entry.flags =
1093                                                 cur_lte->resource_entry.flags |
1094                                                         WIM_RESHDR_FLAG_COMPRESSED;
1095                                 }
1096
1097                                 progress->write_streams.completed_bytes +=
1098                                                 wim_resource_size(cur_lte);
1099                                 progress->write_streams.completed_streams++;
1100
1101                                 if (progress_func) {
1102                                         progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
1103                                                       progress);
1104                                 }
1105
1106                                 FREE(cur_chunk_tab);
1107                                 cur_chunk_tab = NULL;
1108
1109                                 struct list_head *next = cur_lte->staging_list.next;
1110                                 list_del(&cur_lte->staging_list);
1111
1112                                 if (next == &outstanding_resources)
1113                                         cur_lte = NULL;
1114                                 else
1115                                         cur_lte = container_of(cur_lte->staging_list.next,
1116                                                                struct wim_lookup_table_entry,
1117                                                                staging_list);
1118
1119                                 // Since we just finished writing a stream,
1120                                 // write any streams that have been added to the
1121                                 // my_resources list for direct writing by the
1122                                 // main thread (e.g. resources that don't need
1123                                 // to be compressed because the desired
1124                                 // compression type is the same as the previous
1125                                 // compression type).
1126                                 ret = do_write_stream_list(&my_resources,
1127                                                            out_fp,
1128                                                            out_ctype,
1129                                                            progress_func,
1130                                                            progress,
1131                                                            0);
1132                                 if (ret != 0)
1133                                         goto out;
1134                         }
1135                 }
1136         }
1137
1138 out:
1139         if (ret == WIMLIB_ERR_NOMEM) {
1140                 ERROR("Could not allocate enough memory for "
1141                       "multi-threaded compression");
1142         }
1143
1144         if (next_lte) {
1145 #ifdef WITH_NTFS_3G
1146                 end_wim_resource_read(next_lte, ni);
1147 #else
1148                 end_wim_resource_read(next_lte);
1149 #endif
1150         }
1151
1152         if (ret == 0) {
1153                 ret = do_write_stream_list(&my_resources, out_fp,
1154                                            out_ctype, progress_func,
1155                                            progress, 0);
1156         } else {
1157                 if (msgs) {
1158                         size_t num_available_msgs = 0;
1159                         struct list_head *cur;
1160
1161                         list_for_each(cur, &available_msgs) {
1162                                 num_available_msgs++;
1163                         }
1164
1165                         while (num_available_msgs < num_messages) {
1166                                 shared_queue_get(compressed_res_queue);
1167                                 num_available_msgs++;
1168                         }
1169                 }
1170         }
1171
1172         if (msgs) {
1173                 for (size_t i = 0; i < num_messages; i++) {
1174                         for (size_t j = 0; j < MAX_CHUNKS_PER_MSG; j++) {
1175                                 FREE(msgs[i].compressed_chunks[j]);
1176                                 FREE(msgs[i].uncompressed_chunks[j]);
1177                         }
1178                 }
1179                 FREE(msgs);
1180         }
1181
1182         FREE(cur_chunk_tab);
1183         return ret;
1184 }
1185
1186
1187 static int write_stream_list_parallel(struct list_head *stream_list,
1188                                       FILE *out_fp,
1189                                       int out_ctype,
1190                                       int write_flags,
1191                                       unsigned num_threads,
1192                                       wimlib_progress_func_t progress_func,
1193                                       union wimlib_progress_info *progress)
1194 {
1195         int ret;
1196         struct shared_queue res_to_compress_queue;
1197         struct shared_queue compressed_res_queue;
1198         pthread_t *compressor_threads = NULL;
1199
1200         if (num_threads == 0) {
1201                 long nthreads = sysconf(_SC_NPROCESSORS_ONLN);
1202                 if (nthreads < 1) {
1203                         WARNING("Could not determine number of processors! Assuming 1");
1204                         goto out_serial;
1205                 } else {
1206                         num_threads = nthreads;
1207                 }
1208         }
1209
1210         progress->write_streams.num_threads = num_threads;
1211         wimlib_assert(stream_list->next != stream_list);
1212
1213         static const double MESSAGES_PER_THREAD = 2.0;
1214         size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1215
1216         DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1217
1218         ret = shared_queue_init(&res_to_compress_queue, queue_size);
1219         if (ret != 0)
1220                 goto out_serial;
1221
1222         ret = shared_queue_init(&compressed_res_queue, queue_size);
1223         if (ret != 0)
1224                 goto out_destroy_res_to_compress_queue;
1225
1226         struct compressor_thread_params params;
1227         params.res_to_compress_queue = &res_to_compress_queue;
1228         params.compressed_res_queue = &compressed_res_queue;
1229         params.compress = get_compress_func(out_ctype);
1230
1231         compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1232         if (!compressor_threads) {
1233                 ret = WIMLIB_ERR_NOMEM;
1234                 goto out_destroy_compressed_res_queue;
1235         }
1236
1237         for (unsigned i = 0; i < num_threads; i++) {
1238                 DEBUG("pthread_create thread %u", i);
1239                 ret = pthread_create(&compressor_threads[i], NULL,
1240                                      compressor_thread_proc, &params);
1241                 if (ret != 0) {
1242                         ret = -1;
1243                         ERROR_WITH_ERRNO("Failed to create compressor "
1244                                          "thread %u", i);
1245                         num_threads = i;
1246                         goto out_join;
1247                 }
1248         }
1249
1250         if (progress_func)
1251                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS, progress);
1252
1253         ret = main_writer_thread_proc(stream_list,
1254                                       out_fp,
1255                                       out_ctype,
1256                                       &res_to_compress_queue,
1257                                       &compressed_res_queue,
1258                                       queue_size,
1259                                       write_flags,
1260                                       progress_func,
1261                                       progress);
1262 out_join:
1263         for (unsigned i = 0; i < num_threads; i++)
1264                 shared_queue_put(&res_to_compress_queue, NULL);
1265
1266         for (unsigned i = 0; i < num_threads; i++) {
1267                 if (pthread_join(compressor_threads[i], NULL)) {
1268                         WARNING_WITH_ERRNO("Failed to join compressor "
1269                                            "thread %u", i);
1270                 }
1271         }
1272         FREE(compressor_threads);
1273 out_destroy_compressed_res_queue:
1274         shared_queue_destroy(&compressed_res_queue);
1275 out_destroy_res_to_compress_queue:
1276         shared_queue_destroy(&res_to_compress_queue);
1277         if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
1278                 return ret;
1279 out_serial:
1280         WARNING("Falling back to single-threaded compression");
1281         return write_stream_list_serial(stream_list,
1282                                         out_fp,
1283                                         out_ctype,
1284                                         write_flags,
1285                                         progress_func,
1286                                         progress);
1287
1288 }
1289 #endif
1290
1291 /*
1292  * Write a list of streams to a WIM (@out_fp) using the compression type
1293  * @out_ctype and up to @num_threads compressor threads.
1294  */
1295 static int write_stream_list(struct list_head *stream_list, FILE *out_fp,
1296                              int out_ctype, int write_flags,
1297                              unsigned num_threads,
1298                              wimlib_progress_func_t progress_func)
1299 {
1300         struct wim_lookup_table_entry *lte;
1301         size_t num_streams = 0;
1302         u64 total_bytes = 0;
1303         u64 total_compression_bytes = 0;
1304         union wimlib_progress_info progress;
1305
1306         list_for_each_entry(lte, stream_list, staging_list) {
1307                 num_streams++;
1308                 total_bytes += wim_resource_size(lte);
1309                 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
1310                        && (wim_resource_compression_type(lte) != out_ctype ||
1311                            (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)))
1312                 {
1313                         total_compression_bytes += wim_resource_size(lte);
1314                 }
1315         }
1316         progress.write_streams.total_bytes       = total_bytes;
1317         progress.write_streams.total_streams     = num_streams;
1318         progress.write_streams.completed_bytes   = 0;
1319         progress.write_streams.completed_streams = 0;
1320         progress.write_streams.num_threads       = num_threads;
1321         progress.write_streams.compression_type  = out_ctype;
1322
1323 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1324         if (total_compression_bytes >= 1000000 && num_threads != 1)
1325                 return write_stream_list_parallel(stream_list,
1326                                                   out_fp,
1327                                                   out_ctype,
1328                                                   write_flags,
1329                                                   num_threads,
1330                                                   progress_func,
1331                                                   &progress);
1332         else
1333 #endif
1334                 return write_stream_list_serial(stream_list,
1335                                                 out_fp,
1336                                                 out_ctype,
1337                                                 write_flags,
1338                                                 progress_func,
1339                                                 &progress);
1340 }
1341
1342 struct lte_overwrite_prepare_args {
1343         WIMStruct *wim;
1344         off_t end_offset;
1345         struct list_head *stream_list;
1346 };
1347
1348 static int lte_overwrite_prepare(struct wim_lookup_table_entry *lte, void *arg)
1349 {
1350         struct lte_overwrite_prepare_args *args = arg;
1351
1352         if (lte->resource_location == RESOURCE_IN_WIM &&
1353             lte->wim == args->wim &&
1354             lte->resource_entry.offset + lte->resource_entry.size > args->end_offset)
1355         {
1356         #ifdef ENABLE_ERROR_MESSAGES
1357                 ERROR("The following resource is after the XML data:");
1358                 print_lookup_table_entry(lte, stderr);
1359         #endif
1360                 return WIMLIB_ERR_RESOURCE_ORDER;
1361         }
1362
1363         lte->out_refcnt = lte->refcnt;
1364         memcpy(&lte->output_resource_entry, &lte->resource_entry,
1365                sizeof(struct resource_entry));
1366         if (!(lte->resource_entry.flags & WIM_RESHDR_FLAG_METADATA)) {
1367                 wimlib_assert(lte->resource_location != RESOURCE_NONEXISTENT);
1368                 if (lte->resource_location != RESOURCE_IN_WIM || lte->wim != args->wim)
1369                         list_add(&lte->staging_list, args->stream_list);
1370         }
1371         return 0;
1372 }
1373
1374 static int wim_find_new_streams(WIMStruct *wim, off_t end_offset,
1375                                 struct list_head *stream_list)
1376 {
1377         struct lte_overwrite_prepare_args args = {
1378                 .wim         = wim,
1379                 .end_offset  = end_offset,
1380                 .stream_list = stream_list,
1381         };
1382
1383         return for_lookup_table_entry(wim->lookup_table,
1384                                       lte_overwrite_prepare, &args);
1385 }
1386
1387 static int inode_find_streams_to_write(struct wim_inode *inode,
1388                                        struct wim_lookup_table *table,
1389                                        struct list_head *stream_list)
1390 {
1391         struct wim_lookup_table_entry *lte;
1392         for (unsigned i = 0; i <= inode->i_num_ads; i++) {
1393                 lte = inode_stream_lte(inode, i, table);
1394                 if (lte) {
1395                         if (lte->out_refcnt == 0)
1396                                 list_add_tail(&lte->staging_list, stream_list);
1397                         lte->out_refcnt += inode->i_nlink;
1398                 }
1399         }
1400         return 0;
1401 }
1402
1403 static int image_find_streams_to_write(WIMStruct *w)
1404 {
1405         struct wim_inode *inode;
1406         struct hlist_node *cur;
1407         struct hlist_head *inode_list;
1408
1409         inode_list = &wim_get_current_image_metadata(w)->inode_list;
1410         hlist_for_each_entry(inode, cur, inode_list, i_hlist) {
1411                 inode_find_streams_to_write(inode, w->lookup_table,
1412                                             (struct list_head*)w->private);
1413         }
1414         return 0;
1415 }
1416
1417 static int write_wim_streams(WIMStruct *w, int image, int write_flags,
1418                              unsigned num_threads,
1419                              wimlib_progress_func_t progress_func)
1420 {
1421
1422         for_lookup_table_entry(w->lookup_table, lte_zero_out_refcnt, NULL);
1423         LIST_HEAD(stream_list);
1424         w->private = &stream_list;
1425         for_image(w, image, image_find_streams_to_write);
1426         return write_stream_list(&stream_list, w->out_fp,
1427                                  wimlib_get_compression_type(w), write_flags,
1428                                  num_threads, progress_func);
1429 }
1430
1431 /*
1432  * Finish writing a WIM file: write the lookup table, xml data, and integrity
1433  * table (optional), then overwrite the WIM header.
1434  *
1435  * write_flags is a bitwise OR of the following:
1436  *
1437  *      (public)  WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
1438  *              Include an integrity table.
1439  *
1440  *      (public)  WIMLIB_WRITE_FLAG_SHOW_PROGRESS:
1441  *              Show progress information when (if) writing the integrity table.
1442  *
1443  *      (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
1444  *              Don't write the lookup table.
1445  *
1446  *      (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
1447  *              When (if) writing the integrity table, re-use entries from the
1448  *              existing integrity table, if possible.
1449  *
1450  *      (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
1451  *              After writing the XML data but before writing the integrity
1452  *              table, write a temporary WIM header and flush the stream so that
1453  *              the WIM is less likely to become corrupted upon abrupt program
1454  *              termination.
1455  *
1456  *      (private) WIMLIB_WRITE_FLAG_FSYNC:
1457  *              fsync() the output file before closing it.
1458  *
1459  */
1460 int finish_write(WIMStruct *w, int image, int write_flags,
1461                  wimlib_progress_func_t progress_func)
1462 {
1463         int ret;
1464         struct wim_header hdr;
1465         FILE *out = w->out_fp;
1466
1467         /* @hdr will be the header for the new WIM.  First copy all the data
1468          * from the header in the WIMStruct; then set all the fields that may
1469          * have changed, including the resource entries, boot index, and image
1470          * count.  */
1471         memcpy(&hdr, &w->hdr, sizeof(struct wim_header));
1472
1473         if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
1474                 ret = write_lookup_table(w->lookup_table, out, &hdr.lookup_table_res_entry);
1475                 if (ret != 0)
1476                         goto out;
1477         }
1478
1479         ret = write_xml_data(w->wim_info, image, out,
1480                              (write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE) ?
1481                               wim_info_get_total_bytes(w->wim_info) : 0,
1482                              &hdr.xml_res_entry);
1483         if (ret != 0)
1484                 goto out;
1485
1486         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
1487                 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
1488                         struct wim_header checkpoint_hdr;
1489                         memcpy(&checkpoint_hdr, &hdr, sizeof(struct wim_header));
1490                         memset(&checkpoint_hdr.integrity, 0, sizeof(struct resource_entry));
1491                         if (fseeko(out, 0, SEEK_SET) != 0) {
1492                                 ERROR_WITH_ERRNO("Failed to seek to beginning "
1493                                                  "of WIM being written");
1494                                 ret = WIMLIB_ERR_WRITE;
1495                                 goto out;
1496                         }
1497                         ret = write_header(&checkpoint_hdr, out);
1498                         if (ret != 0)
1499                                 goto out;
1500
1501                         if (fflush(out) != 0) {
1502                                 ERROR_WITH_ERRNO("Can't write data to WIM");
1503                                 ret = WIMLIB_ERR_WRITE;
1504                                 goto out;
1505                         }
1506
1507                         if (fseeko(out, 0, SEEK_END) != 0) {
1508                                 ERROR_WITH_ERRNO("Failed to seek to end "
1509                                                  "of WIM being written");
1510                                 ret = WIMLIB_ERR_WRITE;
1511                                 goto out;
1512                         }
1513                 }
1514
1515                 off_t old_lookup_table_end;
1516                 off_t new_lookup_table_end;
1517                 if (write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE) {
1518                         old_lookup_table_end = w->hdr.lookup_table_res_entry.offset +
1519                                                w->hdr.lookup_table_res_entry.size;
1520                 } else {
1521                         old_lookup_table_end = 0;
1522                 }
1523                 new_lookup_table_end = hdr.lookup_table_res_entry.offset +
1524                                        hdr.lookup_table_res_entry.size;
1525
1526                 ret = write_integrity_table(out,
1527                                             &hdr.integrity,
1528                                             new_lookup_table_end,
1529                                             old_lookup_table_end,
1530                                             progress_func);
1531                 if (ret != 0)
1532                         goto out;
1533         } else {
1534                 memset(&hdr.integrity, 0, sizeof(struct resource_entry));
1535         }
1536
1537         /*
1538          * In the WIM header, there is room for the resource entry for a
1539          * metadata resource labeled as the "boot metadata".  This entry should
1540          * be zeroed out if there is no bootable image (boot_idx 0).  Otherwise,
1541          * it should be a copy of the resource entry for the image that is
1542          * marked as bootable.  This is not well documented...
1543          */
1544
1545         /* Set image count and boot index correctly for single image writes */
1546         if (image != WIMLIB_ALL_IMAGES) {
1547                 hdr.image_count = 1;
1548                 if (hdr.boot_idx == image)
1549                         hdr.boot_idx = 1;
1550                 else
1551                         hdr.boot_idx = 0;
1552         }
1553
1554         if (hdr.boot_idx == 0) {
1555                 memset(&hdr.boot_metadata_res_entry, 0,
1556                        sizeof(struct resource_entry));
1557         } else {
1558                 memcpy(&hdr.boot_metadata_res_entry,
1559                        &w->image_metadata[
1560                           hdr.boot_idx - 1].metadata_lte->output_resource_entry,
1561                        sizeof(struct resource_entry));
1562         }
1563
1564         if (fseeko(out, 0, SEEK_SET) != 0) {
1565                 ERROR_WITH_ERRNO("Failed to seek to beginning of WIM "
1566                                  "being written");
1567                 ret = WIMLIB_ERR_WRITE;
1568                 goto out;
1569         }
1570
1571         ret = write_header(&hdr, out);
1572         if (ret != 0)
1573                 goto out;
1574
1575         if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
1576                 if (fflush(out) != 0
1577                     || fsync(fileno(out)) != 0)
1578                 {
1579                         ERROR_WITH_ERRNO("Error flushing data to WIM file");
1580                         ret = WIMLIB_ERR_WRITE;
1581                 }
1582         }
1583 out:
1584         if (fclose(out) != 0) {
1585                 ERROR_WITH_ERRNO("Failed to close the WIM file");
1586                 if (ret == 0)
1587                         ret = WIMLIB_ERR_WRITE;
1588         }
1589         w->out_fp = NULL;
1590         return ret;
1591 }
1592
1593 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
1594 int lock_wim(WIMStruct *w, FILE *fp)
1595 {
1596         int ret = 0;
1597         if (fp && !w->wim_locked) {
1598                 ret = flock(fileno(fp), LOCK_EX | LOCK_NB);
1599                 if (ret != 0) {
1600                         if (errno == EWOULDBLOCK) {
1601                                 ERROR("`%s' is already being modified or has been "
1602                                       "mounted read-write\n"
1603                                       "        by another process!", w->filename);
1604                                 ret = WIMLIB_ERR_ALREADY_LOCKED;
1605                         } else {
1606                                 WARNING_WITH_ERRNO("Failed to lock `%s'",
1607                                                    w->filename);
1608                                 ret = 0;
1609                         }
1610                 } else {
1611                         w->wim_locked = 1;
1612                 }
1613         }
1614         return ret;
1615 }
1616 #endif
1617
1618 static int open_wim_writable(WIMStruct *w, const char *path,
1619                              bool trunc, bool readable)
1620 {
1621         const char *mode;
1622         if (trunc)
1623                 if (readable)
1624                         mode = "w+b";
1625                 else
1626                         mode = "wb";
1627         else
1628                 mode = "r+b";
1629
1630         wimlib_assert(w->out_fp == NULL);
1631         w->out_fp = fopen(path, mode);
1632         if (w->out_fp) {
1633                 return 0;
1634         } else {
1635                 ERROR_WITH_ERRNO("Failed to open `%s' for writing", path);
1636                 return WIMLIB_ERR_OPEN;
1637         }
1638 }
1639
1640
1641 void close_wim_writable(WIMStruct *w)
1642 {
1643         if (w->out_fp) {
1644                 if (fclose(w->out_fp) != 0) {
1645                         WARNING_WITH_ERRNO("Failed to close output WIM");
1646                 }
1647                 w->out_fp = NULL;
1648         }
1649 }
1650
1651 /* Open file stream and write dummy header for WIM. */
1652 int begin_write(WIMStruct *w, const char *path, int write_flags)
1653 {
1654         int ret;
1655         ret = open_wim_writable(w, path, true,
1656                                 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
1657         if (ret != 0)
1658                 return ret;
1659         /* Write dummy header. It will be overwritten later. */
1660         return write_header(&w->hdr, w->out_fp);
1661 }
1662
1663 /* Writes a stand-alone WIM to a file.  */
1664 WIMLIBAPI int wimlib_write(WIMStruct *w, const char *path,
1665                            int image, int write_flags, unsigned num_threads,
1666                            wimlib_progress_func_t progress_func)
1667 {
1668         int ret;
1669
1670         if (!path)
1671                 return WIMLIB_ERR_INVALID_PARAM;
1672
1673         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1674
1675         if (image != WIMLIB_ALL_IMAGES &&
1676              (image < 1 || image > w->hdr.image_count))
1677                 return WIMLIB_ERR_INVALID_IMAGE;
1678
1679         if (w->hdr.total_parts != 1) {
1680                 ERROR("Cannot call wimlib_write() on part of a split WIM");
1681                 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1682         }
1683
1684         ret = begin_write(w, path, write_flags);
1685         if (ret != 0)
1686                 goto out;
1687
1688         ret = write_wim_streams(w, image, write_flags, num_threads,
1689                                 progress_func);
1690         if (ret != 0)
1691                 goto out;
1692
1693         if (progress_func)
1694                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
1695
1696         ret = for_image(w, image, write_metadata_resource);
1697         if (ret != 0)
1698                 goto out;
1699
1700         if (progress_func)
1701                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
1702
1703         ret = finish_write(w, image, write_flags, progress_func);
1704 out:
1705         close_wim_writable(w);
1706         return ret;
1707 }
1708
1709 static bool any_images_modified(WIMStruct *w)
1710 {
1711         for (int i = 0; i < w->hdr.image_count; i++)
1712                 if (w->image_metadata[i].modified)
1713                         return true;
1714         return false;
1715 }
1716
1717 /*
1718  * Overwrite a WIM, possibly appending streams to it.
1719  *
1720  * A WIM looks like (or is supposed to look like) the following:
1721  *
1722  *                   Header (212 bytes)
1723  *                   Streams and metadata resources (variable size)
1724  *                   Lookup table (variable size)
1725  *                   XML data (variable size)
1726  *                   Integrity table (optional) (variable size)
1727  *
1728  * If we are not adding any streams or metadata resources, the lookup table is
1729  * unchanged--- so we only need to overwrite the XML data, integrity table, and
1730  * header.  This operation is potentially unsafe if the program is abruptly
1731  * terminated while the XML data or integrity table are being overwritten, but
1732  * before the new header has been written.  To partially alleviate this problem,
1733  * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
1734  * finish_write() to cause a temporary WIM header to be written after the XML
1735  * data has been written.  This may prevent the WIM from becoming corrupted if
1736  * the program is terminated while the integrity table is being calculated (but
1737  * no guarantees, due to write re-ordering...).
1738  *
1739  * If we are adding new streams or images (metadata resources), the lookup table
1740  * needs to be changed, and those streams need to be written.  In this case, we
1741  * try to perform a safe update of the WIM file by writing the streams *after*
1742  * the end of the previous WIM, then writing the new lookup table, XML data, and
1743  * (optionally) integrity table following the new streams.  This will produce a
1744  * layout like the following:
1745  *
1746  *                   Header (212 bytes)
1747  *                   (OLD) Streams and metadata resources (variable size)
1748  *                   (OLD) Lookup table (variable size)
1749  *                   (OLD) XML data (variable size)
1750  *                   (OLD) Integrity table (optional) (variable size)
1751  *                   (NEW) Streams and metadata resources (variable size)
1752  *                   (NEW) Lookup table (variable size)
1753  *                   (NEW) XML data (variable size)
1754  *                   (NEW) Integrity table (optional) (variable size)
1755  *
1756  * At all points, the WIM is valid as nothing points to the new data yet.  Then,
1757  * the header is overwritten to point to the new lookup table, XML data, and
1758  * integrity table, to produce the following layout:
1759  *
1760  *                   Header (212 bytes)
1761  *                   Streams and metadata resources (variable size)
1762  *                   Nothing (variable size)
1763  *                   More Streams and metadata resources (variable size)
1764  *                   Lookup table (variable size)
1765  *                   XML data (variable size)
1766  *                   Integrity table (optional) (variable size)
1767  *
1768  * This method allows an image to be appended to a large WIM very quickly, and
1769  * is is crash-safe except in the case of write re-ordering, but the
1770  * disadvantage is that a small hole is left in the WIM where the old lookup
1771  * table, xml data, and integrity table were.  (These usually only take up a
1772  * small amount of space compared to the streams, however.)
1773  */
1774 static int overwrite_wim_inplace(WIMStruct *w, int write_flags,
1775                                  unsigned num_threads,
1776                                  wimlib_progress_func_t progress_func)
1777 {
1778         int ret;
1779         struct list_head stream_list;
1780         off_t old_wim_end;
1781         bool found_modified_image;
1782
1783         DEBUG("Overwriting `%s' in-place", w->filename);
1784
1785         /* Make sure that the integrity table (if present) is after the XML
1786          * data, and that there are no stream resources, metadata resources, or
1787          * lookup tables after the XML data.  Otherwise, these data would be
1788          * overwritten. */
1789         if (w->hdr.integrity.offset != 0 &&
1790             w->hdr.integrity.offset < w->hdr.xml_res_entry.offset) {
1791                 ERROR("Didn't expect the integrity table to be before the XML data");
1792                 return WIMLIB_ERR_RESOURCE_ORDER;
1793         }
1794
1795         if (w->hdr.lookup_table_res_entry.offset > w->hdr.xml_res_entry.offset) {
1796                 ERROR("Didn't expect the lookup table to be after the XML data");
1797                 return WIMLIB_ERR_RESOURCE_ORDER;
1798         }
1799
1800
1801         if (w->hdr.integrity.offset)
1802                 old_wim_end = w->hdr.integrity.offset + w->hdr.integrity.size;
1803         else
1804                 old_wim_end = w->hdr.xml_res_entry.offset + w->hdr.xml_res_entry.size;
1805
1806         if (!w->deletion_occurred && !any_images_modified(w)) {
1807                 /* If no images have been modified and no images have been
1808                  * deleted, a new lookup table does not need to be written. */
1809                 old_wim_end = w->hdr.lookup_table_res_entry.offset +
1810                               w->hdr.lookup_table_res_entry.size;
1811                 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
1812                                WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
1813         }
1814         INIT_LIST_HEAD(&stream_list);
1815         ret = wim_find_new_streams(w, old_wim_end, &stream_list);
1816         if (ret != 0)
1817                 return ret;
1818
1819         ret = open_wim_writable(w, w->filename, false,
1820                                 (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) != 0);
1821         if (ret != 0)
1822                 return ret;
1823
1824         ret = lock_wim(w, w->out_fp);
1825         if (ret != 0) {
1826                 fclose(w->out_fp);
1827                 w->out_fp = NULL;
1828                 return ret;
1829         }
1830
1831         if (fseeko(w->out_fp, old_wim_end, SEEK_SET) != 0) {
1832                 ERROR_WITH_ERRNO("Can't seek to end of WIM");
1833                 fclose(w->out_fp);
1834                 w->out_fp = NULL;
1835                 w->wim_locked = 0;
1836                 return WIMLIB_ERR_WRITE;
1837         }
1838
1839         if (!list_empty(&stream_list)) {
1840                 DEBUG("Writing newly added streams (offset = %"PRIu64")",
1841                       old_wim_end);
1842                 ret = write_stream_list(&stream_list, w->out_fp,
1843                                         wimlib_get_compression_type(w),
1844                                         write_flags, num_threads,
1845                                         progress_func);
1846                 if (ret != 0)
1847                         goto out_ftruncate;
1848         } else {
1849                 DEBUG("No new streams were added");
1850         }
1851
1852         found_modified_image = false;
1853         for (int i = 0; i < w->hdr.image_count; i++) {
1854                 if (!found_modified_image)
1855                         found_modified_image = w->image_metadata[i].modified;
1856                 if (found_modified_image) {
1857                         select_wim_image(w, i + 1);
1858                         ret = write_metadata_resource(w);
1859                         if (ret != 0)
1860                                 goto out_ftruncate;
1861                 }
1862         }
1863         write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
1864         ret = finish_write(w, WIMLIB_ALL_IMAGES, write_flags,
1865                            progress_func);
1866 out_ftruncate:
1867         close_wim_writable(w);
1868         if (ret != 0 && !(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
1869                 WARNING("Truncating `%s' to its original size (%"PRIu64" bytes)",
1870                         w->filename, old_wim_end);
1871                 /* Return value of truncate() is ignored because this is already
1872                  * an error path. */
1873                 (void)truncate(w->filename, old_wim_end);
1874         }
1875         w->wim_locked = 0;
1876         return ret;
1877 }
1878
1879 static int overwrite_wim_via_tmpfile(WIMStruct *w, int write_flags,
1880                                      unsigned num_threads,
1881                                      wimlib_progress_func_t progress_func)
1882 {
1883         size_t wim_name_len;
1884         int ret;
1885
1886         DEBUG("Overwriting `%s' via a temporary file", w->filename);
1887
1888         /* Write the WIM to a temporary file in the same directory as the
1889          * original WIM. */
1890         wim_name_len = strlen(w->filename);
1891         char tmpfile[wim_name_len + 10];
1892         memcpy(tmpfile, w->filename, wim_name_len);
1893         randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
1894         tmpfile[wim_name_len + 9] = '\0';
1895
1896         ret = wimlib_write(w, tmpfile, WIMLIB_ALL_IMAGES,
1897                            write_flags | WIMLIB_WRITE_FLAG_FSYNC,
1898                            num_threads, progress_func);
1899         if (ret != 0) {
1900                 ERROR("Failed to write the WIM file `%s'", tmpfile);
1901                 goto err;
1902         }
1903
1904         DEBUG("Renaming `%s' to `%s'", tmpfile, w->filename);
1905
1906         /* Rename the new file to the old file .*/
1907         if (rename(tmpfile, w->filename) != 0) {
1908                 ERROR_WITH_ERRNO("Failed to rename `%s' to `%s'",
1909                                  tmpfile, w->filename);
1910                 ret = WIMLIB_ERR_RENAME;
1911                 goto err;
1912         }
1913
1914         if (progress_func) {
1915                 union wimlib_progress_info progress;
1916                 progress.rename.from = tmpfile;
1917                 progress.rename.to = w->filename;
1918                 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
1919         }
1920
1921         /* Close the original WIM file that was opened for reading. */
1922         if (w->fp != NULL) {
1923                 fclose(w->fp);
1924                 w->fp = NULL;
1925         }
1926
1927         /* Re-open the WIM read-only. */
1928         w->fp = fopen(w->filename, "rb");
1929         if (w->fp == NULL) {
1930                 ret = WIMLIB_ERR_REOPEN;
1931                 WARNING_WITH_ERRNO("Failed to re-open `%s' read-only",
1932                                    w->filename);
1933                 FREE(w->filename);
1934                 w->filename = NULL;
1935         }
1936         return ret;
1937 err:
1938         /* Remove temporary file. */
1939         if (unlink(tmpfile) != 0)
1940                 WARNING_WITH_ERRNO("Failed to remove `%s'", tmpfile);
1941         return ret;
1942 }
1943
1944 /*
1945  * Writes a WIM file to the original file that it was read from, overwriting it.
1946  */
1947 WIMLIBAPI int wimlib_overwrite(WIMStruct *w, int write_flags,
1948                                unsigned num_threads,
1949                                wimlib_progress_func_t progress_func)
1950 {
1951         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
1952
1953         if (!w->filename)
1954                 return WIMLIB_ERR_NO_FILENAME;
1955
1956         if (w->hdr.total_parts != 1) {
1957                 ERROR("Cannot modify a split WIM");
1958                 return WIMLIB_ERR_SPLIT_UNSUPPORTED;
1959         }
1960
1961         if ((!w->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
1962             && !(write_flags & WIMLIB_WRITE_FLAG_REBUILD))
1963         {
1964                 int ret;
1965                 ret = overwrite_wim_inplace(w, write_flags, num_threads,
1966                                             progress_func);
1967                 if (ret == WIMLIB_ERR_RESOURCE_ORDER)
1968                         WARNING("Falling back to re-building entire WIM");
1969                 else
1970                         return ret;
1971         }
1972         return overwrite_wim_via_tmpfile(w, write_flags, num_threads,
1973                                          progress_func);
1974 }