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 #ifdef HAVE_CONFIG_H
  28 #  include "config.h"
  29 #endif
  30
  31 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
  32 /* On BSD, this should be included before "wimlib/list.h" so that "wimlib/list.h" can
  33  * overwrite the LIST_HEAD macro. */
  34 #  include <sys/file.h>
  35 #endif
  36
  37 #include "wimlib/endianness.h"
  38 #include "wimlib/error.h"
  39 #include "wimlib/file_io.h"
  40 #include "wimlib/header.h"
  41 #include "wimlib/integrity.h"
  42 #include "wimlib/lookup_table.h"
  43 #include "wimlib/metadata.h"
  44 #include "wimlib/resource.h"
  45 #include "wimlib/write.h"
  46 #include "wimlib/xml.h"
  47
  48 #ifdef __WIN32__
  49 #  include "wimlib/win32.h" /* win32_get_number_of_processors() */
  50 #endif
  51
  52 #ifdef ENABLE_MULTITHREADED_COMPRESSION
  53 #  include <pthread.h>
  54 #endif
  55
  56 #include <errno.h>
  57 #include <fcntl.h>
  58 #include <limits.h>
  59 #include <stdlib.h>
  60 #include <unistd.h>
  61
  62 #ifdef HAVE_ALLOCA_H
  63 #  include <alloca.h>
  64 #endif
  65
  66
  67 #ifndef __WIN32__
  68 #  include <sys/uio.h> /* for `struct iovec' */
  69 #endif
  70
  71 static unsigned
  72 compress_chunk(const void * uncompressed_data,
  73                unsigned uncompressed_len,
  74                void *compressed_data,
  75                int out_ctype,
  76                struct wimlib_lzx_context *comp_ctx)
  77 {
  78         switch (out_ctype) {
  79         case WIMLIB_COMPRESSION_TYPE_XPRESS:
  80                 return wimlib_xpress_compress(uncompressed_data,
  81                                               uncompressed_len,
  82                                               compressed_data);
  83         case WIMLIB_COMPRESSION_TYPE_LZX:
  84                 return wimlib_lzx_compress2(uncompressed_data,
  85                                             uncompressed_len,
  86                                             compressed_data,
  87                                             comp_ctx);
  88         default:
  89                 wimlib_assert(0);
  90                 return 0;
  91         }
  92 }
  93
  94 /* Chunk table that's located at the beginning of each compressed resource in
  95  * the WIM.  (This is not the on-disk format; the on-disk format just has an
  96  * array of offsets.) */
  97 struct chunk_table {
  98         u64 original_resource_size;
  99         u64 num_chunks;
 100         u64 table_disk_size;
 101         unsigned bytes_per_chunk_entry;
 102         void *cur_offset_p;
 103         union {
 104                 u32 cur_offset_u32;
 105                 u64 cur_offset_u64;
 106         };
 107         /* Beginning of chunk offsets, in either 32-bit or 64-bit little endian
 108          * integers, including the first offset of 0, which will not be written.
 109          * */
 110         u8 offsets[] _aligned_attribute(8);
 111 };
 112
 113 /* Allocate and initializes a chunk table, then reserve space for it in the
 114  * output file unless writing a pipable resource.  */
 115 static int
 116 begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
 117                              struct filedes *out_fd,
 118                              struct chunk_table **chunk_tab_ret,
 119                              int resource_flags)
 120 {
 121         u64 size;
 122         u64 num_chunks;
 123         unsigned bytes_per_chunk_entry;
 124         size_t alloc_size;
 125         struct chunk_table *chunk_tab;
 126         int ret;
 127
 128         size = wim_resource_size(lte);
 129         num_chunks = wim_resource_chunks(lte);
 130         bytes_per_chunk_entry = (size > (1ULL << 32)) ? 8 : 4;
 131         alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
 132         chunk_tab = CALLOC(1, alloc_size);
 133
 134         if (!chunk_tab) {
 135                 ERROR("Failed to allocate chunk table for %"PRIu64" byte "
 136                       "resource", size);
 137                 return WIMLIB_ERR_NOMEM;
 138         }
 139         chunk_tab->num_chunks = num_chunks;
 140         chunk_tab->original_resource_size = size;
 141         chunk_tab->bytes_per_chunk_entry = bytes_per_chunk_entry;
 142         chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
 143                                      (num_chunks - 1);
 144         chunk_tab->cur_offset_p = chunk_tab->offsets;
 145
 146         /* We don't know the correct offsets yet; so just write zeroes to
 147          * reserve space for the table, so we can go back to it later after
 148          * we've written the compressed chunks following it.
 149          *
 150          * Special case: if writing a pipable WIM, compressed resources are in a
 151          * modified format (see comment above write_pipable_wim()) and do not
 152          * have a chunk table at the beginning, so don't reserve any space for
 153          * one.  */
 154         if (!(resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE)) {
 155                 ret = full_write(out_fd, chunk_tab->offsets,
 156                                  chunk_tab->table_disk_size);
 157                 if (ret) {
 158                         ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
 159                                          "file resource");
 160                         FREE(chunk_tab);
 161                         return ret;
 162                 }
 163         }
 164         *chunk_tab_ret = chunk_tab;
 165         return 0;
 166 }
 167
 168 /* Add the offset for the next chunk to the chunk table being constructed for a
 169  * compressed stream. */
 170 static void
 171 chunk_tab_record_chunk(struct chunk_table *chunk_tab, unsigned out_chunk_size)
 172 {
 173         if (chunk_tab->bytes_per_chunk_entry == 4) {
 174                 *(le32*)chunk_tab->cur_offset_p = cpu_to_le32(chunk_tab->cur_offset_u32);
 175                 chunk_tab->cur_offset_p = (le32*)chunk_tab->cur_offset_p + 1;
 176                 chunk_tab->cur_offset_u32 += out_chunk_size;
 177         } else {
 178                 *(le64*)chunk_tab->cur_offset_p = cpu_to_le64(chunk_tab->cur_offset_u64);
 179                 chunk_tab->cur_offset_p = (le64*)chunk_tab->cur_offset_p + 1;
 180                 chunk_tab->cur_offset_u64 += out_chunk_size;
 181         }
 182 }
 183
 184 /* Finishes a WIM chunk table and writes it to the output file at the correct
 185  * offset.  */
 186 static int
 187 finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
 188                               struct filedes *out_fd,
 189                               off_t res_start_offset,
 190                               int write_resource_flags)
 191 {
 192         int ret;
 193
 194         if (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
 195                 ret = full_write(out_fd,
 196                                  chunk_tab->offsets +
 197                                          chunk_tab->bytes_per_chunk_entry,
 198                                  chunk_tab->table_disk_size);
 199         } else {
 200                 ret  = full_pwrite(out_fd,
 201                                    chunk_tab->offsets +
 202                                            chunk_tab->bytes_per_chunk_entry,
 203                                    chunk_tab->table_disk_size,
 204                                    res_start_offset);
 205         }
 206         if (ret) {
 207                 ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
 208                                  "file resource");
 209         }
 210         return ret;
 211 }
 212
 213 /* Write the header for a stream in a pipable WIM.
 214  */
 215 static int
 216 write_pwm_stream_header(const struct wim_lookup_table_entry *lte,
 217                         struct filedes *out_fd,
 218                         int additional_reshdr_flags)
 219 {
 220         struct pwm_stream_hdr stream_hdr;
 221         u32 reshdr_flags;
 222         int ret;
 223
 224         stream_hdr.magic = PWM_STREAM_MAGIC;
 225         stream_hdr.uncompressed_size = cpu_to_le64(lte->resource_entry.original_size);
 226         if (additional_reshdr_flags & PWM_RESHDR_FLAG_UNHASHED) {
 227                 zero_out_hash(stream_hdr.hash);
 228         } else {
 229                 wimlib_assert(!lte->unhashed);
 230                 copy_hash(stream_hdr.hash, lte->hash);
 231         }
 232
 233         reshdr_flags = lte->resource_entry.flags & ~WIM_RESHDR_FLAG_COMPRESSED;
 234         reshdr_flags |= additional_reshdr_flags;
 235         stream_hdr.flags = cpu_to_le32(reshdr_flags);
 236         ret = full_write(out_fd, &stream_hdr, sizeof(stream_hdr));
 237         if (ret)
 238                 ERROR_WITH_ERRNO("Error writing stream header");
 239         return ret;
 240 }
 241
 242 static int
 243 seek_and_truncate(struct filedes *out_fd, off_t offset)
 244 {
 245         if (filedes_seek(out_fd, offset) == -1 ||
 246             ftruncate(out_fd->fd, offset))
 247         {
 248                 ERROR_WITH_ERRNO("Failed to truncate output WIM file");
 249                 return WIMLIB_ERR_WRITE;
 250         }
 251         return 0;
 252 }
 253
 254 static int
 255 finalize_and_check_sha1(SHA_CTX *sha_ctx, struct wim_lookup_table_entry *lte)
 256 {
 257         u8 md[SHA1_HASH_SIZE];
 258
 259         sha1_final(md, sha_ctx);
 260         if (lte->unhashed) {
 261                 copy_hash(lte->hash, md);
 262         } else if (!hashes_equal(md, lte->hash)) {
 263                 ERROR("WIM resource has incorrect hash!");
 264                 if (lte_filename_valid(lte)) {
 265                         ERROR("We were reading it from \"%"TS"\"; maybe "
 266                               "it changed while we were reading it.",
 267                               lte->file_on_disk);
 268                 }
 269                 return WIMLIB_ERR_INVALID_RESOURCE_HASH;
 270         }
 271         return 0;
 272 }
 273
 274 struct write_resource_ctx {
 275         int out_ctype;
 276         struct wimlib_lzx_context *comp_ctx;
 277         struct chunk_table *chunk_tab;
 278         struct filedes *out_fd;
 279         SHA_CTX sha_ctx;
 280         bool doing_sha;
 281         int resource_flags;
 282 };
 283
 284 static int
 285 write_resource_cb(const void *chunk, size_t chunk_size, void *_ctx)
 286 {
 287         struct write_resource_ctx *ctx = _ctx;
 288         const void *out_chunk;
 289         unsigned out_chunk_size;
 290         int ret;
 291
 292         if (ctx->doing_sha)
 293                 sha1_update(&ctx->sha_ctx, chunk, chunk_size);
 294
 295         out_chunk = chunk;
 296         out_chunk_size = chunk_size;
 297         if (ctx->out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
 298                 void *compressed_chunk;
 299                 unsigned compressed_size;
 300
 301                 /* Compress the chunk.  */
 302                 compressed_chunk = alloca(chunk_size);
 303
 304                 compressed_size = compress_chunk(chunk, chunk_size,
 305                                                  compressed_chunk,
 306                                                  ctx->out_ctype,
 307                                                  ctx->comp_ctx);
 308                 /* Use compressed data if compression to less than input size
 309                  * was successful.  */
 310                 if (compressed_size) {
 311                         out_chunk = compressed_chunk;
 312                         out_chunk_size = compressed_size;
 313                 }
 314         }
 315
 316         if (ctx->chunk_tab) {
 317                 /* Update chunk table accounting.  */
 318                 chunk_tab_record_chunk(ctx->chunk_tab, out_chunk_size);
 319
 320                 /* If writing compressed chunks to a pipable WIM, before the
 321                  * chunk data write a chunk header that provides the compressed
 322                  * chunk size.  */
 323                 if (ctx->resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
 324                         struct pwm_chunk_hdr chunk_hdr = {
 325                                 .compressed_size = cpu_to_le32(out_chunk_size),
 326                         };
 327                         ret = full_write(ctx->out_fd, &chunk_hdr,
 328                                          sizeof(chunk_hdr));
 329                         if (ret)
 330                                 goto error;
 331                 }
 332         }
 333
 334         /* Write the chunk data.  */
 335         ret = full_write(ctx->out_fd, out_chunk, out_chunk_size);
 336         if (ret)
 337                 goto error;
 338         return 0;
 339
 340 error:
 341         ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
 342         return ret;
 343 }
 344
 345 /*
 346  * write_wim_resource()-
 347  *
 348  * Write a resource to an output WIM.
 349  *
 350  * @lte:
 351  *      Lookup table entry for the resource, which could be in another WIM, in
 352  *      an external file, or in another location.
 353  *
 354  * @out_fd:
 355  *      File descriptor opened to the output WIM.
 356  *
 357  * @out_ctype:
 358  *      One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate which
 359  *      compression algorithm to use.
 360  *
 361  * @out_res_entry:
 362  *      On success, this is filled in with the offset, flags, compressed size,
 363  *      and uncompressed size of the resource in the output WIM.
 364  *
 365  * @resource_flags:
 366  *      * WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed even
 367  *        if it could otherwise be copied directly from the input;
 368  *      * WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE if writing a resource for a pipable WIM
 369  *        (and the output file descriptor may be a pipe).
 370  *
 371  * @comp_ctx:
 372  *      Location of LZX compression context pointer, which will be allocated or
 373  *      updated if needed.  (Initialize to NULL.)
 374  *
 375  * Additional notes:  The SHA1 message digest of the uncompressed data is
 376  * calculated (except when doing a raw copy --- see below).  If the @unhashed
 377  * flag is set on the lookup table entry, this message digest is simply copied
 378  * to it; otherwise, the message digest is compared with the existing one, and
 379  * the function will fail if they do not match.
 380  */
 381 int
 382 write_wim_resource(struct wim_lookup_table_entry *lte,
 383                    struct filedes *out_fd, int out_ctype,
 384                    struct resource_entry *out_res_entry,
 385                    int resource_flags,
 386                    struct wimlib_lzx_context **comp_ctx)
 387 {
 388         struct write_resource_ctx write_ctx;
 389         off_t res_start_offset;
 390         u64 read_size;
 391         int ret;
 392
 393         /* Mask out any irrelevant flags, since this function also uses this
 394          * variable to store WIMLIB_READ_RESOURCE flags.  */
 395         resource_flags &= WIMLIB_WRITE_RESOURCE_MASK;
 396
 397         /* Get current position in output WIM.  */
 398         res_start_offset = out_fd->offset;
 399
 400         /* If we are not forcing the data to be recompressed, and the input
 401          * resource is located in a WIM with the same compression type as that
 402          * desired other than no compression, we can simply copy the compressed
 403          * data without recompressing it.  This also means we must skip
 404          * calculating the SHA1, as we never will see the uncompressed data.  */
 405         if (lte->resource_location == RESOURCE_IN_WIM &&
 406             out_ctype == wim_resource_compression_type(lte) &&
 407             out_ctype != WIMLIB_COMPRESSION_TYPE_NONE &&
 408             !(resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS))
 409         {
 410                 /* Normally we can request a RAW_FULL read, but if we're reading
 411                  * from a pipable resource and writing a non-pipable resource or
 412                  * vice versa, then a RAW_CHUNKS read needs to be requested so
 413                  * that the written resource can be appropriately formatted.
 414                  * However, in neither case is any actual decompression needed.
 415                  */
 416                 if (lte->is_pipable == !!(resource_flags &
 417                                           WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE))
 418                         resource_flags |= WIMLIB_READ_RESOURCE_FLAG_RAW_FULL;
 419                 else
 420                         resource_flags |= WIMLIB_READ_RESOURCE_FLAG_RAW_CHUNKS;
 421                 write_ctx.doing_sha = false;
 422                 read_size = lte->resource_entry.size;
 423         } else {
 424                 write_ctx.doing_sha = true;
 425                 sha1_init(&write_ctx.sha_ctx);
 426                 read_size = lte->resource_entry.original_size;
 427         }
 428
 429
 430         /* If the output resource is to be compressed, initialize the chunk
 431          * table and set the function to use for chunk compression.  Exceptions:
 432          * no compression function is needed if doing a raw copy; also, no chunk
 433          * table is needed if doing a *full* (not per-chunk) raw copy.  */
 434         write_ctx.out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
 435         write_ctx.chunk_tab = NULL;
 436         if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
 437                 if (!(resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW)) {
 438                         write_ctx.out_ctype = out_ctype;
 439                         if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX) {
 440                                 ret = wimlib_lzx_alloc_context(NULL, comp_ctx);
 441                                 if (ret)
 442                                         goto out;
 443                         }
 444                         write_ctx.comp_ctx = *comp_ctx;
 445                 }
 446                 if (!(resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW_FULL)) {
 447                         ret = begin_wim_resource_chunk_tab(lte, out_fd,
 448                                                            &write_ctx.chunk_tab,
 449                                                            resource_flags);
 450                         if (ret)
 451                                 goto out;
 452                 }
 453         }
 454
 455         /* If writing a pipable resource, write the stream header and update
 456          * @res_start_offset to be the end of the stream header.  */
 457         if (resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
 458                 int reshdr_flags = 0;
 459                 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE)
 460                         reshdr_flags |= WIM_RESHDR_FLAG_COMPRESSED;
 461                 ret = write_pwm_stream_header(lte, out_fd, reshdr_flags);
 462                 if (ret)
 463                         goto out_free_chunk_tab;
 464                 res_start_offset = out_fd->offset;
 465         }
 466
 467         /* Write the entire resource by reading the entire resource and feeding
 468          * the data through the write_resource_cb function. */
 469         write_ctx.out_fd = out_fd;
 470         write_ctx.resource_flags = resource_flags;
 471 try_write_again:
 472         ret = read_resource_prefix(lte, read_size,
 473                                    write_resource_cb, &write_ctx, resource_flags);
 474         if (ret)
 475                 goto out_free_chunk_tab;
 476
 477         /* Verify SHA1 message digest of the resource, or set the hash for the
 478          * first time. */
 479         if (write_ctx.doing_sha) {
 480                 ret = finalize_and_check_sha1(&write_ctx.sha_ctx, lte);
 481                 if (ret)
 482                         goto out_free_chunk_tab;
 483         }
 484
 485         /* Write chunk table if needed.  */
 486         if (write_ctx.chunk_tab) {
 487                 ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab,
 488                                                     out_fd,
 489                                                     res_start_offset,
 490                                                     resource_flags);
 491                 if (ret)
 492                         goto out_free_chunk_tab;
 493         }
 494
 495         /* Fill in out_res_entry with information about the newly written
 496          * resource.  */
 497         out_res_entry->size          = out_fd->offset - res_start_offset;
 498         out_res_entry->flags         = lte->resource_entry.flags;
 499         if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE)
 500                 out_res_entry->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
 501         else
 502                 out_res_entry->flags |= WIM_RESHDR_FLAG_COMPRESSED;
 503         out_res_entry->offset        = res_start_offset;
 504         out_res_entry->original_size = wim_resource_size(lte);
 505
 506         /* Check for resources compressed to greater than their original size
 507          * and write them uncompressed instead.  (But never do this if writing
 508          * to a pipe, and don't bother if we did a raw copy.)  */
 509         if (out_res_entry->size > out_res_entry->original_size &&
 510             !(resource_flags & (WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE |
 511                                 WIMLIB_READ_RESOURCE_FLAG_RAW)))
 512         {
 513                 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
 514                       "writing uncompressed instead",
 515                       out_res_entry->original_size, out_res_entry->size);
 516                 ret = seek_and_truncate(out_fd, res_start_offset);
 517                 if (ret)
 518                         goto out_free_chunk_tab;
 519                 out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
 520                 FREE(write_ctx.chunk_tab);
 521                 write_ctx.out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
 522                 write_ctx.chunk_tab = NULL;
 523                 write_ctx.doing_sha = false;
 524                 goto try_write_again;
 525         }
 526         if (resource_flags & (WIMLIB_READ_RESOURCE_FLAG_RAW)) {
 527                 DEBUG("Copied raw compressed data "
 528                       "(%"PRIu64" => %"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
 529                       out_res_entry->original_size, out_res_entry->size,
 530                       out_res_entry->offset, out_res_entry->flags);
 531         } else if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
 532                 DEBUG("Wrote compressed resource "
 533                       "(%"PRIu64" => %"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
 534                       out_res_entry->original_size, out_res_entry->size,
 535                       out_res_entry->offset, out_res_entry->flags);
 536         } else {
 537                 DEBUG("Wrote uncompressed resource "
 538                       "(%"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
 539                       out_res_entry->original_size,
 540                       out_res_entry->offset, out_res_entry->flags);
 541         }
 542         ret = 0;
 543 out_free_chunk_tab:
 544         FREE(write_ctx.chunk_tab);
 545 out:
 546         return ret;
 547 }
 548
 549 /* Like write_wim_resource(), but the resource is specified by a buffer of
 550  * uncompressed data rather a lookup table entry; also writes the SHA1 hash of
 551  * the buffer to @hash_ret.  */
 552 int
 553 write_wim_resource_from_buffer(const void *buf, size_t buf_size,
 554                                int reshdr_flags, struct filedes *out_fd,
 555                                int out_ctype,
 556                                struct resource_entry *out_res_entry,
 557                                u8 *hash_ret, int write_resource_flags,
 558                                struct wimlib_lzx_context **comp_ctx)
 559 {
 560         /* Set up a temporary lookup table entry to provide to
 561          * write_wim_resource(). */
 562         struct wim_lookup_table_entry lte;
 563         int ret;
 564
 565         lte.resource_location            = RESOURCE_IN_ATTACHED_BUFFER;
 566         lte.attached_buffer              = (void*)buf;
 567         lte.resource_entry.original_size = buf_size;
 568         lte.resource_entry.flags         = reshdr_flags;
 569
 570         if (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
 571                 sha1_buffer(buf, buf_size, lte.hash);
 572                 lte.unhashed = 0;
 573         } else {
 574                 lte.unhashed = 1;
 575         }
 576
 577         ret = write_wim_resource(&lte, out_fd, out_ctype, out_res_entry,
 578                                  write_resource_flags, comp_ctx);
 579         if (ret)
 580                 return ret;
 581         if (hash_ret)
 582                 copy_hash(hash_ret, lte.hash);
 583         return 0;
 584 }
 585
 586
 587 #ifdef ENABLE_MULTITHREADED_COMPRESSION
 588
 589 /* Blocking shared queue (solves the producer-consumer problem) */
 590 struct shared_queue {
 591         unsigned size;
 592         unsigned front;
 593         unsigned back;
 594         unsigned filled_slots;
 595         void **array;
 596         pthread_mutex_t lock;
 597         pthread_cond_t msg_avail_cond;
 598         pthread_cond_t space_avail_cond;
 599 };
 600
 601 static int
 602 shared_queue_init(struct shared_queue *q, unsigned size)
 603 {
 604         wimlib_assert(size != 0);
 605         q->array = CALLOC(sizeof(q->array[0]), size);
 606         if (!q->array)
 607                 goto err;
 608         q->filled_slots = 0;
 609         q->front = 0;
 610         q->back = size - 1;
 611         q->size = size;
 612         if (pthread_mutex_init(&q->lock, NULL)) {
 613                 ERROR_WITH_ERRNO("Failed to initialize mutex");
 614                 goto err;
 615         }
 616         if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
 617                 ERROR_WITH_ERRNO("Failed to initialize condition variable");
 618                 goto err_destroy_lock;
 619         }
 620         if (pthread_cond_init(&q->space_avail_cond, NULL)) {
 621                 ERROR_WITH_ERRNO("Failed to initialize condition variable");
 622                 goto err_destroy_msg_avail_cond;
 623         }
 624         return 0;
 625 err_destroy_msg_avail_cond:
 626         pthread_cond_destroy(&q->msg_avail_cond);
 627 err_destroy_lock:
 628         pthread_mutex_destroy(&q->lock);
 629 err:
 630         return WIMLIB_ERR_NOMEM;
 631 }
 632
 633 static void
 634 shared_queue_destroy(struct shared_queue *q)
 635 {
 636         FREE(q->array);
 637         pthread_mutex_destroy(&q->lock);
 638         pthread_cond_destroy(&q->msg_avail_cond);
 639         pthread_cond_destroy(&q->space_avail_cond);
 640 }
 641
 642 static void
 643 shared_queue_put(struct shared_queue *q, void *obj)
 644 {
 645         pthread_mutex_lock(&q->lock);
 646         while (q->filled_slots == q->size)
 647                 pthread_cond_wait(&q->space_avail_cond, &q->lock);
 648
 649         q->back = (q->back + 1) % q->size;
 650         q->array[q->back] = obj;
 651         q->filled_slots++;
 652
 653         pthread_cond_broadcast(&q->msg_avail_cond);
 654         pthread_mutex_unlock(&q->lock);
 655 }
 656
 657 static void *
 658 shared_queue_get(struct shared_queue *q)
 659 {
 660         void *obj;
 661
 662         pthread_mutex_lock(&q->lock);
 663         while (q->filled_slots == 0)
 664                 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
 665
 666         obj = q->array[q->front];
 667         q->array[q->front] = NULL;
 668         q->front = (q->front + 1) % q->size;
 669         q->filled_slots--;
 670
 671         pthread_cond_broadcast(&q->space_avail_cond);
 672         pthread_mutex_unlock(&q->lock);
 673         return obj;
 674 }
 675
 676 struct compressor_thread_params {
 677         struct shared_queue *res_to_compress_queue;
 678         struct shared_queue *compressed_res_queue;
 679         int out_ctype;
 680         struct wimlib_lzx_context *comp_ctx;
 681 };
 682
 683 #define MAX_CHUNKS_PER_MSG 2
 684
 685 struct message {
 686         struct wim_lookup_table_entry *lte;
 687         u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
 688         u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
 689         unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
 690         struct iovec out_chunks[MAX_CHUNKS_PER_MSG];
 691         unsigned num_chunks;
 692         struct list_head list;
 693         bool complete;
 694         u64 begin_chunk;
 695 };
 696
 697 static void
 698 compress_chunks(struct message *msg, int out_ctype,
 699                 struct wimlib_lzx_context *comp_ctx)
 700 {
 701         for (unsigned i = 0; i < msg->num_chunks; i++) {
 702                 unsigned len;
 703
 704                 len = compress_chunk(msg->uncompressed_chunks[i],
 705                                      msg->uncompressed_chunk_sizes[i],
 706                                      msg->compressed_chunks[i],
 707                                      out_ctype,
 708                                      comp_ctx);
 709
 710                 void *out_chunk;
 711                 unsigned out_len;
 712                 if (len) {
 713                         /* To be written compressed */
 714                         out_chunk = msg->compressed_chunks[i];
 715                         out_len = len;
 716                 } else {
 717                         /* To be written uncompressed */
 718                         out_chunk = msg->uncompressed_chunks[i];
 719                         out_len = msg->uncompressed_chunk_sizes[i];
 720                 }
 721                 msg->out_chunks[i].iov_base = out_chunk;
 722                 msg->out_chunks[i].iov_len = out_len;
 723         }
 724 }
 725
 726 /* Compressor thread routine.  This is a lot simpler than the main thread
 727  * routine: just repeatedly get a group of chunks from the
 728  * res_to_compress_queue, compress them, and put them in the
 729  * compressed_res_queue.  A NULL pointer indicates that the thread should stop.
 730  * */
 731 static void *
 732 compressor_thread_proc(void *arg)
 733 {
 734         struct compressor_thread_params *params = arg;
 735         struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
 736         struct shared_queue *compressed_res_queue = params->compressed_res_queue;
 737         struct message *msg;
 738
 739         DEBUG("Compressor thread ready");
 740         while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
 741                 compress_chunks(msg, params->out_ctype, params->comp_ctx);
 742                 shared_queue_put(compressed_res_queue, msg);
 743         }
 744         DEBUG("Compressor thread terminating");
 745         return NULL;
 746 }
 747 #endif /* ENABLE_MULTITHREADED_COMPRESSION */
 748
 749 struct write_streams_progress_data {
 750         wimlib_progress_func_t progress_func;
 751         union wimlib_progress_info progress;
 752         uint64_t next_progress;
 753         WIMStruct *prev_wim_part;
 754 };
 755
 756 static void
 757 do_write_streams_progress(struct write_streams_progress_data *progress_data,
 758                           struct wim_lookup_table_entry *lte,
 759                           bool stream_discarded)
 760 {
 761         union wimlib_progress_info *progress = &progress_data->progress;
 762         bool new_wim_part;
 763
 764         if (stream_discarded) {
 765                 progress->write_streams.total_bytes -= wim_resource_size(lte);
 766                 if (progress_data->next_progress != ~(uint64_t)0 &&
 767                     progress_data->next_progress > progress->write_streams.total_bytes)
 768                 {
 769                         progress_data->next_progress = progress->write_streams.total_bytes;
 770                 }
 771         } else {
 772                 progress->write_streams.completed_bytes += wim_resource_size(lte);
 773         }
 774         new_wim_part = false;
 775         if (lte->resource_location == RESOURCE_IN_WIM &&
 776             lte->wim != progress_data->prev_wim_part)
 777         {
 778                 if (progress_data->prev_wim_part) {
 779                         new_wim_part = true;
 780                         progress->write_streams.completed_parts++;
 781                 }
 782                 progress_data->prev_wim_part = lte->wim;
 783         }
 784         progress->write_streams.completed_streams++;
 785         if (progress_data->progress_func
 786             && (progress->write_streams.completed_bytes >= progress_data->next_progress
 787                 || new_wim_part))
 788         {
 789                 progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
 790                                              progress);
 791                 if (progress_data->next_progress == progress->write_streams.total_bytes) {
 792                         progress_data->next_progress = ~(uint64_t)0;
 793                 } else {
 794                         progress_data->next_progress =
 795                                 min(progress->write_streams.total_bytes,
 796                                     progress->write_streams.completed_bytes +
 797                                         progress->write_streams.total_bytes / 100);
 798                 }
 799         }
 800 }
 801
 802 struct serial_write_stream_ctx {
 803         struct filedes *out_fd;
 804         int out_ctype;
 805         struct wimlib_lzx_context **comp_ctx;
 806         int write_resource_flags;
 807 };
 808
 809 static int
 810 serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
 811 {
 812         struct serial_write_stream_ctx *ctx = _ctx;
 813         return write_wim_resource(lte, ctx->out_fd,
 814                                   ctx->out_ctype, &lte->output_resource_entry,
 815                                   ctx->write_resource_flags,
 816                                   ctx->comp_ctx);
 817 }
 818
 819
 820 /* Write a list of streams, taking into account that some streams may be
 821  * duplicates that are checksummed and discarded on the fly, and also delegating
 822  * the actual writing of a stream to a function @write_stream_cb, which is
 823  * passed the context @write_stream_ctx. */
 824 static int
 825 do_write_stream_list(struct list_head *stream_list,
 826                      struct wim_lookup_table *lookup_table,
 827                      int (*write_stream_cb)(struct wim_lookup_table_entry *, void *),
 828                      void *write_stream_ctx,
 829                      struct write_streams_progress_data *progress_data)
 830 {
 831         int ret = 0;
 832         struct wim_lookup_table_entry *lte;
 833         bool stream_discarded;
 834
 835         /* For each stream in @stream_list ... */
 836         while (!list_empty(stream_list)) {
 837                 stream_discarded = false;
 838                 lte = container_of(stream_list->next,
 839                                    struct wim_lookup_table_entry,
 840                                    write_streams_list);
 841                 list_del(&lte->write_streams_list);
 842                 if (lte->unhashed && !lte->unique_size) {
 843                         /* Unhashed stream that shares a size with some other
 844                          * stream in the WIM we are writing.  The stream must be
 845                          * checksummed to know if we need to write it or not. */
 846                         struct wim_lookup_table_entry *tmp;
 847                         u32 orig_out_refcnt = lte->out_refcnt;
 848
 849                         ret = hash_unhashed_stream(lte, lookup_table, &tmp);
 850                         if (ret)
 851                                 break;
 852                         if (tmp != lte) {
 853                                 /* We found a duplicate stream.  'lte' was
 854                                  * freed, so replace it with the duplicate.  */
 855                                 lte = tmp;
 856
 857                                 /* 'out_refcnt' was transferred to the
 858                                  * duplicate, and we can detect if the duplicate
 859                                  * stream was already referenced for writing by
 860                                  * checking if its 'out_refcnt' is higher than
 861                                  * that of the original stream.  In such cases,
 862                                  * the current stream can be discarded.  We can
 863                                  * also discard the current stream if it was
 864                                  * previously marked as filtered (e.g. already
 865                                  * present in the WIM being written).  */
 866                                 if (lte->out_refcnt > orig_out_refcnt ||
 867                                     lte->filtered) {
 868                                         DEBUG("Discarding duplicate stream of "
 869                                               "length %"PRIu64,
 870                                               wim_resource_size(lte));
 871                                         lte->no_progress = 0;
 872                                         stream_discarded = true;
 873                                         goto skip_to_progress;
 874                                 }
 875                         }
 876                 }
 877
 878                 /* Here, @lte is either a hashed stream or an unhashed stream
 879                  * with a unique size.  In either case we know that the stream
 880                  * has to be written.  In either case the SHA1 message digest
 881                  * will be calculated over the stream while writing it; however,
 882                  * in the former case this is done merely to check the data,
 883                  * while in the latter case this is done because we do not have
 884                  * the SHA1 message digest yet.  */
 885                 wimlib_assert(lte->out_refcnt != 0);
 886                 lte->deferred = 0;
 887                 lte->no_progress = 0;
 888                 ret = (*write_stream_cb)(lte, write_stream_ctx);
 889                 if (ret)
 890                         break;
 891                 /* In parallel mode, some streams are deferred for later,
 892                  * serialized processing; ignore them here. */
 893                 if (lte->deferred)
 894                         continue;
 895                 if (lte->unhashed) {
 896                         list_del(&lte->unhashed_list);
 897                         lookup_table_insert(lookup_table, lte);
 898                         lte->unhashed = 0;
 899                 }
 900         skip_to_progress:
 901                 if (!lte->no_progress) {
 902                         do_write_streams_progress(progress_data,
 903                                                   lte, stream_discarded);
 904                 }
 905         }
 906         return ret;
 907 }
 908
 909 static int
 910 do_write_stream_list_serial(struct list_head *stream_list,
 911                             struct wim_lookup_table *lookup_table,
 912                             struct filedes *out_fd,
 913                             int out_ctype,
 914                             struct wimlib_lzx_context **comp_ctx,
 915                             int write_resource_flags,
 916                             struct write_streams_progress_data *progress_data)
 917 {
 918         struct serial_write_stream_ctx ctx = {
 919                 .out_fd = out_fd,
 920                 .out_ctype = out_ctype,
 921                 .write_resource_flags = write_resource_flags,
 922                 .comp_ctx = comp_ctx,
 923         };
 924         return do_write_stream_list(stream_list,
 925                                     lookup_table,
 926                                     serial_write_stream,
 927                                     &ctx,
 928                                     progress_data);
 929 }
 930
 931 static inline int
 932 write_flags_to_resource_flags(int write_flags)
 933 {
 934         int resource_flags = 0;
 935
 936         if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
 937                 resource_flags |= WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS;
 938         if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
 939                 resource_flags |= WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE;
 940         return resource_flags;
 941 }
 942
 943 static int
 944 write_stream_list_serial(struct list_head *stream_list,
 945                          struct wim_lookup_table *lookup_table,
 946                          struct filedes *out_fd,
 947                          int out_ctype,
 948                          struct wimlib_lzx_context **comp_ctx,
 949                          int write_resource_flags,
 950                          struct write_streams_progress_data *progress_data)
 951 {
 952         union wimlib_progress_info *progress = &progress_data->progress;
 953         DEBUG("Writing stream list of size %"PRIu64" (serial version)",
 954               progress->write_streams.total_streams);
 955         progress->write_streams.num_threads = 1;
 956         if (progress_data->progress_func) {
 957                 progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
 958                                              progress);
 959         }
 960         return do_write_stream_list_serial(stream_list,
 961                                            lookup_table,
 962                                            out_fd,
 963                                            out_ctype,
 964                                            comp_ctx,
 965                                            write_resource_flags,
 966                                            progress_data);
 967 }
 968
 969 #ifdef ENABLE_MULTITHREADED_COMPRESSION
 970 static int
 971 write_wim_chunks(struct message *msg, struct filedes *out_fd,
 972                  struct chunk_table *chunk_tab,
 973                  int write_resource_flags)
 974 {
 975         struct iovec *vecs;
 976         struct pwm_chunk_hdr *chunk_hdrs;
 977         unsigned nvecs;
 978         int ret;
 979
 980         for (unsigned i = 0; i < msg->num_chunks; i++)
 981                 chunk_tab_record_chunk(chunk_tab, msg->out_chunks[i].iov_len);
 982
 983         if (!(write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE)) {
 984                 nvecs = msg->num_chunks;
 985                 vecs = msg->out_chunks;
 986         } else {
 987                 /* Special case:  If writing a compressed resource to a pipable
 988                  * WIM, prefix each compressed chunk with a header that gives
 989                  * its compressed size.  */
 990                 nvecs = msg->num_chunks * 2;
 991                 vecs = alloca(nvecs * sizeof(vecs[0]));
 992                 chunk_hdrs = alloca(msg->num_chunks * sizeof(chunk_hdrs[0]));
 993
 994                 for (unsigned i = 0; i < msg->num_chunks; i++) {
 995                         chunk_hdrs[i].compressed_size = cpu_to_le32(msg->out_chunks[i].iov_len);
 996                         vecs[i * 2].iov_base = &chunk_hdrs[i];
 997                         vecs[i * 2].iov_len = sizeof(chunk_hdrs[i]);
 998                         vecs[i * 2 + 1].iov_base = msg->out_chunks[i].iov_base;
 999                         vecs[i * 2 + 1].iov_len = msg->out_chunks[i].iov_len;
1000                 }
1001         }
1002         ret = full_writev(out_fd, vecs, nvecs);
1003         if (ret)
1004                 ERROR_WITH_ERRNO("Failed to write WIM chunks");
1005         return ret;
1006 }
1007
1008 struct main_writer_thread_ctx {
1009         struct list_head *stream_list;
1010         struct wim_lookup_table *lookup_table;
1011         struct filedes *out_fd;
1012         off_t res_start_offset;
1013         int out_ctype;
1014         struct wimlib_lzx_context **comp_ctx;
1015         int write_resource_flags;
1016         struct shared_queue *res_to_compress_queue;
1017         struct shared_queue *compressed_res_queue;
1018         size_t num_messages;
1019         struct write_streams_progress_data *progress_data;
1020
1021         struct list_head available_msgs;
1022         struct list_head outstanding_streams;
1023         struct list_head serial_streams;
1024         size_t num_outstanding_messages;
1025
1026         SHA_CTX next_sha_ctx;
1027         u64 next_chunk;
1028         u64 next_num_chunks;
1029         struct wim_lookup_table_entry *next_lte;
1030
1031         struct message *msgs;
1032         struct message *next_msg;
1033         struct chunk_table *cur_chunk_tab;
1034 };
1035
1036 static int
1037 init_message(struct message *msg)
1038 {
1039         for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
1040                 msg->compressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
1041                 msg->uncompressed_chunks[i] = MALLOC(WIM_CHUNK_SIZE);
1042                 if (msg->compressed_chunks[i] == NULL ||
1043                     msg->uncompressed_chunks[i] == NULL)
1044                         return WIMLIB_ERR_NOMEM;
1045         }
1046         return 0;
1047 }
1048
1049 static void
1050 destroy_message(struct message *msg)
1051 {
1052         for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
1053                 FREE(msg->compressed_chunks[i]);
1054                 FREE(msg->uncompressed_chunks[i]);
1055         }
1056 }
1057
1058 static void
1059 free_messages(struct message *msgs, size_t num_messages)
1060 {
1061         if (msgs) {
1062                 for (size_t i = 0; i < num_messages; i++)
1063                         destroy_message(&msgs[i]);
1064                 FREE(msgs);
1065         }
1066 }
1067
1068 static struct message *
1069 allocate_messages(size_t num_messages)
1070 {
1071         struct message *msgs;
1072
1073         msgs = CALLOC(num_messages, sizeof(struct message));
1074         if (!msgs)
1075                 return NULL;
1076         for (size_t i = 0; i < num_messages; i++) {
1077                 if (init_message(&msgs[i])) {
1078                         free_messages(msgs, num_messages);
1079                         return NULL;
1080                 }
1081         }
1082         return msgs;
1083 }
1084
1085 static void
1086 main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
1087 {
1088         while (ctx->num_outstanding_messages--)
1089                 shared_queue_get(ctx->compressed_res_queue);
1090         free_messages(ctx->msgs, ctx->num_messages);
1091         FREE(ctx->cur_chunk_tab);
1092 }
1093
1094 static int
1095 main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
1096 {
1097         /* Pre-allocate all the buffers that will be needed to do the chunk
1098          * compression. */
1099         ctx->msgs = allocate_messages(ctx->num_messages);
1100         if (!ctx->msgs)
1101                 return WIMLIB_ERR_NOMEM;
1102
1103         /* Initially, all the messages are available to use. */
1104         INIT_LIST_HEAD(&ctx->available_msgs);
1105         for (size_t i = 0; i < ctx->num_messages; i++)
1106                 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
1107
1108         /* outstanding_streams is the list of streams that currently have had
1109          * chunks sent off for compression.
1110          *
1111          * The first stream in outstanding_streams is the stream that is
1112          * currently being written.
1113          *
1114          * The last stream in outstanding_streams is the stream that is
1115          * currently being read and having chunks fed to the compressor threads.
1116          * */
1117         INIT_LIST_HEAD(&ctx->outstanding_streams);
1118         ctx->num_outstanding_messages = 0;
1119
1120         ctx->next_msg = NULL;
1121
1122         /* Resources that don't need any chunks compressed are added to this
1123          * list and written directly by the main thread. */
1124         INIT_LIST_HEAD(&ctx->serial_streams);
1125
1126         ctx->cur_chunk_tab = NULL;
1127
1128         return 0;
1129 }
1130
1131 static int
1132 receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
1133 {
1134         struct message *msg;
1135         struct wim_lookup_table_entry *cur_lte;
1136         int ret;
1137
1138         wimlib_assert(!list_empty(&ctx->outstanding_streams));
1139         wimlib_assert(ctx->num_outstanding_messages != 0);
1140
1141         cur_lte = container_of(ctx->outstanding_streams.next,
1142                                struct wim_lookup_table_entry,
1143                                being_compressed_list);
1144
1145         /* Get the next message from the queue and process it.
1146          * The message will contain 1 or more data chunks that have been
1147          * compressed. */
1148         msg = shared_queue_get(ctx->compressed_res_queue);
1149         msg->complete = true;
1150         --ctx->num_outstanding_messages;
1151
1152         /* Is this the next chunk in the current resource?  If it's not
1153          * (i.e., an earlier chunk in a same or different resource
1154          * hasn't been compressed yet), do nothing, and keep this
1155          * message around until all earlier chunks are received.
1156          *
1157          * Otherwise, write all the chunks we can. */
1158         while (cur_lte != NULL &&
1159                !list_empty(&cur_lte->msg_list)
1160                && (msg = container_of(cur_lte->msg_list.next,
1161                                       struct message,
1162                                       list))->complete)
1163         {
1164                 list_move(&msg->list, &ctx->available_msgs);
1165                 if (msg->begin_chunk == 0) {
1166                         /* First set of chunks.  */
1167
1168                         /* Write pipable WIM stream header if needed.  */
1169                         if (ctx->write_resource_flags &
1170                             WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE)
1171                         {
1172                                 ret = write_pwm_stream_header(cur_lte, ctx->out_fd,
1173                                                               WIM_RESHDR_FLAG_COMPRESSED);
1174                                 if (ret)
1175                                         return ret;
1176                         }
1177
1178                         /* Save current offset.  */
1179                         ctx->res_start_offset = ctx->out_fd->offset;
1180
1181                         /* Begin building the chunk table, and leave space for
1182                          * it if needed.  */
1183                         ret = begin_wim_resource_chunk_tab(cur_lte,
1184                                                            ctx->out_fd,
1185                                                            &ctx->cur_chunk_tab,
1186                                                            ctx->write_resource_flags);
1187                         if (ret)
1188                                 return ret;
1189
1190                 }
1191
1192                 /* Write the compressed chunks from the message. */
1193                 ret = write_wim_chunks(msg, ctx->out_fd, ctx->cur_chunk_tab,
1194                                        ctx->write_resource_flags);
1195                 if (ret)
1196                         return ret;
1197
1198                 /* Was this the last chunk of the stream?  If so, finish
1199                  * it. */
1200                 if (list_empty(&cur_lte->msg_list) &&
1201                     msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
1202                 {
1203                         u64 res_csize;
1204
1205                         ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
1206                                                             ctx->out_fd,
1207                                                             ctx->res_start_offset,
1208                                                             ctx->write_resource_flags);
1209                         if (ret)
1210                                 return ret;
1211
1212                         list_del(&cur_lte->being_compressed_list);
1213
1214                         res_csize = ctx->out_fd->offset - ctx->res_start_offset;
1215
1216                         FREE(ctx->cur_chunk_tab);
1217                         ctx->cur_chunk_tab = NULL;
1218
1219                         /* Check for resources compressed to greater than or
1220                          * equal to their original size and write them
1221                          * uncompressed instead.  (But never do this if writing
1222                          * to a pipe.)  */
1223                         if (res_csize >= wim_resource_size(cur_lte) &&
1224                             !(ctx->write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE))
1225                         {
1226                                 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
1227                                       "writing uncompressed instead",
1228                                       wim_resource_size(cur_lte), res_csize);
1229                                 ret = seek_and_truncate(ctx->out_fd, ctx->res_start_offset);
1230                                 if (ret)
1231                                         return ret;
1232                                 ret = write_wim_resource(cur_lte,
1233                                                          ctx->out_fd,
1234                                                          WIMLIB_COMPRESSION_TYPE_NONE,
1235                                                          &cur_lte->output_resource_entry,
1236                                                          ctx->write_resource_flags,
1237                                                          ctx->comp_ctx);
1238                                 if (ret)
1239                                         return ret;
1240                         } else {
1241                                 cur_lte->output_resource_entry.size =
1242                                         res_csize;
1243
1244                                 cur_lte->output_resource_entry.original_size =
1245                                         cur_lte->resource_entry.original_size;
1246
1247                                 cur_lte->output_resource_entry.offset =
1248                                         ctx->res_start_offset;
1249
1250                                 cur_lte->output_resource_entry.flags =
1251                                         cur_lte->resource_entry.flags |
1252                                                 WIM_RESHDR_FLAG_COMPRESSED;
1253
1254                                 DEBUG("Wrote compressed resource "
1255                                       "(%"PRIu64" => %"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
1256                                       cur_lte->output_resource_entry.original_size,
1257                                       cur_lte->output_resource_entry.size,
1258                                       cur_lte->output_resource_entry.offset,
1259                                       cur_lte->output_resource_entry.flags);
1260                         }
1261
1262                         do_write_streams_progress(ctx->progress_data,
1263                                                   cur_lte, false);
1264
1265                         /* Since we just finished writing a stream, write any
1266                          * streams that have been added to the serial_streams
1267                          * list for direct writing by the main thread (e.g.
1268                          * resources that don't need to be compressed because
1269                          * the desired compression type is the same as the
1270                          * previous compression type). */
1271                         if (!list_empty(&ctx->serial_streams)) {
1272                                 ret = do_write_stream_list_serial(&ctx->serial_streams,
1273                                                                   ctx->lookup_table,
1274                                                                   ctx->out_fd,
1275                                                                   ctx->out_ctype,
1276                                                                   ctx->comp_ctx,
1277                                                                   ctx->write_resource_flags,
1278                                                                   ctx->progress_data);
1279                                 if (ret)
1280                                         return ret;
1281                         }
1282
1283                         /* Advance to the next stream to write. */
1284                         if (list_empty(&ctx->outstanding_streams)) {
1285                                 cur_lte = NULL;
1286                         } else {
1287                                 cur_lte = container_of(ctx->outstanding_streams.next,
1288                                                        struct wim_lookup_table_entry,
1289                                                        being_compressed_list);
1290                         }
1291                 }
1292         }
1293         return 0;
1294 }
1295
1296 /* Called when the main thread has read a new chunk of data. */
1297 static int
1298 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
1299 {
1300         struct main_writer_thread_ctx *ctx = _ctx;
1301         int ret;
1302         struct message *next_msg;
1303         u64 next_chunk_in_msg;
1304
1305         /* Update SHA1 message digest for the stream currently being read by the
1306          * main thread. */
1307         sha1_update(&ctx->next_sha_ctx, chunk, chunk_size);
1308
1309         /* We send chunks of data to the compressor chunks in batches which we
1310          * refer to as "messages".  @next_msg is the message that is currently
1311          * being prepared to send off.  If it is NULL, that indicates that we
1312          * need to start a new message. */
1313         next_msg = ctx->next_msg;
1314         if (!next_msg) {
1315                 /* We need to start a new message.  First check to see if there
1316                  * is a message available in the list of available messages.  If
1317                  * so, we can just take one.  If not, all the messages (there is
1318                  * a fixed number of them, proportional to the number of
1319                  * threads) have been sent off to the compressor threads, so we
1320                  * receive messages from the compressor threads containing
1321                  * compressed chunks of data.
1322                  *
1323                  * We may need to receive multiple messages before one is
1324                  * actually available to use because messages received that are
1325                  * *not* for the very next set of chunks to compress must be
1326                  * buffered until it's time to write those chunks. */
1327                 while (list_empty(&ctx->available_msgs)) {
1328                         ret = receive_compressed_chunks(ctx);
1329                         if (ret)
1330                                 return ret;
1331                 }
1332
1333                 next_msg = container_of(ctx->available_msgs.next,
1334                                         struct message, list);
1335                 list_del(&next_msg->list);
1336                 next_msg->complete = false;
1337                 next_msg->begin_chunk = ctx->next_chunk;
1338                 next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
1339                                            ctx->next_num_chunks - ctx->next_chunk);
1340                 ctx->next_msg = next_msg;
1341         }
1342
1343         /* Fill in the next chunk to compress */
1344         next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;
1345
1346         next_msg->uncompressed_chunk_sizes[next_chunk_in_msg] = chunk_size;
1347         memcpy(next_msg->uncompressed_chunks[next_chunk_in_msg],
1348                chunk, chunk_size);
1349         ctx->next_chunk++;
1350         if (++next_chunk_in_msg == next_msg->num_chunks) {
1351                 /* Send off an array of chunks to compress */
1352                 list_add_tail(&next_msg->list, &ctx->next_lte->msg_list);
1353                 shared_queue_put(ctx->res_to_compress_queue, next_msg);
1354                 ++ctx->num_outstanding_messages;
1355                 ctx->next_msg = NULL;
1356         }
1357         return 0;
1358 }
1359
1360 static int
1361 main_writer_thread_finish(void *_ctx)
1362 {
1363         struct main_writer_thread_ctx *ctx = _ctx;
1364         int ret;
1365         while (ctx->num_outstanding_messages != 0) {
1366                 ret = receive_compressed_chunks(ctx);
1367                 if (ret)
1368                         return ret;
1369         }
1370         wimlib_assert(list_empty(&ctx->outstanding_streams));
1371         return do_write_stream_list_serial(&ctx->serial_streams,
1372                                            ctx->lookup_table,
1373                                            ctx->out_fd,
1374                                            ctx->out_ctype,
1375                                            ctx->comp_ctx,
1376                                            ctx->write_resource_flags,
1377                                            ctx->progress_data);
1378 }
1379
1380 static int
1381 submit_stream_for_compression(struct wim_lookup_table_entry *lte,
1382                               struct main_writer_thread_ctx *ctx)
1383 {
1384         int ret;
1385
1386         /* Read the entire stream @lte, feeding its data chunks to the
1387          * compressor threads.  Also SHA1-sum the stream; this is required in
1388          * the case that @lte is unhashed, and a nice additional verification
1389          * when @lte is already hashed. */
1390         sha1_init(&ctx->next_sha_ctx);
1391         ctx->next_chunk = 0;
1392         ctx->next_num_chunks = wim_resource_chunks(lte);
1393         ctx->next_lte = lte;
1394         INIT_LIST_HEAD(&lte->msg_list);
1395         list_add_tail(&lte->being_compressed_list, &ctx->outstanding_streams);
1396         ret = read_resource_prefix(lte, wim_resource_size(lte),
1397                                    main_writer_thread_cb, ctx, 0);
1398         if (ret)
1399                 return ret;
1400         wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
1401         return finalize_and_check_sha1(&ctx->next_sha_ctx, lte);
1402 }
1403
1404 static int
1405 main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
1406 {
1407         struct main_writer_thread_ctx *ctx = _ctx;
1408         int ret;
1409
1410         if (wim_resource_size(lte) < 1000 ||
1411             ctx->out_ctype == WIMLIB_COMPRESSION_TYPE_NONE ||
1412             (lte->resource_location == RESOURCE_IN_WIM &&
1413              !(ctx->write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS) &&
1414              lte->wim->compression_type == ctx->out_ctype))
1415         {
1416                 /* Stream is too small or isn't being compressed.  Process it by
1417                  * the main thread when we have a chance.  We can't necessarily
1418                  * process it right here, as the main thread could be in the
1419                  * middle of writing a different stream. */
1420                 list_add_tail(&lte->write_streams_list, &ctx->serial_streams);
1421                 lte->deferred = 1;
1422                 ret = 0;
1423         } else {
1424                 ret = submit_stream_for_compression(lte, ctx);
1425         }
1426         lte->no_progress = 1;
1427         return ret;
1428 }
1429
1430 static long
1431 get_default_num_threads(void)
1432 {
1433 #ifdef __WIN32__
1434         return win32_get_number_of_processors();
1435 #else
1436         return sysconf(_SC_NPROCESSORS_ONLN);
1437 #endif
1438 }
1439
1440 /* Equivalent to write_stream_list_serial(), except this takes a @num_threads
1441  * parameter and will perform compression using that many threads.  Falls
1442  * back to write_stream_list_serial() on certain errors, such as a failure to
1443  * create the number of threads requested.
1444  *
1445  * High level description of the algorithm for writing compressed streams in
1446  * parallel:  We perform compression on chunks of size WIM_CHUNK_SIZE bytes
1447  * rather than on full files.  The currently executing thread becomes the main
1448  * thread and is entirely in charge of reading the data to compress (which may
1449  * be in any location understood by the resource code--- such as in an external
1450  * file being captured, or in another WIM file from which an image is being
1451  * exported) and actually writing the compressed data to the output file.
1452  * Additional threads are "compressor threads" and all execute the
1453  * compressor_thread_proc, where they repeatedly retrieve buffers of data from
1454  * the main thread, compress them, and hand them back to the main thread.
1455  *
1456  * Certain streams, such as streams that do not need to be compressed (e.g.
1457  * input compression type same as output compression type) or streams of very
1458  * small size are placed in a list (main_writer_thread_ctx.serial_list) and
1459  * handled entirely by the main thread at an appropriate time.
1460  *
1461  * At any given point in time, multiple streams may be having chunks compressed
1462  * concurrently.  The stream that the main thread is currently *reading* may be
1463  * later in the list that the stream that the main thread is currently
1464  * *writing*.
1465  */
1466 static int
1467 write_stream_list_parallel(struct list_head *stream_list,
1468                            struct wim_lookup_table *lookup_table,
1469                            struct filedes *out_fd,
1470                            int out_ctype,
1471                            struct wimlib_lzx_context **comp_ctx,
1472                            int write_resource_flags,
1473                            struct write_streams_progress_data *progress_data,
1474                            unsigned num_threads)
1475 {
1476         int ret;
1477         struct shared_queue res_to_compress_queue;
1478         struct shared_queue compressed_res_queue;
1479         pthread_t *compressor_threads = NULL;
1480         union wimlib_progress_info *progress = &progress_data->progress;
1481
1482         if (num_threads == 0) {
1483                 long nthreads = get_default_num_threads();
1484                 if (nthreads < 1 || nthreads > UINT_MAX) {
1485                         WARNING("Could not determine number of processors! Assuming 1");
1486                         goto out_serial;
1487                 } else if (nthreads == 1) {
1488                         goto out_serial_quiet;
1489                 } else {
1490                         num_threads = nthreads;
1491                 }
1492         }
1493
1494         DEBUG("Writing stream list of size %"PRIu64" "
1495               "(parallel version, num_threads=%u)",
1496               progress->write_streams.total_streams, num_threads);
1497
1498         progress->write_streams.num_threads = num_threads;
1499
1500         static const size_t MESSAGES_PER_THREAD = 2;
1501         size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1502
1503         DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1504
1505         ret = shared_queue_init(&res_to_compress_queue, queue_size);
1506         if (ret)
1507                 goto out_serial;
1508
1509         ret = shared_queue_init(&compressed_res_queue, queue_size);
1510         if (ret)
1511                 goto out_destroy_res_to_compress_queue;
1512
1513         struct compressor_thread_params *params;
1514
1515         params = CALLOC(num_threads, sizeof(params[0]));
1516         if (params == NULL) {
1517                 ret = WIMLIB_ERR_NOMEM;
1518                 goto out_destroy_compressed_res_queue;
1519         }
1520
1521         for (unsigned i = 0; i < num_threads; i++) {
1522                 params[i].res_to_compress_queue = &res_to_compress_queue;
1523                 params[i].compressed_res_queue = &compressed_res_queue;
1524                 params[i].out_ctype = out_ctype;
1525                 if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX) {
1526                         ret = wimlib_lzx_alloc_context(NULL, &params[i].comp_ctx);
1527                         if (ret)
1528                                 goto out_free_params;
1529                 }
1530         }
1531
1532         compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1533         if (!compressor_threads) {
1534                 ret = WIMLIB_ERR_NOMEM;
1535                 goto out_free_params;
1536         }
1537
1538         for (unsigned i = 0; i < num_threads; i++) {
1539                 DEBUG("pthread_create thread %u of %u", i + 1, num_threads);
1540                 ret = pthread_create(&compressor_threads[i], NULL,
1541                                      compressor_thread_proc, &params[i]);
1542                 if (ret != 0) {
1543                         ret = -1;
1544                         ERROR_WITH_ERRNO("Failed to create compressor "
1545                                          "thread %u of %u",
1546                                          i + 1, num_threads);
1547                         num_threads = i;
1548                         goto out_join;
1549                 }
1550         }
1551
1552         if (progress_data->progress_func) {
1553                 progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
1554                                              progress);
1555         }
1556
1557         struct main_writer_thread_ctx ctx;
1558         ctx.stream_list           = stream_list;
1559         ctx.lookup_table          = lookup_table;
1560         ctx.out_fd                = out_fd;
1561         ctx.out_ctype             = out_ctype;
1562         ctx.comp_ctx              = comp_ctx;
1563         ctx.res_to_compress_queue = &res_to_compress_queue;
1564         ctx.compressed_res_queue  = &compressed_res_queue;
1565         ctx.num_messages          = queue_size;
1566         ctx.write_resource_flags  = write_resource_flags;
1567         ctx.progress_data         = progress_data;
1568         ret = main_writer_thread_init_ctx(&ctx);
1569         if (ret)
1570                 goto out_join;
1571         ret = do_write_stream_list(stream_list, lookup_table,
1572                                    main_thread_process_next_stream,
1573                                    &ctx, progress_data);
1574         if (ret)
1575                 goto out_destroy_ctx;
1576
1577         /* The main thread has finished reading all streams that are going to be
1578          * compressed in parallel, and it now needs to wait for all remaining
1579          * chunks to be compressed so that the remaining streams can actually be
1580          * written to the output file.  Furthermore, any remaining streams that
1581          * had processing deferred to the main thread need to be handled.  These
1582          * tasks are done by the main_writer_thread_finish() function. */
1583         ret = main_writer_thread_finish(&ctx);
1584 out_destroy_ctx:
1585         main_writer_thread_destroy_ctx(&ctx);
1586 out_join:
1587         for (unsigned i = 0; i < num_threads; i++)
1588                 shared_queue_put(&res_to_compress_queue, NULL);
1589
1590         for (unsigned i = 0; i < num_threads; i++) {
1591                 if (pthread_join(compressor_threads[i], NULL)) {
1592                         WARNING_WITH_ERRNO("Failed to join compressor "
1593                                            "thread %u of %u",
1594                                            i + 1, num_threads);
1595                 }
1596         }
1597         FREE(compressor_threads);
1598 out_free_params:
1599         for (unsigned i = 0; i < num_threads; i++)
1600                 wimlib_lzx_free_context(params[i].comp_ctx);
1601         FREE(params);
1602 out_destroy_compressed_res_queue:
1603         shared_queue_destroy(&compressed_res_queue);
1604 out_destroy_res_to_compress_queue:
1605         shared_queue_destroy(&res_to_compress_queue);
1606         if (ret >= 0 && ret != WIMLIB_ERR_NOMEM)
1607                 return ret;
1608 out_serial:
1609         WARNING("Falling back to single-threaded compression");
1610 out_serial_quiet:
1611         return write_stream_list_serial(stream_list,
1612                                         lookup_table,
1613                                         out_fd,
1614                                         out_ctype,
1615                                         comp_ctx,
1616                                         write_resource_flags,
1617                                         progress_data);
1618
1619 }
1620 #endif
1621
1622 /*
1623  * Write a list of streams to a WIM (@out_fd) using the compression type
1624  * @out_ctype and up to @num_threads compressor threads.
1625  */
1626 static int
1627 write_stream_list(struct list_head *stream_list,
1628                   struct wim_lookup_table *lookup_table,
1629                   struct filedes *out_fd, int out_ctype,
1630                   struct wimlib_lzx_context **comp_ctx,
1631                   int write_flags,
1632                   unsigned num_threads, wimlib_progress_func_t progress_func)
1633 {
1634         struct wim_lookup_table_entry *lte;
1635         size_t num_streams = 0;
1636         u64 total_bytes = 0;
1637         u64 total_compression_bytes = 0;
1638         struct write_streams_progress_data progress_data;
1639         int ret;
1640         int write_resource_flags;
1641         unsigned total_parts = 0;
1642         WIMStruct *prev_wim_part = NULL;
1643
1644         if (list_empty(stream_list)) {
1645                 DEBUG("No streams to write.");
1646                 return 0;
1647         }
1648
1649         write_resource_flags = write_flags_to_resource_flags(write_flags);
1650
1651         DEBUG("Writing stream list (offset = %"PRIu64", write_resource_flags=0x%08x)",
1652               out_fd->offset, write_resource_flags);
1653
1654         sort_stream_list_by_sequential_order(stream_list,
1655                                              offsetof(struct wim_lookup_table_entry,
1656                                                       write_streams_list));
1657
1658         /* Calculate the total size of the streams to be written.  Note: this
1659          * will be the uncompressed size, as we may not know the compressed size
1660          * yet, and also this will assume that every unhashed stream will be
1661          * written (which will not necessarily be the case). */
1662         list_for_each_entry(lte, stream_list, write_streams_list) {
1663                 num_streams++;
1664                 total_bytes += wim_resource_size(lte);
1665                 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
1666                        && (wim_resource_compression_type(lte) != out_ctype ||
1667                            (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS)))
1668                 {
1669                         total_compression_bytes += wim_resource_size(lte);
1670                 }
1671                 if (lte->resource_location == RESOURCE_IN_WIM) {
1672                         if (prev_wim_part != lte->wim) {
1673                                 prev_wim_part = lte->wim;
1674                                 total_parts++;
1675                         }
1676                 }
1677         }
1678
1679         memset(&progress_data, 0, sizeof(progress_data));
1680         progress_data.progress_func = progress_func;
1681
1682         progress_data.progress.write_streams.total_bytes       = total_bytes;
1683         progress_data.progress.write_streams.total_streams     = num_streams;
1684         progress_data.progress.write_streams.completed_bytes   = 0;
1685         progress_data.progress.write_streams.completed_streams = 0;
1686         progress_data.progress.write_streams.num_threads       = num_threads;
1687         progress_data.progress.write_streams.compression_type  = out_ctype;
1688         progress_data.progress.write_streams.total_parts       = total_parts;
1689         progress_data.progress.write_streams.completed_parts   = 0;
1690
1691         progress_data.next_progress = 0;
1692         progress_data.prev_wim_part = NULL;
1693
1694 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1695         if (total_compression_bytes >= 2000000 && num_threads != 1)
1696                 ret = write_stream_list_parallel(stream_list,
1697                                                  lookup_table,
1698                                                  out_fd,
1699                                                  out_ctype,
1700                                                  comp_ctx,
1701                                                  write_resource_flags,
1702                                                  &progress_data,
1703                                                  num_threads);
1704         else
1705 #endif
1706                 ret = write_stream_list_serial(stream_list,
1707                                                lookup_table,
1708                                                out_fd,
1709                                                out_ctype,
1710                                                comp_ctx,
1711                                                write_resource_flags,
1712                                                &progress_data);
1713         if (ret == 0)
1714                 DEBUG("Successfully wrote stream list.");
1715         else
1716                 DEBUG("Failed to write stream list.");
1717         return ret;
1718 }
1719
1720 struct stream_size_table {
1721         struct hlist_head *array;
1722         size_t num_entries;
1723         size_t capacity;
1724 };
1725
1726 static int
1727 init_stream_size_table(struct stream_size_table *tab, size_t capacity)
1728 {
1729         tab->array = CALLOC(capacity, sizeof(tab->array[0]));
1730         if (!tab->array)
1731                 return WIMLIB_ERR_NOMEM;
1732         tab->num_entries = 0;
1733         tab->capacity = capacity;
1734         return 0;
1735 }
1736
1737 static void
1738 destroy_stream_size_table(struct stream_size_table *tab)
1739 {
1740         FREE(tab->array);
1741 }
1742
1743 static int
1744 stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
1745 {
1746         struct stream_size_table *tab = _tab;
1747         size_t pos;
1748         struct wim_lookup_table_entry *same_size_lte;
1749         struct hlist_node *tmp;
1750
1751         pos = hash_u64(wim_resource_size(lte)) % tab->capacity;
1752         lte->unique_size = 1;
1753         hlist_for_each_entry(same_size_lte, tmp, &tab->array[pos], hash_list_2) {
1754                 if (wim_resource_size(same_size_lte) == wim_resource_size(lte)) {
1755                         lte->unique_size = 0;
1756                         same_size_lte->unique_size = 0;
1757                         break;
1758                 }
1759         }
1760
1761         hlist_add_head(&lte->hash_list_2, &tab->array[pos]);
1762         tab->num_entries++;
1763         return 0;
1764 }
1765
1766 struct find_streams_ctx {
1767         WIMStruct *wim;
1768         int write_flags;
1769         struct list_head stream_list;
1770         struct stream_size_table stream_size_tab;
1771 };
1772
1773 static void
1774 lte_reference_for_logical_write(struct wim_lookup_table_entry *lte,
1775                                 struct find_streams_ctx *ctx,
1776                                 unsigned nref)
1777 {
1778         if (lte->out_refcnt == 0) {
1779                 stream_size_table_insert(lte, &ctx->stream_size_tab);
1780                 list_add_tail(&lte->write_streams_list, &ctx->stream_list);
1781         }
1782         lte->out_refcnt += nref;
1783 }
1784
1785 static int
1786 do_lte_full_reference_for_logical_write(struct wim_lookup_table_entry *lte,
1787                                         void *_ctx)
1788 {
1789         struct find_streams_ctx *ctx = _ctx;
1790         lte->out_refcnt = 0;
1791         lte_reference_for_logical_write(lte, ctx,
1792                                         (lte->refcnt ? lte->refcnt : 1));
1793         return 0;
1794 }
1795
1796 static int
1797 inode_find_streams_to_write(struct wim_inode *inode,
1798                             struct wim_lookup_table *table,
1799                             struct find_streams_ctx *ctx)
1800 {
1801         struct wim_lookup_table_entry *lte;
1802         unsigned i;
1803
1804         for (i = 0; i <= inode->i_num_ads; i++) {
1805                 lte = inode_stream_lte(inode, i, table);
1806                 if (lte)
1807                         lte_reference_for_logical_write(lte, ctx, inode->i_nlink);
1808                 else if (!is_zero_hash(inode_stream_hash(inode, i)))
1809                         return WIMLIB_ERR_RESOURCE_NOT_FOUND;
1810         }
1811         return 0;
1812 }
1813
1814 static int
1815 image_find_streams_to_write(WIMStruct *wim)
1816 {
1817         struct find_streams_ctx *ctx;
1818         struct wim_image_metadata *imd;
1819         struct wim_inode *inode;
1820         struct wim_lookup_table_entry *lte;
1821         int ret;
1822
1823         ctx = wim->private;
1824         imd = wim_get_current_image_metadata(wim);
1825
1826         image_for_each_unhashed_stream(lte, imd)
1827                 lte->out_refcnt = 0;
1828
1829         /* Go through this image's inodes to find any streams that have not been
1830          * found yet. */
1831         image_for_each_inode(inode, imd) {
1832                 ret = inode_find_streams_to_write(inode, wim->lookup_table, ctx);
1833                 if (ret)
1834                         return ret;
1835         }
1836         return 0;
1837 }
1838
1839 /*
1840  * Build a list of streams (via `struct wim_lookup_table_entry's) included in
1841  * the "logical write" of the WIM, meaning all streams that are referenced at
1842  * least once by dentries in the the image(s) being written.  'out_refcnt' on
1843  * each stream being included in the logical write is set to the number of
1844  * references from dentries in the image(s).  Furthermore, 'unique_size' on each
1845  * stream being included in the logical write is set to indicate whether that
1846  * stream has a unique size relative to the streams being included in the
1847  * logical write.  Still furthermore, 'part_number' on each stream being
1848  * included in the logical write is set to the part number given in the
1849  * in-memory header of @p wim.
1850  *
1851  * This is considered a "logical write" because it does not take into account
1852  * filtering out streams already present in the WIM (in the case of an in place
1853  * overwrite) or present in other WIMs (in case of creating delta WIM).
1854  */
1855 static int
1856 prepare_logical_stream_list(WIMStruct *wim, int image, bool streams_ok,
1857                             struct find_streams_ctx *ctx)
1858 {
1859         int ret;
1860         struct wim_lookup_table_entry *lte;
1861
1862         if (streams_ok && (image == WIMLIB_ALL_IMAGES ||
1863                            (image == 1 && wim->hdr.image_count == 1)))
1864         {
1865                 /* Fast case:  Assume that all streams are being written and
1866                  * that the reference counts are correct.  */
1867                 struct wim_lookup_table_entry *lte;
1868                 struct wim_image_metadata *imd;
1869                 unsigned i;
1870
1871                 for_lookup_table_entry(wim->lookup_table,
1872                                        do_lte_full_reference_for_logical_write, ctx);
1873                 for (i = 0; i < wim->hdr.image_count; i++) {
1874                         imd = wim->image_metadata[i];
1875                         image_for_each_unhashed_stream(lte, imd)
1876                                 do_lte_full_reference_for_logical_write(lte, ctx);
1877                 }
1878         } else {
1879                 /* Slow case:  Walk through the images being written and
1880                  * determine the streams referenced.  */
1881                 for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
1882                 wim->private = ctx;
1883                 ret = for_image(wim, image, image_find_streams_to_write);
1884                 if (ret)
1885                         return ret;
1886         }
1887
1888         list_for_each_entry(lte, &ctx->stream_list, write_streams_list)
1889                 lte->part_number = wim->hdr.part_number;
1890         return 0;
1891 }
1892
1893 static int
1894 process_filtered_stream(struct wim_lookup_table_entry *lte, void *_ctx)
1895 {
1896         struct find_streams_ctx *ctx = _ctx;
1897         u16 filtered = 0;
1898
1899         /* Calculate and set lte->filtered.  */
1900         if (lte->resource_location == RESOURCE_IN_WIM) {
1901                 if (lte->wim == ctx->wim &&
1902                     (ctx->write_flags & WIMLIB_WRITE_FLAG_OVERWRITE))
1903                         filtered |= FILTERED_SAME_WIM;
1904                 if (lte->wim != ctx->wim &&
1905                     (ctx->write_flags & WIMLIB_WRITE_FLAG_SKIP_EXTERNAL_WIMS))
1906                         filtered |= FILTERED_EXTERNAL_WIM;
1907         }
1908         lte->filtered = filtered;
1909
1910         /* Filtered streams get inserted into the stream size table too, unless
1911          * they already were.  This is because streams that are checksummed
1912          * on-the-fly during the write should not be written if they are
1913          * duplicates of filtered stream.  */
1914         if (lte->filtered && lte->out_refcnt == 0)
1915                 stream_size_table_insert(lte, &ctx->stream_size_tab);
1916         return 0;
1917 }
1918
1919 static int
1920 mark_stream_not_filtered(struct wim_lookup_table_entry *lte, void *_ignore)
1921 {
1922         lte->filtered = 0;
1923         return 0;
1924 }
1925
1926 /* Given the list of streams to include in a logical write of a WIM, handle
1927  * filtering out streams already present in the WIM or already present in
1928  * external WIMs, depending on the write flags provided.  */
1929 static void
1930 handle_stream_filtering(struct find_streams_ctx *ctx)
1931 {
1932         struct wim_lookup_table_entry *lte, *tmp;
1933
1934         if (!(ctx->write_flags & (WIMLIB_WRITE_FLAG_OVERWRITE |
1935                                   WIMLIB_WRITE_FLAG_SKIP_EXTERNAL_WIMS)))
1936         {
1937                 for_lookup_table_entry(ctx->wim->lookup_table,
1938                                        mark_stream_not_filtered, ctx);
1939                 return;
1940         }
1941
1942         for_lookup_table_entry(ctx->wim->lookup_table,
1943                                process_filtered_stream, ctx);
1944
1945         /* Streams in logical write list that were filtered can be removed.  */
1946         list_for_each_entry_safe(lte, tmp, &ctx->stream_list,
1947                                  write_streams_list)
1948                 if (lte->filtered)
1949                         list_del(&lte->write_streams_list);
1950 }
1951
1952 /* Prepares list of streams to write for the specified WIM image(s).  This wraps
1953  * around prepare_logical_stream_list() to handle filtering out streams already
1954  * present in the WIM or already present in external WIMs, depending on the
1955  * write flags provided.
1956  *
1957  * Note: some additional data is stored in each `struct wim_lookup_table_entry':
1958  *
1959  * - 'out_refcnt' is set to the number of references found for the logical write.
1960  *    This will be nonzero on all streams in the list returned by this function,
1961  *    but will also be nonzero on streams not in the list that were included in
1962  *    the logical write list, but filtered out from the returned list.
1963  * - 'filtered' is set to nonzero if the stream was filtered.  Filtered streams
1964  *   are not included in the list of streams returned by this function.
1965  * - 'unique_size' is set if the stream has a unique size among all streams in
1966  *   the logical write plus any filtered streams in the entire WIM that could
1967  *   potentially turn out to have the same checksum as a yet-to-be-checksummed
1968  *   stream being written.
1969  */
1970 static int
1971 prepare_stream_list(WIMStruct *wim, int image, int write_flags,
1972                     struct list_head *stream_list)
1973 {
1974         int ret;
1975         bool streams_ok;
1976         struct find_streams_ctx ctx;
1977
1978         INIT_LIST_HEAD(&ctx.stream_list);
1979         ret = init_stream_size_table(&ctx.stream_size_tab,
1980                                      wim->lookup_table->capacity);
1981         if (ret)
1982                 return ret;
1983         ctx.write_flags = write_flags;
1984         ctx.wim = wim;
1985
1986         streams_ok = ((write_flags & WIMLIB_WRITE_FLAG_STREAMS_OK) != 0);
1987
1988         ret = prepare_logical_stream_list(wim, image, streams_ok, &ctx);
1989         if (ret)
1990                 goto out_destroy_table;
1991
1992         handle_stream_filtering(&ctx);
1993         list_transfer(&ctx.stream_list, stream_list);
1994         ret = 0;
1995 out_destroy_table:
1996         destroy_stream_size_table(&ctx.stream_size_tab);
1997         return ret;
1998 }
1999
2000 static int
2001 write_wim_streams(WIMStruct *wim, int image, int write_flags,
2002                   unsigned num_threads,
2003                   wimlib_progress_func_t progress_func,
2004                   struct list_head *stream_list_override)
2005 {
2006         int ret;
2007         struct list_head _stream_list;
2008         struct list_head *stream_list;
2009         struct wim_lookup_table_entry *lte;
2010
2011         if (stream_list_override == NULL) {
2012                 /* Normal case: prepare stream list from image(s) being written.
2013                  */
2014                 stream_list = &_stream_list;
2015                 ret = prepare_stream_list(wim, image, write_flags, stream_list);
2016                 if (ret)
2017                         return ret;
2018         } else {
2019                 /* Currently only as a result of wimlib_split() being called:
2020                  * use stream list already explicitly provided.  Use existing
2021                  * reference counts.  */
2022                 stream_list = stream_list_override;
2023                 list_for_each_entry(lte, stream_list, write_streams_list) {
2024                         lte->out_refcnt = (lte->refcnt ? lte->refcnt : 1);
2025                         lte->part_number = wim->hdr.part_number;
2026                 }
2027         }
2028
2029         return write_stream_list(stream_list,
2030                                  wim->lookup_table,
2031                                  &wim->out_fd,
2032                                  wim->out_compression_type,
2033                                  &wim->lzx_context,
2034                                  write_flags,
2035                                  num_threads,
2036                                  progress_func);
2037 }
2038
2039 static int
2040 write_wim_metadata_resources(WIMStruct *wim, int image, int write_flags,
2041                              wimlib_progress_func_t progress_func)
2042 {
2043         int ret;
2044         int start_image;
2045         int end_image;
2046         int write_resource_flags;
2047
2048         if (write_flags & WIMLIB_WRITE_FLAG_NO_METADATA) {
2049                 DEBUG("Not writing any metadata resources.");
2050                 return 0;
2051         }
2052
2053         write_resource_flags = write_flags_to_resource_flags(write_flags);
2054
2055         DEBUG("Writing metadata resources (offset=%"PRIu64")",
2056               wim->out_fd.offset);
2057
2058         if (progress_func)
2059                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
2060
2061         if (image == WIMLIB_ALL_IMAGES) {
2062                 start_image = 1;
2063                 end_image = wim->hdr.image_count;
2064         } else {
2065                 start_image = image;
2066                 end_image = image;
2067         }
2068
2069         for (int i = start_image; i <= end_image; i++) {
2070                 struct wim_image_metadata *imd;
2071
2072                 imd = wim->image_metadata[i - 1];
2073                 /* Build a new metadata resource only if image was modified from
2074                  * the original (or was newly added).  Otherwise just copy the
2075                  * existing one.  */
2076                 if (imd->modified) {
2077                         DEBUG("Image %u was modified; building and writing new "
2078                               "metadata resource", i);
2079                         ret = write_metadata_resource(wim, i,
2080                                                       write_resource_flags);
2081                 } else if (write_flags & WIMLIB_WRITE_FLAG_OVERWRITE) {
2082                         DEBUG("Image %u was not modified; re-using existing "
2083                               "metadata resource.", i);
2084                         copy_resource_entry(&imd->metadata_lte->output_resource_entry,
2085                                             &imd->metadata_lte->resource_entry);
2086                         ret = 0;
2087                 } else {
2088                         DEBUG("Image %u was not modified; copying existing "
2089                               "metadata resource.", i);
2090                         ret = write_wim_resource(imd->metadata_lte,
2091                                                  &wim->out_fd,
2092                                                  wim->out_compression_type,
2093                                                  &imd->metadata_lte->output_resource_entry,
2094                                                  write_resource_flags,
2095                                                  &wim->lzx_context);
2096                 }
2097                 if (ret)
2098                         return ret;
2099         }
2100         if (progress_func)
2101                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
2102         return 0;
2103 }
2104
2105 static int
2106 open_wim_writable(WIMStruct *wim, const tchar *path, int open_flags)
2107 {
2108         int raw_fd;
2109         DEBUG("Opening \"%"TS"\" for writing.", path);
2110
2111         raw_fd = topen(path, open_flags | O_BINARY, 0644);
2112         if (raw_fd < 0) {
2113                 ERROR_WITH_ERRNO("Failed to open \"%"TS"\" for writing", path);
2114                 return WIMLIB_ERR_OPEN;
2115         }
2116         filedes_init(&wim->out_fd, raw_fd);
2117         return 0;
2118 }
2119
2120 static int
2121 close_wim_writable(WIMStruct *wim, int write_flags)
2122 {
2123         int ret = 0;
2124
2125         if (!(write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)) {
2126                 DEBUG("Closing WIM file.");
2127                 if (filedes_valid(&wim->out_fd))
2128                         if (filedes_close(&wim->out_fd))
2129                                 ret = WIMLIB_ERR_WRITE;
2130         }
2131         filedes_invalidate(&wim->out_fd);
2132         return ret;
2133 }
2134
2135 /*
2136  * finish_write():
2137  *
2138  * Finish writing a WIM file: write the lookup table, xml data, and integrity
2139  * table, then overwrite the WIM header.  By default, closes the WIM file
2140  * descriptor (@wim->out_fd) if successful.
2141  *
2142  * write_flags is a bitwise OR of the following:
2143  *
2144  *      (public) WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
2145  *              Include an integrity table.
2146  *
2147  *      (public) WIMLIB_WRITE_FLAG_FSYNC:
2148  *              fsync() the output file before closing it.
2149  *
2150  *      (public) WIMLIB_WRITE_FLAG_PIPABLE:
2151  *              Writing a pipable WIM, possibly to a pipe; include pipable WIM
2152  *              stream headers before the lookup table and XML data, and also
2153  *              write the WIM header at the end instead of seeking to the
2154  *              beginning.  Can't be combined with
2155  *              WIMLIB_WRITE_FLAG_CHECK_INTEGRITY.
2156  *
2157  *      (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
2158  *              Don't write the lookup table.
2159  *
2160  *      (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
2161  *              When (if) writing the integrity table, re-use entries from the
2162  *              existing integrity table, if possible.
2163  *
2164  *      (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
2165  *              After writing the XML data but before writing the integrity
2166  *              table, write a temporary WIM header and flush the stream so that
2167  *              the WIM is less likely to become corrupted upon abrupt program
2168  *              termination.
2169  *      (private) WIMLIB_WRITE_FLAG_HEADER_AT_END:
2170  *              Instead of overwriting the WIM header at the beginning of the
2171  *              file, simply append it to the end of the file.  (Used when
2172  *              writing to pipe.)
2173  *      (private) WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR:
2174  *              Do not close the file descriptor @wim->out_fd on either success
2175  *              on failure.
2176  *      (private) WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES:
2177  *              Use the existing <TOTALBYTES> stored in the in-memory XML
2178  *              information, rather than setting it to the offset of the XML
2179  *              data being written.
2180  */
2181 static int
2182 finish_write(WIMStruct *wim, int image, int write_flags,
2183              wimlib_progress_func_t progress_func,
2184              struct list_head *stream_list_override)
2185 {
2186         int ret;
2187         off_t hdr_offset;
2188         int write_resource_flags;
2189         off_t old_lookup_table_end;
2190         off_t new_lookup_table_end;
2191         u64 xml_totalbytes;
2192
2193         DEBUG("image=%d, write_flags=%08x", image, write_flags);
2194
2195         write_resource_flags = write_flags_to_resource_flags(write_flags);
2196
2197         /* In the WIM header, there is room for the resource entry for a
2198          * metadata resource labeled as the "boot metadata".  This entry should
2199          * be zeroed out if there is no bootable image (boot_idx 0).  Otherwise,
2200          * it should be a copy of the resource entry for the image that is
2201          * marked as bootable.  This is not well documented...  */
2202         if (wim->hdr.boot_idx == 0) {
2203                 zero_resource_entry(&wim->hdr.boot_metadata_res_entry);
2204         } else {
2205                 copy_resource_entry(&wim->hdr.boot_metadata_res_entry,
2206                             &wim->image_metadata[wim->hdr.boot_idx- 1
2207                                         ]->metadata_lte->output_resource_entry);
2208         }
2209
2210         /* Write lookup table.  (Save old position first.)  */
2211         old_lookup_table_end = wim->hdr.lookup_table_res_entry.offset +
2212                                wim->hdr.lookup_table_res_entry.size;
2213         if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
2214                 ret = write_wim_lookup_table(wim, image, write_flags,
2215                                              &wim->hdr.lookup_table_res_entry,
2216                                              stream_list_override);
2217                 if (ret)
2218                         return ret;
2219         }
2220
2221         /* Write XML data.  */
2222         xml_totalbytes = wim->out_fd.offset;
2223         if (write_flags & WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES)
2224                 xml_totalbytes = WIM_TOTALBYTES_USE_EXISTING;
2225         ret = write_wim_xml_data(wim, image, xml_totalbytes,
2226                                  &wim->hdr.xml_res_entry,
2227                                  write_resource_flags);
2228         if (ret)
2229                 return ret;
2230
2231         /* Write integrity table (optional).  */
2232         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
2233                 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
2234                         struct wim_header checkpoint_hdr;
2235                         memcpy(&checkpoint_hdr, &wim->hdr, sizeof(struct wim_header));
2236                         zero_resource_entry(&checkpoint_hdr.integrity);
2237                         checkpoint_hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2238                         ret = write_wim_header_at_offset(&checkpoint_hdr,
2239                                                          &wim->out_fd, 0);
2240                         if (ret)
2241                                 return ret;
2242                 }
2243
2244                 if (!(write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE))
2245                         old_lookup_table_end = 0;
2246
2247                 new_lookup_table_end = wim->hdr.lookup_table_res_entry.offset +
2248                                        wim->hdr.lookup_table_res_entry.size;
2249
2250                 ret = write_integrity_table(wim,
2251                                             new_lookup_table_end,
2252                                             old_lookup_table_end,
2253                                             progress_func);
2254                 if (ret)
2255                         return ret;
2256         } else {
2257                 /* No integrity table.  */
2258                 zero_resource_entry(&wim->hdr.integrity);
2259         }
2260
2261         /* Now that all information in the WIM header has been determined, the
2262          * preliminary header written earlier can be overwritten, the header of
2263          * the existing WIM file can be overwritten, or the final header can be
2264          * written to the end of the pipable WIM.  */
2265         wim->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2266         hdr_offset = 0;
2267         if (write_flags & WIMLIB_WRITE_FLAG_HEADER_AT_END)
2268                 hdr_offset = wim->out_fd.offset;
2269         DEBUG("Writing new header @ %"PRIu64".", hdr_offset);
2270         ret = write_wim_header_at_offset(&wim->hdr, &wim->out_fd, hdr_offset);
2271         if (ret)
2272                 return ret;
2273
2274         /* Possibly sync file data to disk before closing.  On POSIX systems, it
2275          * is necessary to do this before using rename() to overwrite an
2276          * existing file with a new file.  Otherwise, data loss would occur if
2277          * the system is abruptly terminated when the metadata for the rename
2278          * operation has been written to disk, but the new file data has not.
2279          */
2280         if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
2281                 DEBUG("Syncing WIM file.");
2282                 if (fsync(wim->out_fd.fd)) {
2283                         ERROR_WITH_ERRNO("Error syncing data to WIM file");
2284                         return WIMLIB_ERR_WRITE;
2285                 }
2286         }
2287
2288         if (close_wim_writable(wim, write_flags)) {
2289                 ERROR_WITH_ERRNO("Failed to close the output WIM file");
2290                 return WIMLIB_ERR_WRITE;
2291         }
2292
2293         return 0;
2294 }
2295
2296 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
2297 int
2298 lock_wim(WIMStruct *wim, int fd)
2299 {
2300         int ret = 0;
2301         if (fd != -1 && !wim->wim_locked) {
2302                 ret = flock(fd, LOCK_EX | LOCK_NB);
2303                 if (ret != 0) {
2304                         if (errno == EWOULDBLOCK) {
2305                                 ERROR("`%"TS"' is already being modified or has been "
2306                                       "mounted read-write\n"
2307                                       "        by another process!", wim->filename);
2308                                 ret = WIMLIB_ERR_ALREADY_LOCKED;
2309                         } else {
2310                                 WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
2311                                                    wim->filename);
2312                                 ret = 0;
2313                         }
2314                 } else {
2315                         wim->wim_locked = 1;
2316                 }
2317         }
2318         return ret;
2319 }
2320 #endif
2321
2322 /*
2323  * write_pipable_wim():
2324  *
2325  * Perform the intermediate stages of creating a "pipable" WIM (i.e. a WIM
2326  * capable of being applied from a pipe).
2327  *
2328  * Pipable WIMs are a wimlib-specific modification of the WIM format such that
2329  * images can be applied from them sequentially when the file data is sent over
2330  * a pipe.  In addition, a pipable WIM can be written sequentially to a pipe.
2331  * The modifications made to the WIM format for pipable WIMs are:
2332  *
2333  * - Magic characters in header are "WLPWM\0\0\0" (wimlib pipable WIM) instead
2334  *   of "MSWIM\0\0\0".  This lets wimlib know that the WIM is pipable and also
2335  *   stops other software from trying to read the file as a normal WIM.
2336  *
2337  * - The header at the beginning of the file does not contain all the normal
2338  *   information; in particular it will have all 0's for the lookup table and
2339  *   XML data resource entries.  This is because this information cannot be
2340  *   determined until the lookup table and XML data have been written.
2341  *   Consequently, wimlib will write the full header at the very end of the
2342  *   file.  The header at the end, however, is only used when reading the WIM
2343  *   from a seekable file (not a pipe).
2344  *
2345  * - An extra copy of the XML data is placed directly after the header.  This
2346  *   allows image names and sizes to be determined at an appropriate time when
2347  *   reading the WIM from a pipe.  This copy of the XML data is ignored if the
2348  *   WIM is read from a seekable file (not a pipe).
2349  *
2350  * - The format of resources, or streams, has been modified to allow them to be
2351  *   used before the "lookup table" has been read.  Each stream is prefixed with
2352  *   a `struct pwm_stream_hdr' that is basically an abbreviated form of `struct
2353  *   wim_lookup_table_entry_disk' that only contains the SHA1 message digest,
2354  *   uncompressed stream size, and flags that indicate whether the stream is
2355  *   compressed.  The data of uncompressed streams then follows literally, while
2356  *   the data of compressed streams follows in a modified format.  Compressed
2357  *   streams do not begin with a chunk table, since the chunk table cannot be
2358  *   written until all chunks have been compressed.  Instead, each compressed
2359  *   chunk is prefixed by a `struct pwm_chunk_hdr' that gives its size.
2360  *   Furthermore, the chunk table is written at the end of the resource instead
2361  *   of the start.  Note: chunk offsets are given in the chunk table as if the
2362  *   `struct pwm_chunk_hdr's were not present; also, the chunk table is only
2363  *   used if the WIM is being read from a seekable file (not a pipe).
2364  *
2365  * - Metadata resources always come before other file resources (streams).
2366  *   (This does not by itself constitute an incompatibility with normal WIMs,
2367  *   since this is valid in normal WIMs.)
2368  *
2369  * - At least up to the end of the file resources, all components must be packed
2370  *   as tightly as possible; there cannot be any "holes" in the WIM.  (This does
2371  *   not by itself consititute an incompatibility with normal WIMs, since this
2372  *   is valid in normal WIMs.)
2373  *
2374  * Note: the lookup table, XML data, and header at the end are not used when
2375  * applying from a pipe.  They exist to support functionality such as image
2376  * application and export when the WIM is *not* read from a pipe.
2377  *
2378  *   Layout of pipable WIM:
2379  *
2380  * ---------+----------+--------------------+----------------+--------------+-----------+--------+
2381  * | Header | XML data | Metadata resources | File resources | Lookup table | XML data  | Header |
2382  * ---------+----------+--------------------+----------------+--------------+-----------+--------+
2383  *
2384  *   Layout of normal WIM:
2385  *
2386  * +--------+-----------------------------+-------------------------+
2387  * | Header | File and metadata resources | Lookup table | XML data |
2388  * +--------+-----------------------------+-------------------------+
2389  *
2390  * An optional integrity table can follow the final XML data in both normal and
2391  * pipable WIMs.  However, due to implementation details, wimlib currently can
2392  * only include an integrity table in a pipable WIM when writing it to a
2393  * seekable file (not a pipe).
2394  *
2395  * Do note that since pipable WIMs are not supported by Microsoft's software,
2396  * wimlib does not create them unless explicitly requested (with
2397  * WIMLIB_WRITE_FLAG_PIPABLE) and as stated above they use different magic
2398  * characters to identify the file.
2399  */
2400 static int
2401 write_pipable_wim(WIMStruct *wim, int image, int write_flags,
2402                   unsigned num_threads, wimlib_progress_func_t progress_func,
2403                   struct list_head *stream_list_override)
2404 {
2405         int ret;
2406         struct resource_entry xml_res_entry;
2407
2408         WARNING("Creating a pipable WIM, which will "
2409                 "be incompatible\n"
2410                 "          with Microsoft's software (wimgapi/imagex/Dism).");
2411
2412         /* At this point, the header at the beginning of the file has already
2413          * been written.  */
2414
2415         /* For efficiency, when wimlib adds an image to the WIM with
2416          * wimlib_add_image(), the SHA1 message digests of files is not
2417          * calculated; instead, they are calculated while the files are being
2418          * written.  However, this does not work when writing a pipable WIM,
2419          * since when writing a stream to a pipable WIM, its SHA1 message digest
2420          * needs to be known before the stream data is written.  Therefore,
2421          * before getting much farther, we need to pre-calculate the SHA1
2422          * message digests of all streams that will be written.  */
2423         ret = wim_checksum_unhashed_streams(wim);
2424         if (ret)
2425                 return ret;
2426
2427         /* Write extra copy of the XML data.  */
2428         ret = write_wim_xml_data(wim, image, WIM_TOTALBYTES_OMIT,
2429                                  &xml_res_entry,
2430                                  WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE);
2431         if (ret)
2432                 return ret;
2433
2434         /* Write metadata resources for the image(s) being included in the
2435          * output WIM.  */
2436         ret = write_wim_metadata_resources(wim, image, write_flags,
2437                                            progress_func);
2438         if (ret)
2439                 return ret;
2440
2441         /* Write streams needed for the image(s) being included in the output
2442          * WIM, or streams needed for the split WIM part.  */
2443         return write_wim_streams(wim, image, write_flags, num_threads,
2444                                  progress_func, stream_list_override);
2445
2446         /* The lookup table, XML data, and header at end are handled by
2447          * finish_write().  */
2448 }
2449
2450 /* API function documented in wimlib.h  */
2451 WIMLIBAPI int
2452 wimlib_set_output_compression_type(WIMStruct *wim, int ctype)
2453 {
2454         switch (ctype) {
2455         case WIMLIB_COMPRESSION_TYPE_INVALID:
2456                 break;
2457         case WIMLIB_COMPRESSION_TYPE_NONE:
2458         case WIMLIB_COMPRESSION_TYPE_LZX:
2459         case WIMLIB_COMPRESSION_TYPE_XPRESS:
2460                 wim->out_compression_type = ctype;
2461                 return 0;
2462         }
2463         return WIMLIB_ERR_INVALID_PARAM;
2464 }
2465
2466 /* Write a standalone WIM or split WIM (SWM) part to a new file or to a file
2467  * descriptor.  */
2468 int
2469 write_wim_part(WIMStruct *wim,
2470                const void *path_or_fd,
2471                int image,
2472                int write_flags,
2473                unsigned num_threads,
2474                wimlib_progress_func_t progress_func,
2475                unsigned part_number,
2476                unsigned total_parts,
2477                struct list_head *stream_list_override,
2478                const u8 *guid)
2479 {
2480         int ret;
2481         struct wim_header hdr_save;
2482         struct list_head lt_stream_list_override;
2483
2484         if (total_parts == 1)
2485                 DEBUG("Writing standalone WIM.");
2486         else
2487                 DEBUG("Writing split WIM part %u/%u", part_number, total_parts);
2488         if (image == WIMLIB_ALL_IMAGES)
2489                 DEBUG("Including all images.");
2490         else
2491                 DEBUG("Including image %d only.", image);
2492         if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
2493                 DEBUG("File descriptor: %d", *(const int*)path_or_fd);
2494         else
2495                 DEBUG("Path: \"%"TS"\"", (const tchar*)path_or_fd);
2496         DEBUG("Write flags: 0x%08x", write_flags);
2497         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
2498                 DEBUG("\tCHECK_INTEGRITY");
2499         if (write_flags & WIMLIB_WRITE_FLAG_REBUILD)
2500                 DEBUG("\tREBUILD");
2501         if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
2502                 DEBUG("\tRECOMPRESS");
2503         if (write_flags & WIMLIB_WRITE_FLAG_FSYNC)
2504                 DEBUG("\tFSYNC");
2505         if (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE)
2506                 DEBUG("\tFSYNC");
2507         if (write_flags & WIMLIB_WRITE_FLAG_IGNORE_READONLY_FLAG)
2508                 DEBUG("\tIGNORE_READONLY_FLAG");
2509         if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
2510                 DEBUG("\tPIPABLE");
2511         if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
2512                 DEBUG("\tFILE_DESCRIPTOR");
2513         if (write_flags & WIMLIB_WRITE_FLAG_NO_METADATA)
2514                 DEBUG("\tNO_METADATA");
2515         if (write_flags & WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES)
2516                 DEBUG("\tUSE_EXISTING_TOTALBYTES");
2517         if (num_threads == 0)
2518                 DEBUG("Number of threads: autodetect");
2519         else
2520                 DEBUG("Number of threads: %u", num_threads);
2521         DEBUG("Progress function: %s", (progress_func ? "yes" : "no"));
2522         DEBUG("Stream list:       %s", (stream_list_override ? "specified" : "autodetect"));
2523         DEBUG("GUID:              %s", ((guid || wim->guid_set_explicitly) ?
2524                                         "specified" : "generate new"));
2525
2526         /* Internally, this is always called with a valid part number and total
2527          * parts.  */
2528         wimlib_assert(total_parts >= 1);
2529         wimlib_assert(part_number >= 1 && part_number <= total_parts);
2530
2531         /* A valid image (or all images) must be specified.  */
2532         if (image != WIMLIB_ALL_IMAGES &&
2533              (image < 1 || image > wim->hdr.image_count))
2534                 return WIMLIB_ERR_INVALID_IMAGE;
2535
2536         /* If we need to write metadata resources, make sure the ::WIMStruct has
2537          * the needed information attached (e.g. is not a resource-only WIM,
2538          * such as a non-first part of a split WIM).  */
2539         if (!wim_has_metadata(wim) &&
2540             !(write_flags & WIMLIB_WRITE_FLAG_NO_METADATA))
2541                 return WIMLIB_ERR_METADATA_NOT_FOUND;
2542
2543         /* Check for contradictory flags.  */
2544         if ((write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2545                             WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY))
2546                                 == (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2547                                     WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY))
2548                 return WIMLIB_ERR_INVALID_PARAM;
2549
2550         if ((write_flags & (WIMLIB_WRITE_FLAG_PIPABLE |
2551                             WIMLIB_WRITE_FLAG_NOT_PIPABLE))
2552                                 == (WIMLIB_WRITE_FLAG_PIPABLE |
2553                                     WIMLIB_WRITE_FLAG_NOT_PIPABLE))
2554                 return WIMLIB_ERR_INVALID_PARAM;
2555
2556         /* Save previous header, then start initializing the new one.  */
2557         memcpy(&hdr_save, &wim->hdr, sizeof(struct wim_header));
2558
2559         /* Set default integrity and pipable flags.  */
2560         if (!(write_flags & (WIMLIB_WRITE_FLAG_PIPABLE |
2561                              WIMLIB_WRITE_FLAG_NOT_PIPABLE)))
2562                 if (wim_is_pipable(wim))
2563                         write_flags |= WIMLIB_WRITE_FLAG_PIPABLE;
2564
2565         if (!(write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2566                              WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY)))
2567                 if (wim_has_integrity_table(wim))
2568                         write_flags |= WIMLIB_WRITE_FLAG_CHECK_INTEGRITY;
2569
2570         /* Set appropriate magic number.  */
2571         if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
2572                 wim->hdr.magic = PWM_MAGIC;
2573         else
2574                 wim->hdr.magic = WIM_MAGIC;
2575
2576         /* Clear header flags that will be set automatically.  */
2577         wim->hdr.flags &= ~(WIM_HDR_FLAG_METADATA_ONLY          |
2578                             WIM_HDR_FLAG_RESOURCE_ONLY          |
2579                             WIM_HDR_FLAG_SPANNED                |
2580                             WIM_HDR_FLAG_WRITE_IN_PROGRESS);
2581
2582         /* Set SPANNED header flag if writing part of a split WIM.  */
2583         if (total_parts != 1)
2584                 wim->hdr.flags |= WIM_HDR_FLAG_SPANNED;
2585
2586         /* Set part number and total parts of split WIM.  This will be 1 and 1
2587          * if the WIM is standalone.  */
2588         wim->hdr.part_number = part_number;
2589         wim->hdr.total_parts = total_parts;
2590
2591         /* Set compression type if different.  */
2592         if (wim->compression_type != wim->out_compression_type)
2593                 wim->hdr.flags = get_wim_hdr_cflags(wim->out_compression_type);
2594
2595         /* Use GUID if specified; otherwise generate a new one.  */
2596         if (guid)
2597                 memcpy(wim->hdr.guid, guid, WIMLIB_GUID_LEN);
2598         else if (!wim->guid_set_explicitly)
2599                 randomize_byte_array(wim->hdr.guid, WIMLIB_GUID_LEN);
2600
2601         /* Clear references to resources that have not been written yet.  */
2602         zero_resource_entry(&wim->hdr.lookup_table_res_entry);
2603         zero_resource_entry(&wim->hdr.xml_res_entry);
2604         zero_resource_entry(&wim->hdr.boot_metadata_res_entry);
2605         zero_resource_entry(&wim->hdr.integrity);
2606
2607         /* Set image count and boot index correctly for single image writes.  */
2608         if (image != WIMLIB_ALL_IMAGES) {
2609                 wim->hdr.image_count = 1;
2610                 if (wim->hdr.boot_idx == image)
2611                         wim->hdr.boot_idx = 1;
2612                 else
2613                         wim->hdr.boot_idx = 0;
2614         }
2615
2616         /* Split WIMs can't be bootable.  */
2617         if (total_parts != 1)
2618                 wim->hdr.boot_idx = 0;
2619
2620         /* Initialize output file descriptor.  */
2621         if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR) {
2622                 /* File descriptor was explicitly provided.  Return error if
2623                  * file descriptor is not seekable, unless writing a pipable WIM
2624                  * was requested.  */
2625                 wim->out_fd.fd = *(const int*)path_or_fd;
2626                 wim->out_fd.offset = 0;
2627                 if (!filedes_is_seekable(&wim->out_fd)) {
2628                         ret = WIMLIB_ERR_INVALID_PARAM;
2629                         if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE))
2630                                 goto out_restore_hdr;
2631                         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
2632                                 ERROR("Can't include integrity check when "
2633                                       "writing pipable WIM to pipe!");
2634                                 goto out_restore_hdr;
2635                         }
2636                 }
2637
2638         } else {
2639                 /* Filename of WIM to write was provided; open file descriptor
2640                  * to it.  */
2641                 ret = open_wim_writable(wim, (const tchar*)path_or_fd,
2642                                         O_TRUNC | O_CREAT | O_RDWR);
2643                 if (ret)
2644                         goto out_restore_hdr;
2645         }
2646
2647         /* Write initial header.  This is merely a "dummy" header since it
2648          * doesn't have all the information yet, so it will be overwritten later
2649          * (unless writing a pipable WIM).  */
2650         if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE))
2651                 wim->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2652         ret = write_wim_header(&wim->hdr, &wim->out_fd);
2653         wim->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2654         if (ret)
2655                 goto out_restore_hdr;
2656
2657         if (stream_list_override) {
2658                 struct wim_lookup_table_entry *lte;
2659                 INIT_LIST_HEAD(&lt_stream_list_override);
2660                 list_for_each_entry(lte, stream_list_override,
2661                                     write_streams_list)
2662                 {
2663                         list_add_tail(&lte->lookup_table_list,
2664                                       &lt_stream_list_override);
2665                 }
2666         }
2667
2668         /* Write metadata resources and streams.  */
2669         if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE)) {
2670                 /* Default case: create a normal (non-pipable) WIM.  */
2671                 ret = write_wim_streams(wim, image, write_flags, num_threads,
2672                                         progress_func, stream_list_override);
2673                 if (ret)
2674                         goto out_restore_hdr;
2675
2676                 ret = write_wim_metadata_resources(wim, image, write_flags,
2677                                                    progress_func);
2678                 if (ret)
2679                         goto out_restore_hdr;
2680         } else {
2681                 /* Non-default case: create pipable WIM.  */
2682                 ret = write_pipable_wim(wim, image, write_flags, num_threads,
2683                                         progress_func, stream_list_override);
2684                 if (ret)
2685                         goto out_restore_hdr;
2686                 write_flags |= WIMLIB_WRITE_FLAG_HEADER_AT_END;
2687         }
2688
2689         if (stream_list_override)
2690                 stream_list_override = &lt_stream_list_override;
2691
2692         /* Write lookup table, XML data, and (optional) integrity table.  */
2693         ret = finish_write(wim, image, write_flags, progress_func,
2694                            stream_list_override);
2695 out_restore_hdr:
2696         memcpy(&wim->hdr, &hdr_save, sizeof(struct wim_header));
2697         (void)close_wim_writable(wim, write_flags);
2698         DEBUG("ret=%d", ret);
2699         return ret;
2700 }
2701
2702 /* Write a standalone WIM to a file or file descriptor.  */
2703 static int
2704 write_standalone_wim(WIMStruct *wim, const void *path_or_fd,
2705                      int image, int write_flags, unsigned num_threads,
2706                      wimlib_progress_func_t progress_func)
2707 {
2708         return write_wim_part(wim, path_or_fd, image, write_flags,
2709                               num_threads, progress_func, 1, 1, NULL, NULL);
2710 }
2711
2712 /* API function documented in wimlib.h  */
2713 WIMLIBAPI int
2714 wimlib_write(WIMStruct *wim, const tchar *path,
2715              int image, int write_flags, unsigned num_threads,
2716              wimlib_progress_func_t progress_func)
2717 {
2718         if (!path)
2719                 return WIMLIB_ERR_INVALID_PARAM;
2720
2721         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2722
2723         return write_standalone_wim(wim, path, image, write_flags,
2724                                     num_threads, progress_func);
2725 }
2726
2727 /* API function documented in wimlib.h  */
2728 WIMLIBAPI int
2729 wimlib_write_to_fd(WIMStruct *wim, int fd,
2730                    int image, int write_flags, unsigned num_threads,
2731                    wimlib_progress_func_t progress_func)
2732 {
2733         if (fd < 0)
2734                 return WIMLIB_ERR_INVALID_PARAM;
2735
2736         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2737         write_flags |= WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR;
2738
2739         return write_standalone_wim(wim, &fd, image, write_flags,
2740                                     num_threads, progress_func);
2741 }
2742
2743 static bool
2744 any_images_modified(WIMStruct *wim)
2745 {
2746         for (int i = 0; i < wim->hdr.image_count; i++)
2747                 if (wim->image_metadata[i]->modified)
2748                         return true;
2749         return false;
2750 }
2751
2752 static int
2753 check_resource_offset(struct wim_lookup_table_entry *lte, void *_wim)
2754 {
2755         const WIMStruct *wim = _wim;
2756         off_t end_offset = *(const off_t*)wim->private;
2757
2758         if (lte->resource_location == RESOURCE_IN_WIM && lte->wim == wim &&
2759             lte->resource_entry.offset + lte->resource_entry.size > end_offset)
2760                 return WIMLIB_ERR_RESOURCE_ORDER;
2761         return 0;
2762 }
2763
2764 /* Make sure no file or metadata resources are located after the XML data (or
2765  * integrity table if present)--- otherwise we can't safely overwrite the WIM in
2766  * place and we return WIMLIB_ERR_RESOURCE_ORDER.  */
2767 static int
2768 check_resource_offsets(WIMStruct *wim, off_t end_offset)
2769 {
2770         int ret;
2771         unsigned i;
2772
2773         wim->private = &end_offset;
2774         ret = for_lookup_table_entry(wim->lookup_table, check_resource_offset, wim);
2775         if (ret)
2776                 return ret;
2777
2778         for (i = 0; i < wim->hdr.image_count; i++) {
2779                 ret = check_resource_offset(wim->image_metadata[i]->metadata_lte, wim);
2780                 if (ret)
2781                         return ret;
2782         }
2783         return 0;
2784 }
2785
2786 /*
2787  * Overwrite a WIM, possibly appending streams to it.
2788  *
2789  * A WIM looks like (or is supposed to look like) the following:
2790  *
2791  *                   Header (212 bytes)
2792  *                   Streams and metadata resources (variable size)
2793  *                   Lookup table (variable size)
2794  *                   XML data (variable size)
2795  *                   Integrity table (optional) (variable size)
2796  *
2797  * If we are not adding any streams or metadata resources, the lookup table is
2798  * unchanged--- so we only need to overwrite the XML data, integrity table, and
2799  * header.  This operation is potentially unsafe if the program is abruptly
2800  * terminated while the XML data or integrity table are being overwritten, but
2801  * before the new header has been written.  To partially alleviate this problem,
2802  * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
2803  * finish_write() to cause a temporary WIM header to be written after the XML
2804  * data has been written.  This may prevent the WIM from becoming corrupted if
2805  * the program is terminated while the integrity table is being calculated (but
2806  * no guarantees, due to write re-ordering...).
2807  *
2808  * If we are adding new streams or images (metadata resources), the lookup table
2809  * needs to be changed, and those streams need to be written.  In this case, we
2810  * try to perform a safe update of the WIM file by writing the streams *after*
2811  * the end of the previous WIM, then writing the new lookup table, XML data, and
2812  * (optionally) integrity table following the new streams.  This will produce a
2813  * layout like the following:
2814  *
2815  *                   Header (212 bytes)
2816  *                   (OLD) Streams and metadata resources (variable size)
2817  *                   (OLD) Lookup table (variable size)
2818  *                   (OLD) XML data (variable size)
2819  *                   (OLD) Integrity table (optional) (variable size)
2820  *                   (NEW) Streams and metadata resources (variable size)
2821  *                   (NEW) Lookup table (variable size)
2822  *                   (NEW) XML data (variable size)
2823  *                   (NEW) Integrity table (optional) (variable size)
2824  *
2825  * At all points, the WIM is valid as nothing points to the new data yet.  Then,
2826  * the header is overwritten to point to the new lookup table, XML data, and
2827  * integrity table, to produce the following layout:
2828  *
2829  *                   Header (212 bytes)
2830  *                   Streams and metadata resources (variable size)
2831  *                   Nothing (variable size)
2832  *                   More Streams and metadata resources (variable size)
2833  *                   Lookup table (variable size)
2834  *                   XML data (variable size)
2835  *                   Integrity table (optional) (variable size)
2836  *
2837  * This method allows an image to be appended to a large WIM very quickly, and
2838  * is is crash-safe except in the case of write re-ordering, but the
2839  * disadvantage is that a small hole is left in the WIM where the old lookup
2840  * table, xml data, and integrity table were.  (These usually only take up a
2841  * small amount of space compared to the streams, however.)
2842  */
2843 static int
2844 overwrite_wim_inplace(WIMStruct *wim, int write_flags,
2845                       unsigned num_threads,
2846                       wimlib_progress_func_t progress_func)
2847 {
2848         int ret;
2849         struct list_head stream_list;
2850         off_t old_wim_end;
2851         u64 old_lookup_table_end, old_xml_begin, old_xml_end;
2852         struct wim_header hdr_save;
2853
2854         DEBUG("Overwriting `%"TS"' in-place", wim->filename);
2855
2856         /* Set default integrity flag.  */
2857         if (!(write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2858                              WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY)))
2859                 if (wim_has_integrity_table(wim))
2860                         write_flags |= WIMLIB_WRITE_FLAG_CHECK_INTEGRITY;
2861
2862         /* Set additional flags for overwrite.  */
2863         write_flags |= WIMLIB_WRITE_FLAG_OVERWRITE |
2864                        WIMLIB_WRITE_FLAG_STREAMS_OK;
2865
2866         /* Make sure that the integrity table (if present) is after the XML
2867          * data, and that there are no stream resources, metadata resources, or
2868          * lookup tables after the XML data.  Otherwise, these data would be
2869          * overwritten. */
2870         old_xml_begin = wim->hdr.xml_res_entry.offset;
2871         old_xml_end = old_xml_begin + wim->hdr.xml_res_entry.size;
2872         old_lookup_table_end = wim->hdr.lookup_table_res_entry.offset +
2873                                wim->hdr.lookup_table_res_entry.size;
2874         if (wim->hdr.integrity.offset != 0 && wim->hdr.integrity.offset < old_xml_end) {
2875                 WARNING("Didn't expect the integrity table to be before the XML data");
2876                 return WIMLIB_ERR_RESOURCE_ORDER;
2877         }
2878
2879         if (old_lookup_table_end > old_xml_begin) {
2880                 WARNING("Didn't expect the lookup table to be after the XML data");
2881                 return WIMLIB_ERR_RESOURCE_ORDER;
2882         }
2883
2884         /* Set @old_wim_end, which indicates the point beyond which we don't
2885          * allow any file and metadata resources to appear without returning
2886          * WIMLIB_ERR_RESOURCE_ORDER (due to the fact that we would otherwise
2887          * overwrite these resources). */
2888         if (!wim->deletion_occurred && !any_images_modified(wim)) {
2889                 /* If no images have been modified and no images have been
2890                  * deleted, a new lookup table does not need to be written.  We
2891                  * shall write the new XML data and optional integrity table
2892                  * immediately after the lookup table.  Note that this may
2893                  * overwrite an existing integrity table. */
2894                 DEBUG("Skipping writing lookup table "
2895                       "(no images modified or deleted)");
2896                 old_wim_end = old_lookup_table_end;
2897                 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
2898                                WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
2899         } else if (wim->hdr.integrity.offset) {
2900                 /* Old WIM has an integrity table; begin writing new streams
2901                  * after it. */
2902                 old_wim_end = wim->hdr.integrity.offset + wim->hdr.integrity.size;
2903         } else {
2904                 /* No existing integrity table; begin writing new streams after
2905                  * the old XML data. */
2906                 old_wim_end = old_xml_end;
2907         }
2908
2909         ret = check_resource_offsets(wim, old_wim_end);
2910         if (ret)
2911                 return ret;
2912
2913         ret = prepare_stream_list(wim, WIMLIB_ALL_IMAGES, write_flags,
2914                                   &stream_list);
2915         if (ret)
2916                 return ret;
2917
2918         ret = open_wim_writable(wim, wim->filename, O_RDWR);
2919         if (ret)
2920                 return ret;
2921
2922         ret = lock_wim(wim, wim->out_fd.fd);
2923         if (ret)
2924                 goto out_close_wim;
2925
2926         /* Save original header so it can be restored in case of error  */
2927         memcpy(&hdr_save, &wim->hdr, sizeof(struct wim_header));
2928
2929         /* Set WIM_HDR_FLAG_WRITE_IN_PROGRESS flag in header. */
2930         wim->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2931         ret = write_wim_header_flags(wim->hdr.flags, &wim->out_fd);
2932         if (ret) {
2933                 ERROR_WITH_ERRNO("Error updating WIM header flags");
2934                 goto out_restore_memory_hdr;
2935         }
2936
2937         if (filedes_seek(&wim->out_fd, old_wim_end) == -1) {
2938                 ERROR_WITH_ERRNO("Can't seek to end of WIM");
2939                 ret = WIMLIB_ERR_WRITE;
2940                 goto out_restore_physical_hdr;
2941         }
2942
2943         ret = write_stream_list(&stream_list,
2944                                 wim->lookup_table,
2945                                 &wim->out_fd,
2946                                 wim->compression_type,
2947                                 &wim->lzx_context,
2948                                 write_flags,
2949                                 num_threads,
2950                                 progress_func);
2951         if (ret)
2952                 goto out_truncate;
2953
2954         ret = write_wim_metadata_resources(wim, WIMLIB_ALL_IMAGES,
2955                                            write_flags, progress_func);
2956         if (ret)
2957                 goto out_truncate;
2958
2959         write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
2960         ret = finish_write(wim, WIMLIB_ALL_IMAGES, write_flags,
2961                            progress_func, NULL);
2962         if (ret)
2963                 goto out_truncate;
2964
2965         goto out_unlock_wim;
2966
2967 out_truncate:
2968         if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
2969                 WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
2970                         wim->filename, old_wim_end);
2971                 /* Return value of ftruncate() is ignored because this is
2972                  * already an error path.  */
2973                 (void)ftruncate(wim->out_fd.fd, old_wim_end);
2974         }
2975 out_restore_physical_hdr:
2976         (void)write_wim_header_flags(hdr_save.flags, &wim->out_fd);
2977 out_restore_memory_hdr:
2978         memcpy(&wim->hdr, &hdr_save, sizeof(struct wim_header));
2979 out_close_wim:
2980         (void)close_wim_writable(wim, write_flags);
2981 out_unlock_wim:
2982         wim->wim_locked = 0;
2983         return ret;
2984 }
2985
2986 static int
2987 overwrite_wim_via_tmpfile(WIMStruct *wim, int write_flags,
2988                           unsigned num_threads,
2989                           wimlib_progress_func_t progress_func)
2990 {
2991         size_t wim_name_len;
2992         int ret;
2993
2994         DEBUG("Overwriting `%"TS"' via a temporary file", wim->filename);
2995
2996         /* Write the WIM to a temporary file in the same directory as the
2997          * original WIM. */
2998         wim_name_len = tstrlen(wim->filename);
2999         tchar tmpfile[wim_name_len + 10];
3000         tmemcpy(tmpfile, wim->filename, wim_name_len);
3001         randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
3002         tmpfile[wim_name_len + 9] = T('\0');
3003
3004         ret = wimlib_write(wim, tmpfile, WIMLIB_ALL_IMAGES,
3005                            write_flags | WIMLIB_WRITE_FLAG_FSYNC,
3006                            num_threads, progress_func);
3007         if (ret) {
3008                 tunlink(tmpfile);
3009                 return ret;
3010         }
3011
3012         close_wim(wim);
3013
3014         /* Rename the new WIM file to the original WIM file.  Note: on Windows
3015          * this actually calls win32_rename_replacement(), not _wrename(), so
3016          * that removing the existing destination file can be handled.  */
3017         DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, wim->filename);
3018         ret = trename(tmpfile, wim->filename);
3019         if (ret) {
3020                 ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
3021                                  tmpfile, wim->filename);
3022         #ifdef __WIN32__
3023                 if (ret < 0)
3024         #endif
3025                 {
3026                         tunlink(tmpfile);
3027                 }
3028                 return WIMLIB_ERR_RENAME;
3029         }
3030
3031         if (progress_func) {
3032                 union wimlib_progress_info progress;
3033                 progress.rename.from = tmpfile;
3034                 progress.rename.to = wim->filename;
3035                 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
3036         }
3037         return 0;
3038 }
3039
3040 /* API function documented in wimlib.h  */
3041 WIMLIBAPI int
3042 wimlib_overwrite(WIMStruct *wim, int write_flags,
3043                  unsigned num_threads,
3044                  wimlib_progress_func_t progress_func)
3045 {
3046         int ret;
3047         u32 orig_hdr_flags;
3048
3049         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
3050
3051         if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
3052                 return WIMLIB_ERR_INVALID_PARAM;
3053
3054         if (!wim->filename)
3055                 return WIMLIB_ERR_NO_FILENAME;
3056
3057         orig_hdr_flags = wim->hdr.flags;
3058         if (write_flags & WIMLIB_WRITE_FLAG_IGNORE_READONLY_FLAG)
3059                 wim->hdr.flags &= ~WIM_HDR_FLAG_READONLY;
3060         ret = can_modify_wim(wim);
3061         wim->hdr.flags = orig_hdr_flags;
3062         if (ret)
3063                 return ret;
3064
3065         if ((!wim->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
3066             && !(write_flags & (WIMLIB_WRITE_FLAG_REBUILD |
3067                                 WIMLIB_WRITE_FLAG_PIPABLE))
3068             && !(wim_is_pipable(wim))
3069             && wim->compression_type == wim->out_compression_type)
3070         {
3071                 ret = overwrite_wim_inplace(wim, write_flags, num_threads,
3072                                             progress_func);
3073                 if (ret != WIMLIB_ERR_RESOURCE_ORDER)
3074                         return ret;
3075                 WARNING("Falling back to re-building entire WIM");
3076         }
3077         return overwrite_wim_via_tmpfile(wim, write_flags, num_threads,
3078                                          progress_func);
3079 }