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 /* Return true if the specified resource is compressed and the compressed data
  72  * can be reused with the specified output parameters.  */
  73 static bool
  74 can_raw_copy(const struct wim_lookup_table_entry *lte,
  75              int write_resource_flags, int out_ctype, u32 out_chunk_size)
  76 {
  77         if (lte->resource_location != RESOURCE_IN_WIM)
  78                 return false;
  79         if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE)
  80                 return false;
  81         if (lte->rspec->ctype != out_ctype)
  82                 return false;
  83         if (out_chunk_size != lte->rspec->cchunk_size)
  84                 return false;
  85         if (lte->offset_in_res != 0)
  86                 return false;
  87         if (lte->size != lte->rspec->uncompressed_size)
  88                 return false;
  89         return true;
  90 }
  91
  92
  93 /* Return true if the specified resource must be recompressed when the specified
  94  * output parameters are used.  */
  95 static bool
  96 must_compress_stream(const struct wim_lookup_table_entry *lte,
  97                      int write_resource_flags, int out_ctype, u32 out_chunk_size)
  98 {
  99         return (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE
 100                 && ((write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS)
 101                     || !can_raw_copy(lte, write_resource_flags,
 102                                      out_ctype, out_chunk_size)));
 103 }
 104
 105 static unsigned
 106 compress_chunk(const void * uncompressed_data,
 107                unsigned uncompressed_len,
 108                void *compressed_data,
 109                int out_ctype,
 110                struct wimlib_lzx_context *comp_ctx)
 111 {
 112         switch (out_ctype) {
 113         case WIMLIB_COMPRESSION_TYPE_XPRESS:
 114                 return wimlib_xpress_compress(uncompressed_data,
 115                                               uncompressed_len,
 116                                               compressed_data);
 117         case WIMLIB_COMPRESSION_TYPE_LZX:
 118                 return wimlib_lzx_compress2(uncompressed_data,
 119                                             uncompressed_len,
 120                                             compressed_data,
 121                                             comp_ctx);
 122         default:
 123                 wimlib_assert(0);
 124                 return 0;
 125         }
 126 }
 127
 128 /* Chunk table that's located at the beginning of each compressed resource in
 129  * the WIM.  (This is not the on-disk format; the on-disk format just has an
 130  * array of offsets.) */
 131 struct chunk_table {
 132         u64 original_resource_size;
 133         u64 num_chunks;
 134         u64 table_disk_size;
 135         unsigned bytes_per_chunk_entry;
 136         void *cur_offset_p;
 137         union {
 138                 u32 cur_offset_u32;
 139                 u64 cur_offset_u64;
 140         };
 141         /* Beginning of chunk offsets, in either 32-bit or 64-bit little endian
 142          * integers, including the first offset of 0, which will not be written.
 143          * */
 144         u8 offsets[] _aligned_attribute(8);
 145 };
 146
 147 /* Allocate and initializes a chunk table, then reserve space for it in the
 148  * output file unless writing a pipable resource.  */
 149 static int
 150 begin_wim_resource_chunk_tab(const struct wim_lookup_table_entry *lte,
 151                              struct filedes *out_fd,
 152                              u32 out_chunk_size,
 153                              struct chunk_table **chunk_tab_ret,
 154                              int resource_flags)
 155 {
 156         u64 size;
 157         u64 num_chunks;
 158         unsigned bytes_per_chunk_entry;
 159         size_t alloc_size;
 160         struct chunk_table *chunk_tab;
 161         int ret;
 162
 163         size = lte->size;
 164         num_chunks = DIV_ROUND_UP(size, out_chunk_size);
 165         bytes_per_chunk_entry = (size > (1ULL << 32)) ? 8 : 4;
 166         alloc_size = sizeof(struct chunk_table) + num_chunks * sizeof(u64);
 167         chunk_tab = CALLOC(1, alloc_size);
 168
 169         if (!chunk_tab) {
 170                 ERROR("Failed to allocate chunk table for %"PRIu64" byte "
 171                       "resource", size);
 172                 return WIMLIB_ERR_NOMEM;
 173         }
 174         chunk_tab->num_chunks = num_chunks;
 175         chunk_tab->original_resource_size = size;
 176         chunk_tab->bytes_per_chunk_entry = bytes_per_chunk_entry;
 177         chunk_tab->table_disk_size = chunk_tab->bytes_per_chunk_entry *
 178                                      (num_chunks - 1);
 179         chunk_tab->cur_offset_p = chunk_tab->offsets;
 180
 181         /* We don't know the correct offsets yet; so just write zeroes to
 182          * reserve space for the table, so we can go back to it later after
 183          * we've written the compressed chunks following it.
 184          *
 185          * Special case: if writing a pipable WIM, compressed resources are in a
 186          * modified format (see comment above write_pipable_wim()) and do not
 187          * have a chunk table at the beginning, so don't reserve any space for
 188          * one.  */
 189         if (!(resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE)) {
 190                 ret = full_write(out_fd, chunk_tab->offsets,
 191                                  chunk_tab->table_disk_size);
 192                 if (ret) {
 193                         ERROR_WITH_ERRNO("Failed to write chunk table in compressed "
 194                                          "file resource");
 195                         FREE(chunk_tab);
 196                         return ret;
 197                 }
 198         }
 199         *chunk_tab_ret = chunk_tab;
 200         return 0;
 201 }
 202
 203 /* Add the offset for the next chunk to the chunk table being constructed for a
 204  * compressed stream. */
 205 static void
 206 chunk_tab_record_chunk(struct chunk_table *chunk_tab, unsigned out_chunk_size)
 207 {
 208         if (chunk_tab->bytes_per_chunk_entry == 4) {
 209                 *(le32*)chunk_tab->cur_offset_p = cpu_to_le32(chunk_tab->cur_offset_u32);
 210                 chunk_tab->cur_offset_p = (le32*)chunk_tab->cur_offset_p + 1;
 211                 chunk_tab->cur_offset_u32 += out_chunk_size;
 212         } else {
 213                 *(le64*)chunk_tab->cur_offset_p = cpu_to_le64(chunk_tab->cur_offset_u64);
 214                 chunk_tab->cur_offset_p = (le64*)chunk_tab->cur_offset_p + 1;
 215                 chunk_tab->cur_offset_u64 += out_chunk_size;
 216         }
 217 }
 218
 219 /* Finishes a WIM chunk table and writes it to the output file at the correct
 220  * offset.  */
 221 static int
 222 finish_wim_resource_chunk_tab(struct chunk_table *chunk_tab,
 223                               struct filedes *out_fd,
 224                               off_t res_start_offset,
 225                               int write_resource_flags)
 226 {
 227         int ret;
 228
 229         if (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
 230                 ret = full_write(out_fd,
 231                                  chunk_tab->offsets +
 232                                          chunk_tab->bytes_per_chunk_entry,
 233                                  chunk_tab->table_disk_size);
 234         } else {
 235                 ret = full_pwrite(out_fd,
 236                                   chunk_tab->offsets +
 237                                           chunk_tab->bytes_per_chunk_entry,
 238                                   chunk_tab->table_disk_size,
 239                                   res_start_offset);
 240         }
 241         if (ret)
 242                 ERROR_WITH_ERRNO("Write error");
 243         return ret;
 244 }
 245
 246 /* Write the header for a stream in a pipable WIM.
 247  */
 248 static int
 249 write_pwm_stream_header(const struct wim_lookup_table_entry *lte,
 250                         struct filedes *out_fd,
 251                         int additional_reshdr_flags)
 252 {
 253         struct pwm_stream_hdr stream_hdr;
 254         u32 reshdr_flags;
 255         int ret;
 256
 257         stream_hdr.magic = PWM_STREAM_MAGIC;
 258         stream_hdr.uncompressed_size = cpu_to_le64(lte->size);
 259         if (additional_reshdr_flags & PWM_RESHDR_FLAG_UNHASHED) {
 260                 zero_out_hash(stream_hdr.hash);
 261         } else {
 262                 wimlib_assert(!lte->unhashed);
 263                 copy_hash(stream_hdr.hash, lte->hash);
 264         }
 265
 266         reshdr_flags = lte->flags & ~(WIM_RESHDR_FLAG_COMPRESSED | WIM_RESHDR_FLAG_CONCAT);
 267         reshdr_flags |= additional_reshdr_flags;
 268         stream_hdr.flags = cpu_to_le32(reshdr_flags);
 269         ret = full_write(out_fd, &stream_hdr, sizeof(stream_hdr));
 270         if (ret)
 271                 ERROR_WITH_ERRNO("Error writing stream header");
 272         return ret;
 273 }
 274
 275 static int
 276 seek_and_truncate(struct filedes *out_fd, off_t offset)
 277 {
 278         if (filedes_seek(out_fd, offset) == -1 ||
 279             ftruncate(out_fd->fd, offset))
 280         {
 281                 ERROR_WITH_ERRNO("Failed to truncate output WIM file");
 282                 return WIMLIB_ERR_WRITE;
 283         }
 284         return 0;
 285 }
 286
 287 static int
 288 finalize_and_check_sha1(SHA_CTX *sha_ctx, struct wim_lookup_table_entry *lte)
 289 {
 290         u8 md[SHA1_HASH_SIZE];
 291
 292         sha1_final(md, sha_ctx);
 293         if (lte->unhashed) {
 294                 copy_hash(lte->hash, md);
 295         } else if (!hashes_equal(md, lte->hash)) {
 296                 ERROR("WIM resource has incorrect hash!");
 297                 if (lte_filename_valid(lte)) {
 298                         ERROR("We were reading it from \"%"TS"\"; maybe "
 299                               "it changed while we were reading it.",
 300                               lte->file_on_disk);
 301                 }
 302                 return WIMLIB_ERR_INVALID_RESOURCE_HASH;
 303         }
 304         return 0;
 305 }
 306
 307 struct write_resource_ctx {
 308         int out_ctype;
 309         u32 out_chunk_size;
 310         struct wimlib_lzx_context *comp_ctx;
 311         struct chunk_table *chunk_tab;
 312         struct filedes *out_fd;
 313         SHA_CTX sha_ctx;
 314         bool doing_sha;
 315         int resource_flags;
 316 };
 317
 318 static int
 319 write_resource_cb(const void *chunk, size_t chunk_size, void *_ctx)
 320 {
 321         struct write_resource_ctx *ctx = _ctx;
 322         const void *out_chunk;
 323         unsigned out_chunk_size;
 324         int ret;
 325         void *compressed_chunk = NULL;
 326         unsigned compressed_size;
 327         bool compressed_chunk_malloced = false;
 328
 329         if (ctx->doing_sha)
 330                 sha1_update(&ctx->sha_ctx, chunk, chunk_size);
 331
 332         out_chunk = chunk;
 333         out_chunk_size = chunk_size;
 334         if (ctx->out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
 335
 336                 /* Compress the chunk.  */
 337                 if (chunk_size <= STACK_MAX) {
 338                         compressed_chunk = alloca(chunk_size);
 339                 } else {
 340                         compressed_chunk = MALLOC(chunk_size);
 341                         if (compressed_chunk == NULL)
 342                                 return WIMLIB_ERR_NOMEM;
 343                         compressed_chunk_malloced = true;
 344                 }
 345
 346                 compressed_size = compress_chunk(chunk, chunk_size,
 347                                                  compressed_chunk,
 348                                                  ctx->out_ctype,
 349                                                  ctx->comp_ctx);
 350                 /* Use compressed data if compression to less than input size
 351                  * was successful.  */
 352                 if (compressed_size) {
 353                         out_chunk = compressed_chunk;
 354                         out_chunk_size = compressed_size;
 355                 }
 356         }
 357
 358         if (ctx->chunk_tab) {
 359                 /* Update chunk table accounting.  */
 360                 chunk_tab_record_chunk(ctx->chunk_tab, out_chunk_size);
 361
 362                 /* If writing compressed chunks to a pipable WIM, before the
 363                  * chunk data write a chunk header that provides the compressed
 364                  * chunk size.  */
 365                 if (ctx->resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
 366                         struct pwm_chunk_hdr chunk_hdr = {
 367                                 .compressed_size = cpu_to_le32(out_chunk_size),
 368                         };
 369                         ret = full_write(ctx->out_fd, &chunk_hdr,
 370                                          sizeof(chunk_hdr));
 371                         if (ret)
 372                                 goto error;
 373                 }
 374         }
 375
 376         /* Write the chunk data.  */
 377         ret = full_write(ctx->out_fd, out_chunk, out_chunk_size);
 378         if (ret)
 379                 goto error;
 380
 381 out_free_memory:
 382         if (compressed_chunk_malloced)
 383                 FREE(compressed_chunk);
 384         return ret;
 385
 386 error:
 387         ERROR_WITH_ERRNO("Failed to write WIM resource chunk");
 388         goto out_free_memory;
 389 }
 390
 391 /*
 392  * write_wim_resource()-
 393  *
 394  * Write a resource to an output WIM.
 395  *
 396  * @lte:
 397  *      Lookup table entry for the resource, which could be in another WIM, in
 398  *      an external file, or in another location.
 399  *
 400  * @out_fd:
 401  *      File descriptor opened to the output WIM.
 402  *
 403  * @out_ctype:
 404  *      One of the WIMLIB_COMPRESSION_TYPE_* constants to indicate which
 405  *      compression algorithm to use.
 406  *
 407  * @out_chunk_size:
 408  *      Compressed chunk size to use.
 409  *
 410  * @out_reshdr:
 411  *      On success, this is filled in with the offset, flags, compressed size,
 412  *      and uncompressed size of the resource in the output WIM.
 413  *
 414  * @resource_flags:
 415  *      * WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS to force data to be recompressed even
 416  *        if it could otherwise be copied directly from the input;
 417  *      * WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE if writing a resource for a pipable WIM
 418  *        (and the output file descriptor may be a pipe).
 419  *
 420  * @comp_ctx:
 421  *      Location of LZX compression context pointer, which will be allocated or
 422  *      updated if needed.  (Initialize to NULL.)
 423  *
 424  * Additional notes:  The SHA1 message digest of the uncompressed data is
 425  * calculated (except when doing a raw copy --- see below).  If the @unhashed
 426  * flag is set on the lookup table entry, this message digest is simply copied
 427  * to it; otherwise, the message digest is compared with the existing one, and
 428  * this function will fail if they do not match.
 429  */
 430 static int
 431 write_wim_resource(struct wim_lookup_table_entry *lte,
 432                    struct filedes *out_fd, int out_ctype,
 433                    u32 out_chunk_size,
 434                    struct wim_reshdr *out_reshdr,
 435                    int resource_flags,
 436                    struct wimlib_lzx_context **comp_ctx)
 437 {
 438         struct write_resource_ctx write_ctx;
 439         off_t res_start_offset;
 440         u32 in_chunk_size;
 441         u64 read_size;
 442         int ret;
 443
 444         /* Mask out any irrelevant flags, since this function also uses this
 445          * variable to store WIMLIB_READ_RESOURCE flags.  */
 446         resource_flags &= WIMLIB_WRITE_RESOURCE_MASK;
 447
 448         /* Get current position in output WIM.  */
 449         res_start_offset = out_fd->offset;
 450
 451         /* If we are not forcing the data to be recompressed, and the input
 452          * resource is located in a WIM with a compression mode compatible with
 453          * the output, we can simply copy the compressed data without
 454          * recompressing it.  This also means we must skip calculating the SHA1,
 455          * as we never will see the uncompressed data.  */
 456         if (can_raw_copy(lte, resource_flags, out_ctype, out_chunk_size)) {
 457                 /* Normally, for raw copies we can request a RAW_FULL read, but
 458                  * if we're reading from a pipable resource and writing a
 459                  * non-pipable resource or vice versa, then a RAW_CHUNKS read
 460                  * needs to be requested so that the written resource can be
 461                  * appropriately formatted.  However, in neither case is any
 462                  * actual decompression needed.  */
 463                 if (lte->rspec->is_pipable == !!(resource_flags &
 464                                                  WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE))
 465                 {
 466                         resource_flags |= WIMLIB_READ_RESOURCE_FLAG_RAW_FULL;
 467                         read_size = lte->rspec->size_in_wim;
 468                 } else {
 469                         resource_flags |= WIMLIB_READ_RESOURCE_FLAG_RAW_CHUNKS;
 470                         read_size = lte->size;
 471                 }
 472                 write_ctx.doing_sha = false;
 473         } else {
 474                 write_ctx.doing_sha = true;
 475                 sha1_init(&write_ctx.sha_ctx);
 476                 read_size = lte->size;
 477         }
 478
 479         /* Set the output compression mode and initialize chunk table if needed.
 480          */
 481         write_ctx.out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
 482         write_ctx.out_chunk_size = out_chunk_size;
 483         write_ctx.chunk_tab = NULL;
 484         if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
 485                 wimlib_assert(out_chunk_size > 0);
 486                 if (!(resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW)) {
 487                         /* Compression needed.  */
 488                         write_ctx.out_ctype = out_ctype;
 489                         if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX) {
 490                                 ret = wimlib_lzx_alloc_context(out_chunk_size,
 491                                                                NULL, comp_ctx);
 492                                 if (ret)
 493                                         goto out;
 494                         }
 495                         write_ctx.comp_ctx = *comp_ctx;
 496                 }
 497                 if (!(resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW_FULL)) {
 498                         /* Chunk table needed.  */
 499                         ret = begin_wim_resource_chunk_tab(lte, out_fd,
 500                                                            out_chunk_size,
 501                                                            &write_ctx.chunk_tab,
 502                                                            resource_flags);
 503                         if (ret)
 504                                 goto out;
 505                 }
 506         }
 507
 508         /* If writing a pipable resource, write the stream header and update
 509          * @res_start_offset to be the end of the stream header.  */
 510         if (resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
 511                 int reshdr_flags = 0;
 512                 if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE)
 513                         reshdr_flags |= WIM_RESHDR_FLAG_COMPRESSED;
 514                 ret = write_pwm_stream_header(lte, out_fd, reshdr_flags);
 515                 if (ret)
 516                         goto out_free_chunk_tab;
 517                 res_start_offset = out_fd->offset;
 518         }
 519
 520         /* Write the entire resource by reading the entire resource and feeding
 521          * the data through write_resource_cb().  */
 522         write_ctx.out_fd = out_fd;
 523         write_ctx.resource_flags = resource_flags;
 524 try_write_again:
 525         if (write_ctx.out_ctype == WIMLIB_COMPRESSION_TYPE_NONE)
 526                 in_chunk_size = lte_cchunk_size(lte);
 527         else
 528                 in_chunk_size = out_chunk_size;
 529         ret = read_resource_prefix(lte, read_size,
 530                                    write_resource_cb,
 531                                    in_chunk_size, &write_ctx, resource_flags);
 532         if (ret)
 533                 goto out_free_chunk_tab;
 534
 535         /* Verify SHA1 message digest of the resource, or set the hash for the
 536          * first time. */
 537         if (write_ctx.doing_sha) {
 538                 ret = finalize_and_check_sha1(&write_ctx.sha_ctx, lte);
 539                 if (ret)
 540                         goto out_free_chunk_tab;
 541         }
 542
 543         /* Write chunk table if needed.  */
 544         if (write_ctx.chunk_tab) {
 545                 ret = finish_wim_resource_chunk_tab(write_ctx.chunk_tab,
 546                                                     out_fd,
 547                                                     res_start_offset,
 548                                                     resource_flags);
 549                 if (ret)
 550                         goto out_free_chunk_tab;
 551         }
 552
 553         /* Fill in out_reshdr with information about the newly written
 554          * resource.  */
 555         out_reshdr->size_in_wim   = out_fd->offset - res_start_offset;
 556         out_reshdr->flags         = lte->flags & ~WIM_RESHDR_FLAG_CONCAT;
 557         if (out_ctype == WIMLIB_COMPRESSION_TYPE_NONE)
 558                 out_reshdr->flags &= ~WIM_RESHDR_FLAG_COMPRESSED;
 559         else
 560                 out_reshdr->flags |= WIM_RESHDR_FLAG_COMPRESSED;
 561         out_reshdr->offset_in_wim  = res_start_offset;
 562         out_reshdr->uncompressed_size = lte->size;
 563
 564         /* Check for resources compressed to greater than their original size
 565          * and write them uncompressed instead.  (But never do this if writing
 566          * to a pipe, and don't bother if we did a raw copy.)  */
 567         if (out_reshdr->size_in_wim > out_reshdr->uncompressed_size &&
 568             !(resource_flags & (WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE |
 569                                 WIMLIB_READ_RESOURCE_FLAG_RAW)))
 570         {
 571                 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
 572                       "writing uncompressed instead",
 573                       out_reshdr->uncompressed_size, out_reshdr->size_in_wim);
 574                 ret = seek_and_truncate(out_fd, res_start_offset);
 575                 if (ret)
 576                         goto out_free_chunk_tab;
 577                 out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
 578                 FREE(write_ctx.chunk_tab);
 579                 write_ctx.out_ctype = WIMLIB_COMPRESSION_TYPE_NONE;
 580                 write_ctx.chunk_tab = NULL;
 581                 write_ctx.doing_sha = false;
 582                 goto try_write_again;
 583         }
 584         if (resource_flags & WIMLIB_READ_RESOURCE_FLAG_RAW) {
 585                 DEBUG("Copied raw compressed data "
 586                       "(%"PRIu64" => %"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
 587                       out_reshdr->uncompressed_size, out_reshdr->size_in_wim,
 588                       out_reshdr->offset_in_wim, out_reshdr->flags);
 589         } else if (out_ctype != WIMLIB_COMPRESSION_TYPE_NONE) {
 590                 DEBUG("Wrote compressed resource "
 591                       "(%"PRIu64" => %"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
 592                       out_reshdr->uncompressed_size, out_reshdr->size_in_wim,
 593                       out_reshdr->offset_in_wim, out_reshdr->flags);
 594         } else {
 595                 DEBUG("Wrote uncompressed resource "
 596                       "(%"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
 597                       out_reshdr->uncompressed_size,
 598                       out_reshdr->offset_in_wim, out_reshdr->flags);
 599         }
 600         ret = 0;
 601 out_free_chunk_tab:
 602         FREE(write_ctx.chunk_tab);
 603 out:
 604         return ret;
 605 }
 606
 607 /* Like write_wim_resource(), but the resource is specified by a buffer of
 608  * uncompressed data rather a lookup table entry.  Also writes the SHA1 message
 609  * digest of the buffer to @hash_ret if it is non-NULL.  */
 610 int
 611 write_wim_resource_from_buffer(const void *buf, size_t buf_size,
 612                                int reshdr_flags, struct filedes *out_fd,
 613                                int out_ctype,
 614                                u32 out_chunk_size,
 615                                struct wim_reshdr *out_reshdr,
 616                                u8 *hash_ret, int write_resource_flags,
 617                                struct wimlib_lzx_context **comp_ctx)
 618 {
 619         int ret;
 620         struct wim_lookup_table_entry *lte;
 621
 622         /* Set up a temporary lookup table entry to provide to
 623          * write_wim_resource().  */
 624
 625         lte = new_lookup_table_entry();
 626         if (lte == NULL)
 627                 return WIMLIB_ERR_NOMEM;
 628
 629         lte->resource_location  = RESOURCE_IN_ATTACHED_BUFFER;
 630         lte->attached_buffer    = (void*)buf;
 631         lte->size               = buf_size;
 632         lte->flags              = reshdr_flags;
 633
 634         if (write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE) {
 635                 sha1_buffer(buf, buf_size, lte->hash);
 636                 lte->unhashed = 0;
 637         } else {
 638                 lte->unhashed = 1;
 639         }
 640
 641         ret = write_wim_resource(lte, out_fd, out_ctype, out_chunk_size,
 642                                  out_reshdr, write_resource_flags, comp_ctx);
 643         if (ret)
 644                 goto out_free_lte;
 645         if (hash_ret)
 646                 copy_hash(hash_ret, lte->hash);
 647         ret = 0;
 648 out_free_lte:
 649         lte->resource_location = RESOURCE_NONEXISTENT;
 650         free_lookup_table_entry(lte);
 651         return ret;
 652 }
 653
 654
 655 #ifdef ENABLE_MULTITHREADED_COMPRESSION
 656
 657 /* Blocking shared queue (solves the producer-consumer problem) */
 658 struct shared_queue {
 659         unsigned size;
 660         unsigned front;
 661         unsigned back;
 662         unsigned filled_slots;
 663         void **array;
 664         pthread_mutex_t lock;
 665         pthread_cond_t msg_avail_cond;
 666         pthread_cond_t space_avail_cond;
 667 };
 668
 669 static int
 670 shared_queue_init(struct shared_queue *q, unsigned size)
 671 {
 672         wimlib_assert(size != 0);
 673         q->array = CALLOC(sizeof(q->array[0]), size);
 674         if (!q->array)
 675                 goto err;
 676         q->filled_slots = 0;
 677         q->front = 0;
 678         q->back = size - 1;
 679         q->size = size;
 680         if (pthread_mutex_init(&q->lock, NULL)) {
 681                 ERROR_WITH_ERRNO("Failed to initialize mutex");
 682                 goto err;
 683         }
 684         if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
 685                 ERROR_WITH_ERRNO("Failed to initialize condition variable");
 686                 goto err_destroy_lock;
 687         }
 688         if (pthread_cond_init(&q->space_avail_cond, NULL)) {
 689                 ERROR_WITH_ERRNO("Failed to initialize condition variable");
 690                 goto err_destroy_msg_avail_cond;
 691         }
 692         return 0;
 693 err_destroy_msg_avail_cond:
 694         pthread_cond_destroy(&q->msg_avail_cond);
 695 err_destroy_lock:
 696         pthread_mutex_destroy(&q->lock);
 697 err:
 698         return WIMLIB_ERR_NOMEM;
 699 }
 700
 701 static void
 702 shared_queue_destroy(struct shared_queue *q)
 703 {
 704         FREE(q->array);
 705         pthread_mutex_destroy(&q->lock);
 706         pthread_cond_destroy(&q->msg_avail_cond);
 707         pthread_cond_destroy(&q->space_avail_cond);
 708 }
 709
 710 static void
 711 shared_queue_put(struct shared_queue *q, void *obj)
 712 {
 713         pthread_mutex_lock(&q->lock);
 714         while (q->filled_slots == q->size)
 715                 pthread_cond_wait(&q->space_avail_cond, &q->lock);
 716
 717         q->back = (q->back + 1) % q->size;
 718         q->array[q->back] = obj;
 719         q->filled_slots++;
 720
 721         pthread_cond_broadcast(&q->msg_avail_cond);
 722         pthread_mutex_unlock(&q->lock);
 723 }
 724
 725 static void *
 726 shared_queue_get(struct shared_queue *q)
 727 {
 728         void *obj;
 729
 730         pthread_mutex_lock(&q->lock);
 731         while (q->filled_slots == 0)
 732                 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
 733
 734         obj = q->array[q->front];
 735         q->array[q->front] = NULL;
 736         q->front = (q->front + 1) % q->size;
 737         q->filled_slots--;
 738
 739         pthread_cond_broadcast(&q->space_avail_cond);
 740         pthread_mutex_unlock(&q->lock);
 741         return obj;
 742 }
 743
 744 struct compressor_thread_params {
 745         struct shared_queue *res_to_compress_queue;
 746         struct shared_queue *compressed_res_queue;
 747         int out_ctype;
 748         struct wimlib_lzx_context *comp_ctx;
 749 };
 750
 751 #define MAX_CHUNKS_PER_MSG 2
 752
 753 struct message {
 754         struct wim_lookup_table_entry *lte;
 755         u32 out_chunk_size;
 756         u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
 757         u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
 758         unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
 759         struct iovec out_chunks[MAX_CHUNKS_PER_MSG];
 760         unsigned num_chunks;
 761         struct list_head list;
 762         bool complete;
 763         u64 begin_chunk;
 764 };
 765
 766 static void
 767 compress_chunks(struct message *msg, int out_ctype,
 768                 struct wimlib_lzx_context *comp_ctx)
 769 {
 770         for (unsigned i = 0; i < msg->num_chunks; i++) {
 771                 unsigned len;
 772
 773                 len = compress_chunk(msg->uncompressed_chunks[i],
 774                                      msg->uncompressed_chunk_sizes[i],
 775                                      msg->compressed_chunks[i],
 776                                      out_ctype,
 777                                      comp_ctx);
 778
 779                 void *out_chunk;
 780                 unsigned out_len;
 781                 if (len) {
 782                         /* To be written compressed */
 783                         out_chunk = msg->compressed_chunks[i];
 784                         out_len = len;
 785                 } else {
 786                         /* To be written uncompressed */
 787                         out_chunk = msg->uncompressed_chunks[i];
 788                         out_len = msg->uncompressed_chunk_sizes[i];
 789                 }
 790                 msg->out_chunks[i].iov_base = out_chunk;
 791                 msg->out_chunks[i].iov_len = out_len;
 792         }
 793 }
 794
 795 /* Compressor thread routine.  This is a lot simpler than the main thread
 796  * routine: just repeatedly get a group of chunks from the
 797  * res_to_compress_queue, compress them, and put them in the
 798  * compressed_res_queue.  A NULL pointer indicates that the thread should stop.
 799  * */
 800 static void *
 801 compressor_thread_proc(void *arg)
 802 {
 803         struct compressor_thread_params *params = arg;
 804         struct shared_queue *res_to_compress_queue = params->res_to_compress_queue;
 805         struct shared_queue *compressed_res_queue = params->compressed_res_queue;
 806         struct message *msg;
 807
 808         DEBUG("Compressor thread ready");
 809         while ((msg = shared_queue_get(res_to_compress_queue)) != NULL) {
 810                 compress_chunks(msg, params->out_ctype, params->comp_ctx);
 811                 shared_queue_put(compressed_res_queue, msg);
 812         }
 813         DEBUG("Compressor thread terminating");
 814         return NULL;
 815 }
 816 #endif /* ENABLE_MULTITHREADED_COMPRESSION */
 817
 818 struct write_streams_progress_data {
 819         wimlib_progress_func_t progress_func;
 820         union wimlib_progress_info progress;
 821         uint64_t next_progress;
 822         WIMStruct *prev_wim_part;
 823 };
 824
 825 static void
 826 do_write_streams_progress(struct write_streams_progress_data *progress_data,
 827                           struct wim_lookup_table_entry *lte,
 828                           bool stream_discarded)
 829 {
 830         union wimlib_progress_info *progress = &progress_data->progress;
 831         bool new_wim_part;
 832
 833         if (stream_discarded) {
 834                 progress->write_streams.total_bytes -= lte->size;
 835                 if (progress_data->next_progress != ~(uint64_t)0 &&
 836                     progress_data->next_progress > progress->write_streams.total_bytes)
 837                 {
 838                         progress_data->next_progress = progress->write_streams.total_bytes;
 839                 }
 840         } else {
 841                 progress->write_streams.completed_bytes += lte->size;
 842         }
 843         new_wim_part = false;
 844         if (lte->resource_location == RESOURCE_IN_WIM &&
 845             lte->rspec->wim != progress_data->prev_wim_part)
 846         {
 847                 if (progress_data->prev_wim_part) {
 848                         new_wim_part = true;
 849                         progress->write_streams.completed_parts++;
 850                 }
 851                 progress_data->prev_wim_part = lte->rspec->wim;
 852         }
 853         progress->write_streams.completed_streams++;
 854         if (progress_data->progress_func
 855             && (progress->write_streams.completed_bytes >= progress_data->next_progress
 856                 || new_wim_part))
 857         {
 858                 progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
 859                                              progress);
 860                 if (progress_data->next_progress == progress->write_streams.total_bytes) {
 861                         progress_data->next_progress = ~(uint64_t)0;
 862                 } else {
 863                         progress_data->next_progress =
 864                                 min(progress->write_streams.total_bytes,
 865                                     progress->write_streams.completed_bytes +
 866                                         progress->write_streams.total_bytes / 100);
 867                 }
 868         }
 869 }
 870
 871 struct serial_write_stream_ctx {
 872         struct filedes *out_fd;
 873         int out_ctype;
 874         u32 out_chunk_size;
 875         struct wimlib_lzx_context **comp_ctx;
 876         int write_resource_flags;
 877 };
 878
 879 static int
 880 serial_write_stream(struct wim_lookup_table_entry *lte, void *_ctx)
 881 {
 882         struct serial_write_stream_ctx *ctx = _ctx;
 883         return write_wim_resource(lte, ctx->out_fd,
 884                                   ctx->out_ctype,
 885                                   ctx->out_chunk_size,
 886                                   &lte->out_reshdr,
 887                                   ctx->write_resource_flags,
 888                                   ctx->comp_ctx);
 889 }
 890
 891
 892 /* Write a list of streams, taking into account that some streams may be
 893  * duplicates that are checksummed and discarded on the fly, and also delegating
 894  * the actual writing of a stream to a function @write_stream_cb, which is
 895  * passed the context @write_stream_ctx. */
 896 static int
 897 do_write_stream_list(struct list_head *stream_list,
 898                      struct wim_lookup_table *lookup_table,
 899                      int (*write_stream_cb)(struct wim_lookup_table_entry *, void *),
 900                      void *write_stream_ctx,
 901                      struct write_streams_progress_data *progress_data)
 902 {
 903         int ret = 0;
 904         struct wim_lookup_table_entry *lte;
 905         bool stream_discarded;
 906
 907         /* For each stream in @stream_list ... */
 908         while (!list_empty(stream_list)) {
 909                 stream_discarded = false;
 910                 lte = container_of(stream_list->next,
 911                                    struct wim_lookup_table_entry,
 912                                    write_streams_list);
 913                 list_del(&lte->write_streams_list);
 914                 if (lte->unhashed && !lte->unique_size) {
 915                         /* Unhashed stream that shares a size with some other
 916                          * stream in the WIM we are writing.  The stream must be
 917                          * checksummed to know if we need to write it or not. */
 918                         struct wim_lookup_table_entry *tmp;
 919                         u32 orig_out_refcnt = lte->out_refcnt;
 920
 921                         ret = hash_unhashed_stream(lte, lookup_table, &tmp);
 922                         if (ret)
 923                                 break;
 924                         if (tmp != lte) {
 925                                 /* We found a duplicate stream.  'lte' was
 926                                  * freed, so replace it with the duplicate.  */
 927                                 lte = tmp;
 928
 929                                 /* 'out_refcnt' was transferred to the
 930                                  * duplicate, and we can detect if the duplicate
 931                                  * stream was already referenced for writing by
 932                                  * checking if its 'out_refcnt' is higher than
 933                                  * that of the original stream.  In such cases,
 934                                  * the current stream can be discarded.  We can
 935                                  * also discard the current stream if it was
 936                                  * previously marked as filtered (e.g. already
 937                                  * present in the WIM being written).  */
 938                                 if (lte->out_refcnt > orig_out_refcnt ||
 939                                     lte->filtered) {
 940                                         DEBUG("Discarding duplicate stream of "
 941                                               "length %"PRIu64,
 942                                               lte->size);
 943                                         lte->no_progress = 0;
 944                                         stream_discarded = true;
 945                                         goto skip_to_progress;
 946                                 }
 947                         }
 948                 }
 949
 950                 /* Here, @lte is either a hashed stream or an unhashed stream
 951                  * with a unique size.  In either case we know that the stream
 952                  * has to be written.  In either case the SHA1 message digest
 953                  * will be calculated over the stream while writing it; however,
 954                  * in the former case this is done merely to check the data,
 955                  * while in the latter case this is done because we do not have
 956                  * the SHA1 message digest yet.  */
 957                 wimlib_assert(lte->out_refcnt != 0);
 958                 lte->deferred = 0;
 959                 lte->no_progress = 0;
 960                 ret = (*write_stream_cb)(lte, write_stream_ctx);
 961                 if (ret)
 962                         break;
 963                 /* In parallel mode, some streams are deferred for later,
 964                  * serialized processing; ignore them here.  */
 965                 if (lte->deferred)
 966                         continue;
 967                 if (lte->unhashed) {
 968                         list_del(&lte->unhashed_list);
 969                         lookup_table_insert(lookup_table, lte);
 970                         lte->unhashed = 0;
 971                 }
 972         skip_to_progress:
 973                 if (!lte->no_progress) {
 974                         do_write_streams_progress(progress_data,
 975                                                   lte, stream_discarded);
 976                 }
 977         }
 978         return ret;
 979 }
 980
 981 static int
 982 do_write_stream_list_serial(struct list_head *stream_list,
 983                             struct wim_lookup_table *lookup_table,
 984                             struct filedes *out_fd,
 985                             int out_ctype,
 986                             u32 out_chunk_size,
 987                             struct wimlib_lzx_context **comp_ctx,
 988                             int write_resource_flags,
 989                             struct write_streams_progress_data *progress_data)
 990 {
 991         struct serial_write_stream_ctx ctx = {
 992                 .out_fd = out_fd,
 993                 .out_ctype = out_ctype,
 994                 .out_chunk_size = out_chunk_size,
 995                 .write_resource_flags = write_resource_flags,
 996                 .comp_ctx = comp_ctx,
 997         };
 998         return do_write_stream_list(stream_list,
 999                                     lookup_table,
1000                                     serial_write_stream,
1001                                     &ctx,
1002                                     progress_data);
1003 }
1004
1005 static inline int
1006 write_flags_to_resource_flags(int write_flags)
1007 {
1008         int resource_flags = 0;
1009
1010         if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
1011                 resource_flags |= WIMLIB_WRITE_RESOURCE_FLAG_RECOMPRESS;
1012         if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
1013                 resource_flags |= WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE;
1014         return resource_flags;
1015 }
1016
1017 static int
1018 write_stream_list_serial(struct list_head *stream_list,
1019                          struct wim_lookup_table *lookup_table,
1020                          struct filedes *out_fd,
1021                          int out_ctype,
1022                          u32 out_chunk_size,
1023                          struct wimlib_lzx_context **comp_ctx,
1024                          int write_resource_flags,
1025                          struct write_streams_progress_data *progress_data)
1026 {
1027         union wimlib_progress_info *progress = &progress_data->progress;
1028         DEBUG("Writing stream list of size %"PRIu64" (serial version)",
1029               progress->write_streams.total_streams);
1030         progress->write_streams.num_threads = 1;
1031         if (progress_data->progress_func) {
1032                 progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
1033                                              progress);
1034         }
1035         return do_write_stream_list_serial(stream_list,
1036                                            lookup_table,
1037                                            out_fd,
1038                                            out_ctype,
1039                                            out_chunk_size,
1040                                            comp_ctx,
1041                                            write_resource_flags,
1042                                            progress_data);
1043 }
1044
1045 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1046 static int
1047 write_wim_chunks(struct message *msg, struct filedes *out_fd,
1048                  struct chunk_table *chunk_tab,
1049                  int write_resource_flags)
1050 {
1051         struct iovec *vecs;
1052         struct pwm_chunk_hdr *chunk_hdrs;
1053         unsigned nvecs;
1054         int ret;
1055
1056         for (unsigned i = 0; i < msg->num_chunks; i++)
1057                 chunk_tab_record_chunk(chunk_tab, msg->out_chunks[i].iov_len);
1058
1059         if (!(write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE)) {
1060                 nvecs = msg->num_chunks;
1061                 vecs = msg->out_chunks;
1062         } else {
1063                 /* Special case:  If writing a compressed resource to a pipable
1064                  * WIM, prefix each compressed chunk with a header that gives
1065                  * its compressed size.  */
1066                 nvecs = msg->num_chunks * 2;
1067                 vecs = alloca(nvecs * sizeof(vecs[0]));
1068                 chunk_hdrs = alloca(msg->num_chunks * sizeof(chunk_hdrs[0]));
1069
1070                 for (unsigned i = 0; i < msg->num_chunks; i++) {
1071                         chunk_hdrs[i].compressed_size = cpu_to_le32(msg->out_chunks[i].iov_len);
1072                         vecs[i * 2].iov_base = &chunk_hdrs[i];
1073                         vecs[i * 2].iov_len = sizeof(chunk_hdrs[i]);
1074                         vecs[i * 2 + 1].iov_base = msg->out_chunks[i].iov_base;
1075                         vecs[i * 2 + 1].iov_len = msg->out_chunks[i].iov_len;
1076                 }
1077         }
1078         ret = full_writev(out_fd, vecs, nvecs);
1079         if (ret)
1080                 ERROR_WITH_ERRNO("Write error");
1081         return ret;
1082 }
1083
1084 struct main_writer_thread_ctx {
1085         struct list_head *stream_list;
1086         struct wim_lookup_table *lookup_table;
1087         struct filedes *out_fd;
1088         off_t res_start_offset;
1089         int out_ctype;
1090         u32 out_chunk_size;
1091         struct wimlib_lzx_context **comp_ctx;
1092         int write_resource_flags;
1093         struct shared_queue *res_to_compress_queue;
1094         struct shared_queue *compressed_res_queue;
1095         size_t num_messages;
1096         struct write_streams_progress_data *progress_data;
1097
1098         struct list_head available_msgs;
1099         struct list_head outstanding_streams;
1100         struct list_head serial_streams;
1101         size_t num_outstanding_messages;
1102
1103         SHA_CTX next_sha_ctx;
1104         u64 next_chunk;
1105         u64 next_num_chunks;
1106         struct wim_lookup_table_entry *next_lte;
1107
1108         struct message *msgs;
1109         struct message *next_msg;
1110         struct chunk_table *cur_chunk_tab;
1111 };
1112
1113 static int
1114 init_message(struct message *msg, u32 out_chunk_size)
1115 {
1116         msg->out_chunk_size = out_chunk_size;
1117         for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
1118                 msg->compressed_chunks[i] = MALLOC(out_chunk_size);
1119                 msg->uncompressed_chunks[i] = MALLOC(out_chunk_size);
1120                 if (msg->compressed_chunks[i] == NULL ||
1121                     msg->uncompressed_chunks[i] == NULL)
1122                         return WIMLIB_ERR_NOMEM;
1123         }
1124         return 0;
1125 }
1126
1127 static void
1128 destroy_message(struct message *msg)
1129 {
1130         for (size_t i = 0; i < MAX_CHUNKS_PER_MSG; i++) {
1131                 FREE(msg->compressed_chunks[i]);
1132                 FREE(msg->uncompressed_chunks[i]);
1133         }
1134 }
1135
1136 static void
1137 free_messages(struct message *msgs, size_t num_messages)
1138 {
1139         if (msgs) {
1140                 for (size_t i = 0; i < num_messages; i++)
1141                         destroy_message(&msgs[i]);
1142                 FREE(msgs);
1143         }
1144 }
1145
1146 static struct message *
1147 allocate_messages(size_t num_messages, u32 out_chunk_size)
1148 {
1149         struct message *msgs;
1150
1151         msgs = CALLOC(num_messages, sizeof(struct message));
1152         if (msgs == NULL)
1153                 return NULL;
1154         for (size_t i = 0; i < num_messages; i++) {
1155                 if (init_message(&msgs[i], out_chunk_size)) {
1156                         free_messages(msgs, num_messages);
1157                         return NULL;
1158                 }
1159         }
1160         return msgs;
1161 }
1162
1163 static void
1164 main_writer_thread_destroy_ctx(struct main_writer_thread_ctx *ctx)
1165 {
1166         while (ctx->num_outstanding_messages--)
1167                 shared_queue_get(ctx->compressed_res_queue);
1168         free_messages(ctx->msgs, ctx->num_messages);
1169         FREE(ctx->cur_chunk_tab);
1170 }
1171
1172 static int
1173 main_writer_thread_init_ctx(struct main_writer_thread_ctx *ctx)
1174 {
1175         /* Pre-allocate all the buffers that will be needed to do the chunk
1176          * compression. */
1177         ctx->msgs = allocate_messages(ctx->num_messages, ctx->out_chunk_size);
1178         if (ctx->msgs == NULL)
1179                 return WIMLIB_ERR_NOMEM;
1180
1181         /* Initially, all the messages are available to use. */
1182         INIT_LIST_HEAD(&ctx->available_msgs);
1183         for (size_t i = 0; i < ctx->num_messages; i++)
1184                 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
1185
1186         /* outstanding_streams is the list of streams that currently have had
1187          * chunks sent off for compression.
1188          *
1189          * The first stream in outstanding_streams is the stream that is
1190          * currently being written.
1191          *
1192          * The last stream in outstanding_streams is the stream that is
1193          * currently being read and having chunks fed to the compressor threads.
1194          * */
1195         INIT_LIST_HEAD(&ctx->outstanding_streams);
1196         ctx->num_outstanding_messages = 0;
1197
1198         /* Message currently being prepared.  */
1199         ctx->next_msg = NULL;
1200
1201         /* Resources that don't need any chunks compressed are added to this
1202          * list and written directly by the main thread.  */
1203         INIT_LIST_HEAD(&ctx->serial_streams);
1204
1205         /* Pointer to chunk table for stream currently being written.  */
1206         ctx->cur_chunk_tab = NULL;
1207
1208         return 0;
1209 }
1210
1211 static int
1212 receive_compressed_chunks(struct main_writer_thread_ctx *ctx)
1213 {
1214         struct message *msg;
1215         struct wim_lookup_table_entry *cur_lte;
1216         int ret;
1217
1218         wimlib_assert(!list_empty(&ctx->outstanding_streams));
1219         wimlib_assert(ctx->num_outstanding_messages != 0);
1220
1221         cur_lte = container_of(ctx->outstanding_streams.next,
1222                                struct wim_lookup_table_entry,
1223                                being_compressed_list);
1224
1225         /* Get the next message from the queue and process it.
1226          * The message will contain 1 or more data chunks that have been
1227          * compressed.  */
1228         msg = shared_queue_get(ctx->compressed_res_queue);
1229         msg->complete = true;
1230         --ctx->num_outstanding_messages;
1231
1232         /* Is this the next chunk in the current resource?  If it's not (i.e.,
1233          * an earlier chunk in a same or different resource hasn't been
1234          * compressed yet), do nothing, and keep this message around until all
1235          * earlier chunks are received.
1236          *
1237          * Otherwise, write all the chunks we can.  */
1238         while (cur_lte != NULL &&
1239                !list_empty(&cur_lte->msg_list)
1240                && (msg = container_of(cur_lte->msg_list.next,
1241                                       struct message,
1242                                       list))->complete)
1243         {
1244                 list_move(&msg->list, &ctx->available_msgs);
1245                 if (msg->begin_chunk == 0) {
1246                         /* First set of chunks.  */
1247
1248                         /* Write pipable WIM stream header if needed.  */
1249                         if (ctx->write_resource_flags &
1250                             WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE)
1251                         {
1252                                 ret = write_pwm_stream_header(cur_lte, ctx->out_fd,
1253                                                               WIM_RESHDR_FLAG_COMPRESSED);
1254                                 if (ret)
1255                                         return ret;
1256                         }
1257
1258                         /* Save current offset.  */
1259                         ctx->res_start_offset = ctx->out_fd->offset;
1260
1261                         /* Begin building the chunk table, and leave space for
1262                          * it if needed.  */
1263                         ret = begin_wim_resource_chunk_tab(cur_lte,
1264                                                            ctx->out_fd,
1265                                                            ctx->out_chunk_size,
1266                                                            &ctx->cur_chunk_tab,
1267                                                            ctx->write_resource_flags);
1268                         if (ret)
1269                                 return ret;
1270                 }
1271
1272                 /* Write the compressed chunks from the message.  */
1273                 ret = write_wim_chunks(msg, ctx->out_fd, ctx->cur_chunk_tab,
1274                                        ctx->write_resource_flags);
1275                 if (ret)
1276                         return ret;
1277
1278                 /* Was this the last chunk of the stream?  If so, finish the
1279                  * stream by writing the chunk table.  */
1280                 if (list_empty(&cur_lte->msg_list) &&
1281                     msg->begin_chunk + msg->num_chunks == ctx->cur_chunk_tab->num_chunks)
1282                 {
1283                         u64 res_csize;
1284
1285                         ret = finish_wim_resource_chunk_tab(ctx->cur_chunk_tab,
1286                                                             ctx->out_fd,
1287                                                             ctx->res_start_offset,
1288                                                             ctx->write_resource_flags);
1289                         if (ret)
1290                                 return ret;
1291
1292                         list_del(&cur_lte->being_compressed_list);
1293
1294                         res_csize = ctx->out_fd->offset - ctx->res_start_offset;
1295
1296                         FREE(ctx->cur_chunk_tab);
1297                         ctx->cur_chunk_tab = NULL;
1298
1299                         /* Check for resources compressed to greater than or
1300                          * equal to their original size and write them
1301                          * uncompressed instead.  (But never do this if writing
1302                          * to a pipe.)  */
1303                         if (res_csize >= cur_lte->size &&
1304                             !(ctx->write_resource_flags & WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE))
1305                         {
1306                                 DEBUG("Compressed %"PRIu64" => %"PRIu64" bytes; "
1307                                       "writing uncompressed instead",
1308                                       cur_lte->size, res_csize);
1309                                 ret = seek_and_truncate(ctx->out_fd, ctx->res_start_offset);
1310                                 if (ret)
1311                                         return ret;
1312                                 ret = write_wim_resource(cur_lte,
1313                                                          ctx->out_fd,
1314                                                          WIMLIB_COMPRESSION_TYPE_NONE,
1315                                                          0,
1316                                                          &cur_lte->out_reshdr,
1317                                                          ctx->write_resource_flags,
1318                                                          ctx->comp_ctx);
1319                                 if (ret)
1320                                         return ret;
1321                         } else {
1322                                 cur_lte->out_reshdr.size_in_wim =
1323                                         res_csize;
1324
1325                                 cur_lte->out_reshdr.uncompressed_size =
1326                                         cur_lte->size;
1327
1328                                 cur_lte->out_reshdr.offset_in_wim =
1329                                         ctx->res_start_offset;
1330
1331                                 cur_lte->out_reshdr.flags =
1332                                         cur_lte->flags |
1333                                                 WIM_RESHDR_FLAG_COMPRESSED;
1334
1335                                 DEBUG("Wrote compressed resource "
1336                                       "(%"PRIu64" => %"PRIu64" bytes @ +%"PRIu64", flags=0x%02x)",
1337                                       cur_lte->out_reshdr.uncompressed_size,
1338                                       cur_lte->out_reshdr.size_in_wim,
1339                                       cur_lte->out_reshdr.offset_in_wim,
1340                                       cur_lte->out_reshdr.flags);
1341                         }
1342
1343                         do_write_streams_progress(ctx->progress_data,
1344                                                   cur_lte, false);
1345
1346                         /* Since we just finished writing a stream, write any
1347                          * streams that have been added to the serial_streams
1348                          * list for direct writing by the main thread (e.g.
1349                          * resources that don't need to be compressed because
1350                          * the desired compression type is the same as the
1351                          * previous compression type).  */
1352                         if (!list_empty(&ctx->serial_streams)) {
1353                                 ret = do_write_stream_list_serial(&ctx->serial_streams,
1354                                                                   ctx->lookup_table,
1355                                                                   ctx->out_fd,
1356                                                                   ctx->out_ctype,
1357                                                                   ctx->out_chunk_size,
1358                                                                   ctx->comp_ctx,
1359                                                                   ctx->write_resource_flags,
1360                                                                   ctx->progress_data);
1361                                 if (ret)
1362                                         return ret;
1363                         }
1364
1365                         /* Advance to the next stream to write.  */
1366                         if (list_empty(&ctx->outstanding_streams)) {
1367                                 cur_lte = NULL;
1368                         } else {
1369                                 cur_lte = container_of(ctx->outstanding_streams.next,
1370                                                        struct wim_lookup_table_entry,
1371                                                        being_compressed_list);
1372                         }
1373                 }
1374         }
1375         return 0;
1376 }
1377
1378 /* Called when the main thread has read a new chunk of data.  */
1379 static int
1380 main_writer_thread_cb(const void *chunk, size_t chunk_size, void *_ctx)
1381 {
1382         struct main_writer_thread_ctx *ctx = _ctx;
1383         int ret;
1384         struct message *next_msg;
1385         u64 next_chunk_in_msg;
1386
1387         /* Update SHA1 message digest for the stream currently being read by the
1388          * main thread. */
1389         sha1_update(&ctx->next_sha_ctx, chunk, chunk_size);
1390
1391         /* We send chunks of data to the compressor chunks in batches which we
1392          * refer to as "messages".  @next_msg is the message that is currently
1393          * being prepared to send off.  If it is NULL, that indicates that we
1394          * need to start a new message. */
1395         next_msg = ctx->next_msg;
1396         if (!next_msg) {
1397                 /* We need to start a new message.  First check to see if there
1398                  * is a message available in the list of available messages.  If
1399                  * so, we can just take one.  If not, all the messages (there is
1400                  * a fixed number of them, proportional to the number of
1401                  * threads) have been sent off to the compressor threads, so we
1402                  * receive messages from the compressor threads containing
1403                  * compressed chunks of data.
1404                  *
1405                  * We may need to receive multiple messages before one is
1406                  * actually available to use because messages received that are
1407                  * *not* for the very next set of chunks to compress must be
1408                  * buffered until it's time to write those chunks. */
1409                 while (list_empty(&ctx->available_msgs)) {
1410                         ret = receive_compressed_chunks(ctx);
1411                         if (ret)
1412                                 return ret;
1413                 }
1414
1415                 next_msg = container_of(ctx->available_msgs.next,
1416                                         struct message, list);
1417                 list_del(&next_msg->list);
1418                 next_msg->complete = false;
1419                 next_msg->begin_chunk = ctx->next_chunk;
1420                 next_msg->num_chunks = min(MAX_CHUNKS_PER_MSG,
1421                                            ctx->next_num_chunks - ctx->next_chunk);
1422                 ctx->next_msg = next_msg;
1423         }
1424
1425         /* Fill in the next chunk to compress */
1426         next_chunk_in_msg = ctx->next_chunk - next_msg->begin_chunk;
1427
1428         next_msg->uncompressed_chunk_sizes[next_chunk_in_msg] = chunk_size;
1429         memcpy(next_msg->uncompressed_chunks[next_chunk_in_msg],
1430                chunk, chunk_size);
1431         ctx->next_chunk++;
1432         if (++next_chunk_in_msg == next_msg->num_chunks) {
1433                 /* Send off an array of chunks to compress */
1434                 list_add_tail(&next_msg->list, &ctx->next_lte->msg_list);
1435                 shared_queue_put(ctx->res_to_compress_queue, next_msg);
1436                 ++ctx->num_outstanding_messages;
1437                 ctx->next_msg = NULL;
1438         }
1439         return 0;
1440 }
1441
1442 static int
1443 main_writer_thread_finish(void *_ctx)
1444 {
1445         struct main_writer_thread_ctx *ctx = _ctx;
1446         int ret;
1447         while (ctx->num_outstanding_messages != 0) {
1448                 ret = receive_compressed_chunks(ctx);
1449                 if (ret)
1450                         return ret;
1451         }
1452         wimlib_assert(list_empty(&ctx->outstanding_streams));
1453         return do_write_stream_list_serial(&ctx->serial_streams,
1454                                            ctx->lookup_table,
1455                                            ctx->out_fd,
1456                                            ctx->out_ctype,
1457                                            ctx->out_chunk_size,
1458                                            ctx->comp_ctx,
1459                                            ctx->write_resource_flags,
1460                                            ctx->progress_data);
1461 }
1462
1463 static int
1464 submit_stream_for_compression(struct wim_lookup_table_entry *lte,
1465                               struct main_writer_thread_ctx *ctx)
1466 {
1467         int ret;
1468
1469         /* Read the entire stream @lte, feeding its data chunks to the
1470          * compressor threads.  Also SHA1-sum the stream; this is required in
1471          * the case that @lte is unhashed, and a nice additional verification
1472          * when @lte is already hashed.  */
1473         sha1_init(&ctx->next_sha_ctx);
1474         ctx->next_chunk = 0;
1475         ctx->next_num_chunks = DIV_ROUND_UP(lte->size, ctx->out_chunk_size);
1476         ctx->next_lte = lte;
1477         INIT_LIST_HEAD(&lte->msg_list);
1478         list_add_tail(&lte->being_compressed_list, &ctx->outstanding_streams);
1479         ret = read_resource_prefix(lte, lte->size,
1480                                    main_writer_thread_cb,
1481                                    ctx->out_chunk_size, ctx, 0);
1482         if (ret)
1483                 return ret;
1484         wimlib_assert(ctx->next_chunk == ctx->next_num_chunks);
1485         return finalize_and_check_sha1(&ctx->next_sha_ctx, lte);
1486 }
1487
1488 static int
1489 main_thread_process_next_stream(struct wim_lookup_table_entry *lte, void *_ctx)
1490 {
1491         struct main_writer_thread_ctx *ctx = _ctx;
1492         int ret;
1493
1494         if (lte->size < 1000 ||
1495             !must_compress_stream(lte, ctx->write_resource_flags,
1496                                   ctx->out_ctype, ctx->out_chunk_size))
1497         {
1498                 /* Stream is too small or isn't being compressed.  Process it by
1499                  * the main thread when we have a chance.  We can't necessarily
1500                  * process it right here, as the main thread could be in the
1501                  * middle of writing a different stream.  */
1502                 list_add_tail(&lte->write_streams_list, &ctx->serial_streams);
1503                 lte->deferred = 1;
1504                 ret = 0;
1505         } else {
1506                 ret = submit_stream_for_compression(lte, ctx);
1507         }
1508         lte->no_progress = 1;
1509         return ret;
1510 }
1511
1512 static long
1513 get_default_num_threads(void)
1514 {
1515 #ifdef __WIN32__
1516         return win32_get_number_of_processors();
1517 #else
1518         return sysconf(_SC_NPROCESSORS_ONLN);
1519 #endif
1520 }
1521
1522 /* Equivalent to write_stream_list_serial(), except this takes a @num_threads
1523  * parameter and will perform compression using that many threads.  Falls
1524  * back to write_stream_list_serial() on certain errors, such as a failure to
1525  * create the number of threads requested.
1526  *
1527  * High level description of the algorithm for writing compressed streams in
1528  * parallel:  We perform compression on chunks rather than on full files.  The
1529  * currently executing thread becomes the main thread and is entirely in charge
1530  * of reading the data to compress (which may be in any location understood by
1531  * the resource code--- such as in an external file being captured, or in
1532  * another WIM file from which an image is being exported) and actually writing
1533  * the compressed data to the output file.  Additional threads are "compressor
1534  * threads" and all execute the compressor_thread_proc, where they repeatedly
1535  * retrieve buffers of data from the main thread, compress them, and hand them
1536  * back to the main thread.
1537  *
1538  * Certain streams, such as streams that do not need to be compressed (e.g.
1539  * input compression type same as output compression type) or streams of very
1540  * small size are placed in a list (main_writer_thread_ctx.serial_list) and
1541  * handled entirely by the main thread at an appropriate time.
1542  *
1543  * At any given point in time, multiple streams may be having chunks compressed
1544  * concurrently.  The stream that the main thread is currently *reading* may be
1545  * later in the list that the stream that the main thread is currently
1546  * *writing*.  */
1547 static int
1548 write_stream_list_parallel(struct list_head *stream_list,
1549                            struct wim_lookup_table *lookup_table,
1550                            struct filedes *out_fd,
1551                            int out_ctype,
1552                            u32 out_chunk_size,
1553                            struct wimlib_lzx_context **comp_ctx,
1554                            int write_resource_flags,
1555                            struct write_streams_progress_data *progress_data,
1556                            unsigned num_threads)
1557 {
1558         int ret;
1559         struct shared_queue res_to_compress_queue;
1560         struct shared_queue compressed_res_queue;
1561         pthread_t *compressor_threads = NULL;
1562         union wimlib_progress_info *progress = &progress_data->progress;
1563         unsigned num_started_threads;
1564         bool can_retry = true;
1565
1566         if (num_threads == 0) {
1567                 long nthreads = get_default_num_threads();
1568                 if (nthreads < 1 || nthreads > UINT_MAX) {
1569                         WARNING("Could not determine number of processors! Assuming 1");
1570                         goto out_serial_quiet;
1571                 } else if (nthreads == 1) {
1572                         goto out_serial_quiet;
1573                 } else {
1574                         num_threads = nthreads;
1575                 }
1576         }
1577
1578         DEBUG("Writing stream list of size %"PRIu64" "
1579               "(parallel version, num_threads=%u)",
1580               progress->write_streams.total_streams, num_threads);
1581
1582         progress->write_streams.num_threads = num_threads;
1583
1584         static const size_t MESSAGES_PER_THREAD = 2;
1585         size_t queue_size = (size_t)(num_threads * MESSAGES_PER_THREAD);
1586
1587         DEBUG("Initializing shared queues (queue_size=%zu)", queue_size);
1588
1589         ret = shared_queue_init(&res_to_compress_queue, queue_size);
1590         if (ret)
1591                 goto out_serial;
1592
1593         ret = shared_queue_init(&compressed_res_queue, queue_size);
1594         if (ret)
1595                 goto out_destroy_res_to_compress_queue;
1596
1597         struct compressor_thread_params *params;
1598
1599         params = CALLOC(num_threads, sizeof(params[0]));
1600         if (params == NULL) {
1601                 ret = WIMLIB_ERR_NOMEM;
1602                 goto out_destroy_compressed_res_queue;
1603         }
1604
1605         for (unsigned i = 0; i < num_threads; i++) {
1606                 params[i].res_to_compress_queue = &res_to_compress_queue;
1607                 params[i].compressed_res_queue = &compressed_res_queue;
1608                 params[i].out_ctype = out_ctype;
1609                 if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX) {
1610                         ret = wimlib_lzx_alloc_context(out_chunk_size,
1611                                                        NULL, &params[i].comp_ctx);
1612                         if (ret)
1613                                 goto out_free_params;
1614                 }
1615         }
1616
1617         compressor_threads = MALLOC(num_threads * sizeof(pthread_t));
1618         if (compressor_threads == NULL) {
1619                 ret = WIMLIB_ERR_NOMEM;
1620                 goto out_free_params;
1621         }
1622
1623         for (unsigned i = 0; i < num_threads; i++) {
1624                 DEBUG("pthread_create thread %u of %u", i + 1, num_threads);
1625                 ret = pthread_create(&compressor_threads[i], NULL,
1626                                      compressor_thread_proc, &params[i]);
1627                 if (ret) {
1628                         errno = ret;
1629                         ret = -1;
1630                         ERROR_WITH_ERRNO("Failed to create compressor "
1631                                          "thread %u of %u",
1632                                          i + 1, num_threads);
1633                         num_started_threads = i;
1634                         goto out_join;
1635                 }
1636         }
1637         num_started_threads = num_threads;
1638
1639         if (progress_data->progress_func) {
1640                 progress_data->progress_func(WIMLIB_PROGRESS_MSG_WRITE_STREAMS,
1641                                              progress);
1642         }
1643
1644         struct main_writer_thread_ctx ctx;
1645
1646         memset(&ctx, 0, sizeof(ctx));
1647
1648         ctx.stream_list           = stream_list;
1649         ctx.lookup_table          = lookup_table;
1650         ctx.out_fd                = out_fd;
1651         ctx.out_ctype             = out_ctype;
1652         ctx.out_chunk_size        = out_chunk_size;
1653         ctx.comp_ctx              = comp_ctx;
1654         ctx.res_to_compress_queue = &res_to_compress_queue;
1655         ctx.compressed_res_queue  = &compressed_res_queue;
1656         ctx.num_messages          = queue_size;
1657         ctx.write_resource_flags  = write_resource_flags;
1658         ctx.progress_data         = progress_data;
1659         ret = main_writer_thread_init_ctx(&ctx);
1660         if (ret)
1661                 goto out_join;
1662
1663         can_retry = false;
1664         ret = do_write_stream_list(stream_list, lookup_table,
1665                                    main_thread_process_next_stream,
1666                                    &ctx, progress_data);
1667         if (ret)
1668                 goto out_destroy_ctx;
1669
1670         /* The main thread has finished reading all streams that are going to be
1671          * compressed in parallel, and it now needs to wait for all remaining
1672          * chunks to be compressed so that the remaining streams can actually be
1673          * written to the output file.  Furthermore, any remaining streams that
1674          * had processing deferred to the main thread need to be handled.  These
1675          * tasks are done by the main_writer_thread_finish() function.  */
1676         ret = main_writer_thread_finish(&ctx);
1677 out_destroy_ctx:
1678         main_writer_thread_destroy_ctx(&ctx);
1679 out_join:
1680         for (unsigned i = 0; i < num_started_threads; i++)
1681                 shared_queue_put(&res_to_compress_queue, NULL);
1682
1683         for (unsigned i = 0; i < num_started_threads; i++) {
1684                 if (pthread_join(compressor_threads[i], NULL)) {
1685                         WARNING_WITH_ERRNO("Failed to join compressor "
1686                                            "thread %u of %u",
1687                                            i + 1, num_threads);
1688                 }
1689         }
1690         FREE(compressor_threads);
1691 out_free_params:
1692         for (unsigned i = 0; i < num_threads; i++)
1693                 wimlib_lzx_free_context(params[i].comp_ctx);
1694         FREE(params);
1695 out_destroy_compressed_res_queue:
1696         shared_queue_destroy(&compressed_res_queue);
1697 out_destroy_res_to_compress_queue:
1698         shared_queue_destroy(&res_to_compress_queue);
1699         if (!can_retry || (ret >= 0 && ret != WIMLIB_ERR_NOMEM))
1700                 return ret;
1701 out_serial:
1702         WARNING("Falling back to single-threaded compression");
1703 out_serial_quiet:
1704         return write_stream_list_serial(stream_list,
1705                                         lookup_table,
1706                                         out_fd,
1707                                         out_ctype,
1708                                         out_chunk_size,
1709                                         comp_ctx,
1710                                         write_resource_flags,
1711                                         progress_data);
1712
1713 }
1714 #endif
1715
1716 /* Write a list of streams to a WIM (@out_fd) using the compression type
1717  * @out_ctype, chunk size @out_chunk_size, and up to @num_threads compressor
1718  * threads.  */
1719 static int
1720 write_stream_list(struct list_head *stream_list,
1721                   struct wim_lookup_table *lookup_table,
1722                   struct filedes *out_fd, int out_ctype,
1723                   u32 out_chunk_size,
1724                   struct wimlib_lzx_context **comp_ctx,
1725                   int write_flags,
1726                   unsigned num_threads, wimlib_progress_func_t progress_func)
1727 {
1728         int ret;
1729         int write_resource_flags;
1730         u64 total_bytes;
1731         u64 total_compression_bytes;
1732         unsigned total_parts;
1733         WIMStruct *prev_wim_part;
1734         size_t num_streams;
1735         struct wim_lookup_table_entry *lte;
1736         struct write_streams_progress_data progress_data;
1737
1738         if (list_empty(stream_list)) {
1739                 DEBUG("No streams to write.");
1740                 return 0;
1741         }
1742
1743         write_resource_flags = write_flags_to_resource_flags(write_flags);
1744
1745         DEBUG("Writing stream list (offset = %"PRIu64", write_resource_flags=0x%08x)",
1746               out_fd->offset, write_resource_flags);
1747
1748         /* Sort the stream list into a good order for reading.  */
1749         ret = sort_stream_list_by_sequential_order(stream_list,
1750                                                    offsetof(struct wim_lookup_table_entry,
1751                                                             write_streams_list));
1752         if (ret)
1753                 return ret;
1754
1755         /* Calculate the total size of the streams to be written.  Note: this
1756          * will be the uncompressed size, as we may not know the compressed size
1757          * yet, and also this will assume that every unhashed stream will be
1758          * written (which will not necessarily be the case).  */
1759         total_bytes = 0;
1760         total_compression_bytes = 0;
1761         num_streams = 0;
1762         total_parts = 0;
1763         prev_wim_part = NULL;
1764         list_for_each_entry(lte, stream_list, write_streams_list) {
1765                 num_streams++;
1766                 total_bytes += lte->size;
1767                 if (must_compress_stream(lte, write_resource_flags,
1768                                          out_ctype, out_chunk_size))
1769                         total_compression_bytes += lte->size;
1770                 if (lte->resource_location == RESOURCE_IN_WIM) {
1771                         if (prev_wim_part != lte->rspec->wim) {
1772                                 prev_wim_part = lte->rspec->wim;
1773                                 total_parts++;
1774                         }
1775                 }
1776         }
1777
1778         memset(&progress_data, 0, sizeof(progress_data));
1779         progress_data.progress_func = progress_func;
1780
1781         progress_data.progress.write_streams.total_bytes       = total_bytes;
1782         progress_data.progress.write_streams.total_streams     = num_streams;
1783         progress_data.progress.write_streams.completed_bytes   = 0;
1784         progress_data.progress.write_streams.completed_streams = 0;
1785         progress_data.progress.write_streams.num_threads       = num_threads;
1786         progress_data.progress.write_streams.compression_type  = out_ctype;
1787         progress_data.progress.write_streams.total_parts       = total_parts;
1788         progress_data.progress.write_streams.completed_parts   = 0;
1789
1790         progress_data.next_progress = 0;
1791         progress_data.prev_wim_part = NULL;
1792
1793 #ifdef ENABLE_MULTITHREADED_COMPRESSION
1794         if (total_compression_bytes >= 2000000 && num_threads != 1)
1795                 ret = write_stream_list_parallel(stream_list,
1796                                                  lookup_table,
1797                                                  out_fd,
1798                                                  out_ctype,
1799                                                  out_chunk_size,
1800                                                  comp_ctx,
1801                                                  write_resource_flags,
1802                                                  &progress_data,
1803                                                  num_threads);
1804         else
1805 #endif
1806                 ret = write_stream_list_serial(stream_list,
1807                                                lookup_table,
1808                                                out_fd,
1809                                                out_ctype,
1810                                                out_chunk_size,
1811                                                comp_ctx,
1812                                                write_resource_flags,
1813                                                &progress_data);
1814         if (ret == 0)
1815                 DEBUG("Successfully wrote stream list.");
1816         else
1817                 DEBUG("Failed to write stream list (ret=%d).", ret);
1818         return ret;
1819 }
1820
1821 struct stream_size_table {
1822         struct hlist_head *array;
1823         size_t num_entries;
1824         size_t capacity;
1825 };
1826
1827 static int
1828 init_stream_size_table(struct stream_size_table *tab, size_t capacity)
1829 {
1830         tab->array = CALLOC(capacity, sizeof(tab->array[0]));
1831         if (!tab->array)
1832                 return WIMLIB_ERR_NOMEM;
1833         tab->num_entries = 0;
1834         tab->capacity = capacity;
1835         return 0;
1836 }
1837
1838 static void
1839 destroy_stream_size_table(struct stream_size_table *tab)
1840 {
1841         FREE(tab->array);
1842 }
1843
1844 static int
1845 stream_size_table_insert(struct wim_lookup_table_entry *lte, void *_tab)
1846 {
1847         struct stream_size_table *tab = _tab;
1848         size_t pos;
1849         struct wim_lookup_table_entry *same_size_lte;
1850         struct hlist_node *tmp;
1851
1852         pos = hash_u64(lte->size) % tab->capacity;
1853         lte->unique_size = 1;
1854         hlist_for_each_entry(same_size_lte, tmp, &tab->array[pos], hash_list_2) {
1855                 if (same_size_lte->size == lte->size) {
1856                         lte->unique_size = 0;
1857                         same_size_lte->unique_size = 0;
1858                         break;
1859                 }
1860         }
1861
1862         hlist_add_head(&lte->hash_list_2, &tab->array[pos]);
1863         tab->num_entries++;
1864         return 0;
1865 }
1866
1867 struct find_streams_ctx {
1868         WIMStruct *wim;
1869         int write_flags;
1870         struct list_head stream_list;
1871         struct stream_size_table stream_size_tab;
1872 };
1873
1874 static void
1875 lte_reference_for_logical_write(struct wim_lookup_table_entry *lte,
1876                                 struct find_streams_ctx *ctx,
1877                                 unsigned nref)
1878 {
1879         if (lte->out_refcnt == 0) {
1880                 stream_size_table_insert(lte, &ctx->stream_size_tab);
1881                 list_add_tail(&lte->write_streams_list, &ctx->stream_list);
1882         }
1883         lte->out_refcnt += nref;
1884 }
1885
1886 static int
1887 do_lte_full_reference_for_logical_write(struct wim_lookup_table_entry *lte,
1888                                         void *_ctx)
1889 {
1890         struct find_streams_ctx *ctx = _ctx;
1891         lte->out_refcnt = 0;
1892         lte_reference_for_logical_write(lte, ctx,
1893                                         (lte->refcnt ? lte->refcnt : 1));
1894         return 0;
1895 }
1896
1897 static int
1898 inode_find_streams_to_write(struct wim_inode *inode,
1899                             struct wim_lookup_table *table,
1900                             struct find_streams_ctx *ctx)
1901 {
1902         struct wim_lookup_table_entry *lte;
1903         unsigned i;
1904
1905         for (i = 0; i <= inode->i_num_ads; i++) {
1906                 lte = inode_stream_lte(inode, i, table);
1907                 if (lte)
1908                         lte_reference_for_logical_write(lte, ctx, inode->i_nlink);
1909                 else if (!is_zero_hash(inode_stream_hash(inode, i)))
1910                         return WIMLIB_ERR_RESOURCE_NOT_FOUND;
1911         }
1912         return 0;
1913 }
1914
1915 static int
1916 image_find_streams_to_write(WIMStruct *wim)
1917 {
1918         struct find_streams_ctx *ctx;
1919         struct wim_image_metadata *imd;
1920         struct wim_inode *inode;
1921         struct wim_lookup_table_entry *lte;
1922         int ret;
1923
1924         ctx = wim->private;
1925         imd = wim_get_current_image_metadata(wim);
1926
1927         image_for_each_unhashed_stream(lte, imd)
1928                 lte->out_refcnt = 0;
1929
1930         /* Go through this image's inodes to find any streams that have not been
1931          * found yet. */
1932         image_for_each_inode(inode, imd) {
1933                 ret = inode_find_streams_to_write(inode, wim->lookup_table, ctx);
1934                 if (ret)
1935                         return ret;
1936         }
1937         return 0;
1938 }
1939
1940 /*
1941  * Build a list of streams (via `struct wim_lookup_table_entry's) included in
1942  * the "logical write" of the WIM, meaning all streams that are referenced at
1943  * least once by dentries in the the image(s) being written.  'out_refcnt' on
1944  * each stream being included in the logical write is set to the number of
1945  * references from dentries in the image(s).  Furthermore, 'unique_size' on each
1946  * stream being included in the logical write is set to indicate whether that
1947  * stream has a unique size relative to the streams being included in the
1948  * logical write.  Still furthermore, 'part_number' on each stream being
1949  * included in the logical write is set to the part number given in the
1950  * in-memory header of @p wim.
1951  *
1952  * This is considered a "logical write" because it does not take into account
1953  * filtering out streams already present in the WIM (in the case of an in place
1954  * overwrite) or present in other WIMs (in case of creating delta WIM).
1955  */
1956 static int
1957 prepare_logical_stream_list(WIMStruct *wim, int image, bool streams_ok,
1958                             struct find_streams_ctx *ctx)
1959 {
1960         int ret;
1961
1962         if (streams_ok && (image == WIMLIB_ALL_IMAGES ||
1963                            (image == 1 && wim->hdr.image_count == 1)))
1964         {
1965                 /* Fast case:  Assume that all streams are being written and
1966                  * that the reference counts are correct.  */
1967                 struct wim_lookup_table_entry *lte;
1968                 struct wim_image_metadata *imd;
1969                 unsigned i;
1970
1971                 for_lookup_table_entry(wim->lookup_table,
1972                                        do_lte_full_reference_for_logical_write, ctx);
1973                 for (i = 0; i < wim->hdr.image_count; i++) {
1974                         imd = wim->image_metadata[i];
1975                         image_for_each_unhashed_stream(lte, imd)
1976                                 do_lte_full_reference_for_logical_write(lte, ctx);
1977                 }
1978         } else {
1979                 /* Slow case:  Walk through the images being written and
1980                  * determine the streams referenced.  */
1981                 for_lookup_table_entry(wim->lookup_table, lte_zero_out_refcnt, NULL);
1982                 wim->private = ctx;
1983                 ret = for_image(wim, image, image_find_streams_to_write);
1984                 if (ret)
1985                         return ret;
1986         }
1987
1988         return 0;
1989 }
1990
1991 static int
1992 process_filtered_stream(struct wim_lookup_table_entry *lte, void *_ctx)
1993 {
1994         struct find_streams_ctx *ctx = _ctx;
1995         u16 filtered = 0;
1996
1997         /* Calculate and set lte->filtered.  */
1998         if (lte->resource_location == RESOURCE_IN_WIM) {
1999                 if (lte->rspec->wim == ctx->wim &&
2000                     (ctx->write_flags & WIMLIB_WRITE_FLAG_OVERWRITE))
2001                         filtered |= FILTERED_SAME_WIM;
2002                 if (lte->rspec->wim != ctx->wim &&
2003                     (ctx->write_flags & WIMLIB_WRITE_FLAG_SKIP_EXTERNAL_WIMS))
2004                         filtered |= FILTERED_EXTERNAL_WIM;
2005         }
2006         lte->filtered = filtered;
2007
2008         /* Filtered streams get inserted into the stream size table too, unless
2009          * they already were.  This is because streams that are checksummed
2010          * on-the-fly during the write should not be written if they are
2011          * duplicates of filtered stream.  */
2012         if (lte->filtered && lte->out_refcnt == 0)
2013                 stream_size_table_insert(lte, &ctx->stream_size_tab);
2014         return 0;
2015 }
2016
2017 static int
2018 mark_stream_not_filtered(struct wim_lookup_table_entry *lte, void *_ignore)
2019 {
2020         lte->filtered = 0;
2021         return 0;
2022 }
2023
2024 /* Given the list of streams to include in a logical write of a WIM, handle
2025  * filtering out streams already present in the WIM or already present in
2026  * external WIMs, depending on the write flags provided.  */
2027 static void
2028 handle_stream_filtering(struct find_streams_ctx *ctx)
2029 {
2030         struct wim_lookup_table_entry *lte, *tmp;
2031
2032         if (!(ctx->write_flags & (WIMLIB_WRITE_FLAG_OVERWRITE |
2033                                   WIMLIB_WRITE_FLAG_SKIP_EXTERNAL_WIMS)))
2034         {
2035                 for_lookup_table_entry(ctx->wim->lookup_table,
2036                                        mark_stream_not_filtered, ctx);
2037                 return;
2038         }
2039
2040         for_lookup_table_entry(ctx->wim->lookup_table,
2041                                process_filtered_stream, ctx);
2042
2043         /* Streams in logical write list that were filtered can be removed.  */
2044         list_for_each_entry_safe(lte, tmp, &ctx->stream_list,
2045                                  write_streams_list)
2046                 if (lte->filtered)
2047                         list_del(&lte->write_streams_list);
2048 }
2049
2050 /* Prepares list of streams to write for the specified WIM image(s).  This wraps
2051  * around prepare_logical_stream_list() to handle filtering out streams already
2052  * present in the WIM or already present in external WIMs, depending on the
2053  * write flags provided.
2054  *
2055  * Note: some additional data is stored in each `struct wim_lookup_table_entry':
2056  *
2057  * - 'out_refcnt' is set to the number of references found for the logical write.
2058  *    This will be nonzero on all streams in the list returned by this function,
2059  *    but will also be nonzero on streams not in the list that were included in
2060  *    the logical write list, but filtered out from the returned list.
2061  * - 'filtered' is set to nonzero if the stream was filtered.  Filtered streams
2062  *   are not included in the list of streams returned by this function.
2063  * - 'unique_size' is set if the stream has a unique size among all streams in
2064  *   the logical write plus any filtered streams in the entire WIM that could
2065  *   potentially turn out to have the same checksum as a yet-to-be-checksummed
2066  *   stream being written.
2067  */
2068 static int
2069 prepare_stream_list(WIMStruct *wim, int image, int write_flags,
2070                     struct list_head *stream_list)
2071 {
2072         int ret;
2073         bool streams_ok;
2074         struct find_streams_ctx ctx;
2075
2076         INIT_LIST_HEAD(&ctx.stream_list);
2077         ret = init_stream_size_table(&ctx.stream_size_tab,
2078                                      wim->lookup_table->capacity);
2079         if (ret)
2080                 return ret;
2081         ctx.write_flags = write_flags;
2082         ctx.wim = wim;
2083
2084         streams_ok = ((write_flags & WIMLIB_WRITE_FLAG_STREAMS_OK) != 0);
2085
2086         ret = prepare_logical_stream_list(wim, image, streams_ok, &ctx);
2087         if (ret)
2088                 goto out_destroy_table;
2089
2090         handle_stream_filtering(&ctx);
2091         list_transfer(&ctx.stream_list, stream_list);
2092         ret = 0;
2093 out_destroy_table:
2094         destroy_stream_size_table(&ctx.stream_size_tab);
2095         return ret;
2096 }
2097
2098 static int
2099 write_wim_streams(WIMStruct *wim, int image, int write_flags,
2100                   unsigned num_threads,
2101                   wimlib_progress_func_t progress_func,
2102                   struct list_head *stream_list_override)
2103 {
2104         int ret;
2105         struct list_head _stream_list;
2106         struct list_head *stream_list;
2107         struct wim_lookup_table_entry *lte;
2108
2109         if (stream_list_override == NULL) {
2110                 /* Normal case: prepare stream list from image(s) being written.
2111                  */
2112                 stream_list = &_stream_list;
2113                 ret = prepare_stream_list(wim, image, write_flags, stream_list);
2114                 if (ret)
2115                         return ret;
2116         } else {
2117                 /* Currently only as a result of wimlib_split() being called:
2118                  * use stream list already explicitly provided.  Use existing
2119                  * reference counts.  */
2120                 stream_list = stream_list_override;
2121                 list_for_each_entry(lte, stream_list, write_streams_list)
2122                         lte->out_refcnt = (lte->refcnt ? lte->refcnt : 1);
2123         }
2124
2125         return write_stream_list(stream_list,
2126                                  wim->lookup_table,
2127                                  &wim->out_fd,
2128                                  wim->out_compression_type,
2129                                  wim->out_chunk_size,
2130                                  &wim->lzx_context,
2131                                  write_flags,
2132                                  num_threads,
2133                                  progress_func);
2134 }
2135
2136 static int
2137 write_wim_metadata_resources(WIMStruct *wim, int image, int write_flags,
2138                              wimlib_progress_func_t progress_func)
2139 {
2140         int ret;
2141         int start_image;
2142         int end_image;
2143         int write_resource_flags;
2144
2145         if (write_flags & WIMLIB_WRITE_FLAG_NO_METADATA) {
2146                 DEBUG("Not writing any metadata resources.");
2147                 return 0;
2148         }
2149
2150         write_resource_flags = write_flags_to_resource_flags(write_flags);
2151
2152         DEBUG("Writing metadata resources (offset=%"PRIu64")",
2153               wim->out_fd.offset);
2154
2155         if (progress_func)
2156                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_BEGIN, NULL);
2157
2158         if (image == WIMLIB_ALL_IMAGES) {
2159                 start_image = 1;
2160                 end_image = wim->hdr.image_count;
2161         } else {
2162                 start_image = image;
2163                 end_image = image;
2164         }
2165
2166         for (int i = start_image; i <= end_image; i++) {
2167                 struct wim_image_metadata *imd;
2168
2169                 imd = wim->image_metadata[i - 1];
2170                 /* Build a new metadata resource only if image was modified from
2171                  * the original (or was newly added).  Otherwise just copy the
2172                  * existing one.  */
2173                 if (imd->modified) {
2174                         DEBUG("Image %u was modified; building and writing new "
2175                               "metadata resource", i);
2176                         ret = write_metadata_resource(wim, i,
2177                                                       write_resource_flags);
2178                 } else if (write_flags & WIMLIB_WRITE_FLAG_OVERWRITE) {
2179                         DEBUG("Image %u was not modified; re-using existing "
2180                               "metadata resource.", i);
2181                         wim_res_spec_to_hdr(imd->metadata_lte->rspec,
2182                                             &imd->metadata_lte->out_reshdr);
2183                         ret = 0;
2184                 } else {
2185                         DEBUG("Image %u was not modified; copying existing "
2186                               "metadata resource.", i);
2187                         ret = write_wim_resource(imd->metadata_lte,
2188                                                  &wim->out_fd,
2189                                                  wim->out_compression_type,
2190                                                  wim->out_chunk_size,
2191                                                  &imd->metadata_lte->out_reshdr,
2192                                                  write_resource_flags,
2193                                                  &wim->lzx_context);
2194                 }
2195                 if (ret)
2196                         return ret;
2197         }
2198         if (progress_func)
2199                 progress_func(WIMLIB_PROGRESS_MSG_WRITE_METADATA_END, NULL);
2200         return 0;
2201 }
2202
2203 static int
2204 open_wim_writable(WIMStruct *wim, const tchar *path, int open_flags)
2205 {
2206         int raw_fd;
2207         DEBUG("Opening \"%"TS"\" for writing.", path);
2208
2209         raw_fd = topen(path, open_flags | O_BINARY, 0644);
2210         if (raw_fd < 0) {
2211                 ERROR_WITH_ERRNO("Failed to open \"%"TS"\" for writing", path);
2212                 return WIMLIB_ERR_OPEN;
2213         }
2214         filedes_init(&wim->out_fd, raw_fd);
2215         return 0;
2216 }
2217
2218 static int
2219 close_wim_writable(WIMStruct *wim, int write_flags)
2220 {
2221         int ret = 0;
2222
2223         if (!(write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)) {
2224                 DEBUG("Closing WIM file.");
2225                 if (filedes_valid(&wim->out_fd))
2226                         if (filedes_close(&wim->out_fd))
2227                                 ret = WIMLIB_ERR_WRITE;
2228         }
2229         filedes_invalidate(&wim->out_fd);
2230         return ret;
2231 }
2232
2233 /*
2234  * finish_write():
2235  *
2236  * Finish writing a WIM file: write the lookup table, xml data, and integrity
2237  * table, then overwrite the WIM header.  By default, closes the WIM file
2238  * descriptor (@wim->out_fd) if successful.
2239  *
2240  * write_flags is a bitwise OR of the following:
2241  *
2242  *      (public) WIMLIB_WRITE_FLAG_CHECK_INTEGRITY:
2243  *              Include an integrity table.
2244  *
2245  *      (public) WIMLIB_WRITE_FLAG_FSYNC:
2246  *              fsync() the output file before closing it.
2247  *
2248  *      (public) WIMLIB_WRITE_FLAG_PIPABLE:
2249  *              Writing a pipable WIM, possibly to a pipe; include pipable WIM
2250  *              stream headers before the lookup table and XML data, and also
2251  *              write the WIM header at the end instead of seeking to the
2252  *              beginning.  Can't be combined with
2253  *              WIMLIB_WRITE_FLAG_CHECK_INTEGRITY.
2254  *
2255  *      (private) WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE:
2256  *              Don't write the lookup table.
2257  *
2258  *      (private) WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE:
2259  *              When (if) writing the integrity table, re-use entries from the
2260  *              existing integrity table, if possible.
2261  *
2262  *      (private) WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML:
2263  *              After writing the XML data but before writing the integrity
2264  *              table, write a temporary WIM header and flush the stream so that
2265  *              the WIM is less likely to become corrupted upon abrupt program
2266  *              termination.
2267  *      (private) WIMLIB_WRITE_FLAG_HEADER_AT_END:
2268  *              Instead of overwriting the WIM header at the beginning of the
2269  *              file, simply append it to the end of the file.  (Used when
2270  *              writing to pipe.)
2271  *      (private) WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR:
2272  *              Do not close the file descriptor @wim->out_fd on either success
2273  *              on failure.
2274  *      (private) WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES:
2275  *              Use the existing <TOTALBYTES> stored in the in-memory XML
2276  *              information, rather than setting it to the offset of the XML
2277  *              data being written.
2278  */
2279 static int
2280 finish_write(WIMStruct *wim, int image, int write_flags,
2281              wimlib_progress_func_t progress_func,
2282              struct list_head *stream_list_override)
2283 {
2284         int ret;
2285         off_t hdr_offset;
2286         int write_resource_flags;
2287         off_t old_lookup_table_end;
2288         off_t new_lookup_table_end;
2289         u64 xml_totalbytes;
2290
2291         DEBUG("image=%d, write_flags=%08x", image, write_flags);
2292
2293         write_resource_flags = write_flags_to_resource_flags(write_flags);
2294
2295         /* In the WIM header, there is room for the resource entry for a
2296          * metadata resource labeled as the "boot metadata".  This entry should
2297          * be zeroed out if there is no bootable image (boot_idx 0).  Otherwise,
2298          * it should be a copy of the resource entry for the image that is
2299          * marked as bootable.  This is not well documented...  */
2300         if (wim->hdr.boot_idx == 0) {
2301                 zero_reshdr(&wim->hdr.boot_metadata_reshdr);
2302         } else {
2303                 copy_reshdr(&wim->hdr.boot_metadata_reshdr,
2304                             &wim->image_metadata[wim->hdr.boot_idx- 1
2305                                         ]->metadata_lte->out_reshdr);
2306         }
2307
2308         /* Write lookup table.  (Save old position first.)  */
2309         old_lookup_table_end = wim->hdr.lookup_table_reshdr.offset_in_wim +
2310                                wim->hdr.lookup_table_reshdr.size_in_wim;
2311         if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
2312                 ret = write_wim_lookup_table(wim, image, write_flags,
2313                                              &wim->hdr.lookup_table_reshdr,
2314                                              stream_list_override);
2315                 if (ret)
2316                         return ret;
2317         }
2318
2319         /* Write XML data.  */
2320         xml_totalbytes = wim->out_fd.offset;
2321         if (write_flags & WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES)
2322                 xml_totalbytes = WIM_TOTALBYTES_USE_EXISTING;
2323         ret = write_wim_xml_data(wim, image, xml_totalbytes,
2324                                  &wim->hdr.xml_data_reshdr,
2325                                  write_resource_flags);
2326         if (ret)
2327                 return ret;
2328
2329         /* Write integrity table (optional).  */
2330         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
2331                 if (write_flags & WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) {
2332                         struct wim_header checkpoint_hdr;
2333                         memcpy(&checkpoint_hdr, &wim->hdr, sizeof(struct wim_header));
2334                         zero_reshdr(&checkpoint_hdr.integrity_table_reshdr);
2335                         checkpoint_hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2336                         ret = write_wim_header_at_offset(&checkpoint_hdr,
2337                                                          &wim->out_fd, 0);
2338                         if (ret)
2339                                 return ret;
2340                 }
2341
2342                 if (!(write_flags & WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE))
2343                         old_lookup_table_end = 0;
2344
2345                 new_lookup_table_end = wim->hdr.lookup_table_reshdr.offset_in_wim +
2346                                        wim->hdr.lookup_table_reshdr.size_in_wim;
2347
2348                 ret = write_integrity_table(wim,
2349                                             new_lookup_table_end,
2350                                             old_lookup_table_end,
2351                                             progress_func);
2352                 if (ret)
2353                         return ret;
2354         } else {
2355                 /* No integrity table.  */
2356                 zero_reshdr(&wim->hdr.integrity_table_reshdr);
2357         }
2358
2359         /* Now that all information in the WIM header has been determined, the
2360          * preliminary header written earlier can be overwritten, the header of
2361          * the existing WIM file can be overwritten, or the final header can be
2362          * written to the end of the pipable WIM.  */
2363         wim->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2364         hdr_offset = 0;
2365         if (write_flags & WIMLIB_WRITE_FLAG_HEADER_AT_END)
2366                 hdr_offset = wim->out_fd.offset;
2367         DEBUG("Writing new header @ %"PRIu64".", hdr_offset);
2368         ret = write_wim_header_at_offset(&wim->hdr, &wim->out_fd, hdr_offset);
2369         if (ret)
2370                 return ret;
2371
2372         /* Possibly sync file data to disk before closing.  On POSIX systems, it
2373          * is necessary to do this before using rename() to overwrite an
2374          * existing file with a new file.  Otherwise, data loss would occur if
2375          * the system is abruptly terminated when the metadata for the rename
2376          * operation has been written to disk, but the new file data has not.
2377          */
2378         if (write_flags & WIMLIB_WRITE_FLAG_FSYNC) {
2379                 DEBUG("Syncing WIM file.");
2380                 if (fsync(wim->out_fd.fd)) {
2381                         ERROR_WITH_ERRNO("Error syncing data to WIM file");
2382                         return WIMLIB_ERR_WRITE;
2383                 }
2384         }
2385
2386         if (close_wim_writable(wim, write_flags)) {
2387                 ERROR_WITH_ERRNO("Failed to close the output WIM file");
2388                 return WIMLIB_ERR_WRITE;
2389         }
2390
2391         return 0;
2392 }
2393
2394 #if defined(HAVE_SYS_FILE_H) && defined(HAVE_FLOCK)
2395 int
2396 lock_wim(WIMStruct *wim, int fd)
2397 {
2398         int ret = 0;
2399         if (fd != -1 && !wim->wim_locked) {
2400                 ret = flock(fd, LOCK_EX | LOCK_NB);
2401                 if (ret != 0) {
2402                         if (errno == EWOULDBLOCK) {
2403                                 ERROR("`%"TS"' is already being modified or has been "
2404                                       "mounted read-write\n"
2405                                       "        by another process!", wim->filename);
2406                                 ret = WIMLIB_ERR_ALREADY_LOCKED;
2407                         } else {
2408                                 WARNING_WITH_ERRNO("Failed to lock `%"TS"'",
2409                                                    wim->filename);
2410                                 ret = 0;
2411                         }
2412                 } else {
2413                         wim->wim_locked = 1;
2414                 }
2415         }
2416         return ret;
2417 }
2418 #endif
2419
2420 /*
2421  * write_pipable_wim():
2422  *
2423  * Perform the intermediate stages of creating a "pipable" WIM (i.e. a WIM
2424  * capable of being applied from a pipe).
2425  *
2426  * Pipable WIMs are a wimlib-specific modification of the WIM format such that
2427  * images can be applied from them sequentially when the file data is sent over
2428  * a pipe.  In addition, a pipable WIM can be written sequentially to a pipe.
2429  * The modifications made to the WIM format for pipable WIMs are:
2430  *
2431  * - Magic characters in header are "WLPWM\0\0\0" (wimlib pipable WIM) instead
2432  *   of "MSWIM\0\0\0".  This lets wimlib know that the WIM is pipable and also
2433  *   stops other software from trying to read the file as a normal WIM.
2434  *
2435  * - The header at the beginning of the file does not contain all the normal
2436  *   information; in particular it will have all 0's for the lookup table and
2437  *   XML data resource entries.  This is because this information cannot be
2438  *   determined until the lookup table and XML data have been written.
2439  *   Consequently, wimlib will write the full header at the very end of the
2440  *   file.  The header at the end, however, is only used when reading the WIM
2441  *   from a seekable file (not a pipe).
2442  *
2443  * - An extra copy of the XML data is placed directly after the header.  This
2444  *   allows image names and sizes to be determined at an appropriate time when
2445  *   reading the WIM from a pipe.  This copy of the XML data is ignored if the
2446  *   WIM is read from a seekable file (not a pipe).
2447  *
2448  * - The format of resources, or streams, has been modified to allow them to be
2449  *   used before the "lookup table" has been read.  Each stream is prefixed with
2450  *   a `struct pwm_stream_hdr' that is basically an abbreviated form of `struct
2451  *   wim_lookup_table_entry_disk' that only contains the SHA1 message digest,
2452  *   uncompressed stream size, and flags that indicate whether the stream is
2453  *   compressed.  The data of uncompressed streams then follows literally, while
2454  *   the data of compressed streams follows in a modified format.  Compressed
2455  *   streams do not begin with a chunk table, since the chunk table cannot be
2456  *   written until all chunks have been compressed.  Instead, each compressed
2457  *   chunk is prefixed by a `struct pwm_chunk_hdr' that gives its size.
2458  *   Furthermore, the chunk table is written at the end of the resource instead
2459  *   of the start.  Note: chunk offsets are given in the chunk table as if the
2460  *   `struct pwm_chunk_hdr's were not present; also, the chunk table is only
2461  *   used if the WIM is being read from a seekable file (not a pipe).
2462  *
2463  * - Metadata resources always come before other file resources (streams).
2464  *   (This does not by itself constitute an incompatibility with normal WIMs,
2465  *   since this is valid in normal WIMs.)
2466  *
2467  * - At least up to the end of the file resources, all components must be packed
2468  *   as tightly as possible; there cannot be any "holes" in the WIM.  (This does
2469  *   not by itself consititute an incompatibility with normal WIMs, since this
2470  *   is valid in normal WIMs.)
2471  *
2472  * Note: the lookup table, XML data, and header at the end are not used when
2473  * applying from a pipe.  They exist to support functionality such as image
2474  * application and export when the WIM is *not* read from a pipe.
2475  *
2476  *   Layout of pipable WIM:
2477  *
2478  * ---------+----------+--------------------+----------------+--------------+-----------+--------+
2479  * | Header | XML data | Metadata resources | File resources | Lookup table | XML data  | Header |
2480  * ---------+----------+--------------------+----------------+--------------+-----------+--------+
2481  *
2482  *   Layout of normal WIM:
2483  *
2484  * +--------+-----------------------------+-------------------------+
2485  * | Header | File and metadata resources | Lookup table | XML data |
2486  * +--------+-----------------------------+-------------------------+
2487  *
2488  * An optional integrity table can follow the final XML data in both normal and
2489  * pipable WIMs.  However, due to implementation details, wimlib currently can
2490  * only include an integrity table in a pipable WIM when writing it to a
2491  * seekable file (not a pipe).
2492  *
2493  * Do note that since pipable WIMs are not supported by Microsoft's software,
2494  * wimlib does not create them unless explicitly requested (with
2495  * WIMLIB_WRITE_FLAG_PIPABLE) and as stated above they use different magic
2496  * characters to identify the file.
2497  */
2498 static int
2499 write_pipable_wim(WIMStruct *wim, int image, int write_flags,
2500                   unsigned num_threads, wimlib_progress_func_t progress_func,
2501                   struct list_head *stream_list_override)
2502 {
2503         int ret;
2504         struct wim_reshdr xml_reshdr;
2505
2506         WARNING("Creating a pipable WIM, which will "
2507                 "be incompatible\n"
2508                 "          with Microsoft's software (wimgapi/imagex/Dism).");
2509
2510         /* At this point, the header at the beginning of the file has already
2511          * been written.  */
2512
2513         /* For efficiency, when wimlib adds an image to the WIM with
2514          * wimlib_add_image(), the SHA1 message digests of files is not
2515          * calculated; instead, they are calculated while the files are being
2516          * written.  However, this does not work when writing a pipable WIM,
2517          * since when writing a stream to a pipable WIM, its SHA1 message digest
2518          * needs to be known before the stream data is written.  Therefore,
2519          * before getting much farther, we need to pre-calculate the SHA1
2520          * message digests of all streams that will be written.  */
2521         ret = wim_checksum_unhashed_streams(wim);
2522         if (ret)
2523                 return ret;
2524
2525         /* Write extra copy of the XML data.  */
2526         ret = write_wim_xml_data(wim, image, WIM_TOTALBYTES_OMIT,
2527                                  &xml_reshdr,
2528                                  WIMLIB_WRITE_RESOURCE_FLAG_PIPABLE);
2529         if (ret)
2530                 return ret;
2531
2532         /* Write metadata resources for the image(s) being included in the
2533          * output WIM.  */
2534         ret = write_wim_metadata_resources(wim, image, write_flags,
2535                                            progress_func);
2536         if (ret)
2537                 return ret;
2538
2539         /* Write streams needed for the image(s) being included in the output
2540          * WIM, or streams needed for the split WIM part.  */
2541         return write_wim_streams(wim, image, write_flags, num_threads,
2542                                  progress_func, stream_list_override);
2543
2544         /* The lookup table, XML data, and header at end are handled by
2545          * finish_write().  */
2546 }
2547
2548 /* Write a standalone WIM or split WIM (SWM) part to a new file or to a file
2549  * descriptor.  */
2550 int
2551 write_wim_part(WIMStruct *wim,
2552                const void *path_or_fd,
2553                int image,
2554                int write_flags,
2555                unsigned num_threads,
2556                wimlib_progress_func_t progress_func,
2557                unsigned part_number,
2558                unsigned total_parts,
2559                struct list_head *stream_list_override,
2560                const u8 *guid)
2561 {
2562         int ret;
2563         struct wim_header hdr_save;
2564         struct list_head lt_stream_list_override;
2565
2566         if (total_parts == 1)
2567                 DEBUG("Writing standalone WIM.");
2568         else
2569                 DEBUG("Writing split WIM part %u/%u", part_number, total_parts);
2570         if (image == WIMLIB_ALL_IMAGES)
2571                 DEBUG("Including all images.");
2572         else
2573                 DEBUG("Including image %d only.", image);
2574         if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
2575                 DEBUG("File descriptor: %d", *(const int*)path_or_fd);
2576         else
2577                 DEBUG("Path: \"%"TS"\"", (const tchar*)path_or_fd);
2578         DEBUG("Write flags: 0x%08x", write_flags);
2579         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY)
2580                 DEBUG("\tCHECK_INTEGRITY");
2581         if (write_flags & WIMLIB_WRITE_FLAG_REBUILD)
2582                 DEBUG("\tREBUILD");
2583         if (write_flags & WIMLIB_WRITE_FLAG_RECOMPRESS)
2584                 DEBUG("\tRECOMPRESS");
2585         if (write_flags & WIMLIB_WRITE_FLAG_FSYNC)
2586                 DEBUG("\tFSYNC");
2587         if (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE)
2588                 DEBUG("\tFSYNC");
2589         if (write_flags & WIMLIB_WRITE_FLAG_IGNORE_READONLY_FLAG)
2590                 DEBUG("\tIGNORE_READONLY_FLAG");
2591         if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
2592                 DEBUG("\tPIPABLE");
2593         if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
2594                 DEBUG("\tFILE_DESCRIPTOR");
2595         if (write_flags & WIMLIB_WRITE_FLAG_NO_METADATA)
2596                 DEBUG("\tNO_METADATA");
2597         if (write_flags & WIMLIB_WRITE_FLAG_USE_EXISTING_TOTALBYTES)
2598                 DEBUG("\tUSE_EXISTING_TOTALBYTES");
2599         if (num_threads == 0)
2600                 DEBUG("Number of threads: autodetect");
2601         else
2602                 DEBUG("Number of threads: %u", num_threads);
2603         DEBUG("Progress function: %s", (progress_func ? "yes" : "no"));
2604         DEBUG("Stream list:       %s", (stream_list_override ? "specified" : "autodetect"));
2605         DEBUG("GUID:              %s", ((guid || wim->guid_set_explicitly) ?
2606                                         "specified" : "generate new"));
2607
2608         /* Internally, this is always called with a valid part number and total
2609          * parts.  */
2610         wimlib_assert(total_parts >= 1);
2611         wimlib_assert(part_number >= 1 && part_number <= total_parts);
2612
2613         /* A valid image (or all images) must be specified.  */
2614         if (image != WIMLIB_ALL_IMAGES &&
2615              (image < 1 || image > wim->hdr.image_count))
2616                 return WIMLIB_ERR_INVALID_IMAGE;
2617
2618         /* If we need to write metadata resources, make sure the ::WIMStruct has
2619          * the needed information attached (e.g. is not a resource-only WIM,
2620          * such as a non-first part of a split WIM).  */
2621         if (!wim_has_metadata(wim) &&
2622             !(write_flags & WIMLIB_WRITE_FLAG_NO_METADATA))
2623                 return WIMLIB_ERR_METADATA_NOT_FOUND;
2624
2625         /* Check for contradictory flags.  */
2626         if ((write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2627                             WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY))
2628                                 == (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2629                                     WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY))
2630                 return WIMLIB_ERR_INVALID_PARAM;
2631
2632         if ((write_flags & (WIMLIB_WRITE_FLAG_PIPABLE |
2633                             WIMLIB_WRITE_FLAG_NOT_PIPABLE))
2634                                 == (WIMLIB_WRITE_FLAG_PIPABLE |
2635                                     WIMLIB_WRITE_FLAG_NOT_PIPABLE))
2636                 return WIMLIB_ERR_INVALID_PARAM;
2637
2638         /* Save previous header, then start initializing the new one.  */
2639         memcpy(&hdr_save, &wim->hdr, sizeof(struct wim_header));
2640
2641         /* Set default integrity and pipable flags.  */
2642         if (!(write_flags & (WIMLIB_WRITE_FLAG_PIPABLE |
2643                              WIMLIB_WRITE_FLAG_NOT_PIPABLE)))
2644                 if (wim_is_pipable(wim))
2645                         write_flags |= WIMLIB_WRITE_FLAG_PIPABLE;
2646
2647         if (!(write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2648                              WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY)))
2649                 if (wim_has_integrity_table(wim))
2650                         write_flags |= WIMLIB_WRITE_FLAG_CHECK_INTEGRITY;
2651
2652         /* Set appropriate magic number.  */
2653         if (write_flags & WIMLIB_WRITE_FLAG_PIPABLE)
2654                 wim->hdr.magic = PWM_MAGIC;
2655         else
2656                 wim->hdr.magic = WIM_MAGIC;
2657
2658         /* Clear header flags that will be set automatically.  */
2659         wim->hdr.flags &= ~(WIM_HDR_FLAG_METADATA_ONLY          |
2660                             WIM_HDR_FLAG_RESOURCE_ONLY          |
2661                             WIM_HDR_FLAG_SPANNED                |
2662                             WIM_HDR_FLAG_WRITE_IN_PROGRESS);
2663
2664         /* Set SPANNED header flag if writing part of a split WIM.  */
2665         if (total_parts != 1)
2666                 wim->hdr.flags |= WIM_HDR_FLAG_SPANNED;
2667
2668         /* Set part number and total parts of split WIM.  This will be 1 and 1
2669          * if the WIM is standalone.  */
2670         wim->hdr.part_number = part_number;
2671         wim->hdr.total_parts = total_parts;
2672
2673         /* Set compression type if different.  */
2674         if (wim->compression_type != wim->out_compression_type) {
2675                 ret = set_wim_hdr_cflags(wim->out_compression_type, &wim->hdr);
2676                 wimlib_assert(ret == 0);
2677         }
2678
2679         /* Set chunk size if different.  */
2680         wim->hdr.chunk_size = wim->out_chunk_size;
2681
2682         /* Use GUID if specified; otherwise generate a new one.  */
2683         if (guid)
2684                 memcpy(wim->hdr.guid, guid, WIMLIB_GUID_LEN);
2685         else if (!wim->guid_set_explicitly)
2686                 randomize_byte_array(wim->hdr.guid, WIMLIB_GUID_LEN);
2687
2688         /* Clear references to resources that have not been written yet.  */
2689         zero_reshdr(&wim->hdr.lookup_table_reshdr);
2690         zero_reshdr(&wim->hdr.xml_data_reshdr);
2691         zero_reshdr(&wim->hdr.boot_metadata_reshdr);
2692         zero_reshdr(&wim->hdr.integrity_table_reshdr);
2693
2694         /* Set image count and boot index correctly for single image writes.  */
2695         if (image != WIMLIB_ALL_IMAGES) {
2696                 wim->hdr.image_count = 1;
2697                 if (wim->hdr.boot_idx == image)
2698                         wim->hdr.boot_idx = 1;
2699                 else
2700                         wim->hdr.boot_idx = 0;
2701         }
2702
2703         /* Split WIMs can't be bootable.  */
2704         if (total_parts != 1)
2705                 wim->hdr.boot_idx = 0;
2706
2707         /* Initialize output file descriptor.  */
2708         if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR) {
2709                 /* File descriptor was explicitly provided.  Return error if
2710                  * file descriptor is not seekable, unless writing a pipable WIM
2711                  * was requested.  */
2712                 wim->out_fd.fd = *(const int*)path_or_fd;
2713                 wim->out_fd.offset = 0;
2714                 if (!filedes_is_seekable(&wim->out_fd)) {
2715                         ret = WIMLIB_ERR_INVALID_PARAM;
2716                         if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE))
2717                                 goto out_restore_hdr;
2718                         if (write_flags & WIMLIB_WRITE_FLAG_CHECK_INTEGRITY) {
2719                                 ERROR("Can't include integrity check when "
2720                                       "writing pipable WIM to pipe!");
2721                                 goto out_restore_hdr;
2722                         }
2723                 }
2724
2725         } else {
2726                 /* Filename of WIM to write was provided; open file descriptor
2727                  * to it.  */
2728                 ret = open_wim_writable(wim, (const tchar*)path_or_fd,
2729                                         O_TRUNC | O_CREAT | O_RDWR);
2730                 if (ret)
2731                         goto out_restore_hdr;
2732         }
2733
2734         /* Write initial header.  This is merely a "dummy" header since it
2735          * doesn't have all the information yet, so it will be overwritten later
2736          * (unless writing a pipable WIM).  */
2737         if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE))
2738                 wim->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2739         ret = write_wim_header(&wim->hdr, &wim->out_fd);
2740         wim->hdr.flags &= ~WIM_HDR_FLAG_WRITE_IN_PROGRESS;
2741         if (ret)
2742                 goto out_restore_hdr;
2743
2744         if (stream_list_override) {
2745                 struct wim_lookup_table_entry *lte;
2746                 INIT_LIST_HEAD(&lt_stream_list_override);
2747                 list_for_each_entry(lte, stream_list_override,
2748                                     write_streams_list)
2749                 {
2750                         list_add_tail(&lte->lookup_table_list,
2751                                       &lt_stream_list_override);
2752                 }
2753         }
2754
2755         /* Write metadata resources and streams.  */
2756         if (!(write_flags & WIMLIB_WRITE_FLAG_PIPABLE)) {
2757                 /* Default case: create a normal (non-pipable) WIM.  */
2758                 ret = write_wim_streams(wim, image, write_flags, num_threads,
2759                                         progress_func, stream_list_override);
2760                 if (ret)
2761                         goto out_restore_hdr;
2762
2763                 ret = write_wim_metadata_resources(wim, image, write_flags,
2764                                                    progress_func);
2765                 if (ret)
2766                         goto out_restore_hdr;
2767         } else {
2768                 /* Non-default case: create pipable WIM.  */
2769                 ret = write_pipable_wim(wim, image, write_flags, num_threads,
2770                                         progress_func, stream_list_override);
2771                 if (ret)
2772                         goto out_restore_hdr;
2773                 write_flags |= WIMLIB_WRITE_FLAG_HEADER_AT_END;
2774         }
2775
2776         if (stream_list_override)
2777                 stream_list_override = &lt_stream_list_override;
2778
2779         /* Write lookup table, XML data, and (optional) integrity table.  */
2780         ret = finish_write(wim, image, write_flags, progress_func,
2781                            stream_list_override);
2782 out_restore_hdr:
2783         memcpy(&wim->hdr, &hdr_save, sizeof(struct wim_header));
2784         (void)close_wim_writable(wim, write_flags);
2785         DEBUG("ret=%d", ret);
2786         return ret;
2787 }
2788
2789 /* Write a standalone WIM to a file or file descriptor.  */
2790 static int
2791 write_standalone_wim(WIMStruct *wim, const void *path_or_fd,
2792                      int image, int write_flags, unsigned num_threads,
2793                      wimlib_progress_func_t progress_func)
2794 {
2795         return write_wim_part(wim, path_or_fd, image, write_flags,
2796                               num_threads, progress_func, 1, 1, NULL, NULL);
2797 }
2798
2799 /* API function documented in wimlib.h  */
2800 WIMLIBAPI int
2801 wimlib_write(WIMStruct *wim, const tchar *path,
2802              int image, int write_flags, unsigned num_threads,
2803              wimlib_progress_func_t progress_func)
2804 {
2805         if (!path)
2806                 return WIMLIB_ERR_INVALID_PARAM;
2807
2808         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2809
2810         return write_standalone_wim(wim, path, image, write_flags,
2811                                     num_threads, progress_func);
2812 }
2813
2814 /* API function documented in wimlib.h  */
2815 WIMLIBAPI int
2816 wimlib_write_to_fd(WIMStruct *wim, int fd,
2817                    int image, int write_flags, unsigned num_threads,
2818                    wimlib_progress_func_t progress_func)
2819 {
2820         if (fd < 0)
2821                 return WIMLIB_ERR_INVALID_PARAM;
2822
2823         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
2824         write_flags |= WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR;
2825
2826         return write_standalone_wim(wim, &fd, image, write_flags,
2827                                     num_threads, progress_func);
2828 }
2829
2830 static bool
2831 any_images_modified(WIMStruct *wim)
2832 {
2833         for (int i = 0; i < wim->hdr.image_count; i++)
2834                 if (wim->image_metadata[i]->modified)
2835                         return true;
2836         return false;
2837 }
2838
2839 static int
2840 check_resource_offset(struct wim_lookup_table_entry *lte, void *_wim)
2841 {
2842         const WIMStruct *wim = _wim;
2843         off_t end_offset = *(const off_t*)wim->private;
2844
2845         if (lte->resource_location == RESOURCE_IN_WIM && lte->rspec->wim == wim &&
2846             lte->rspec->offset_in_wim + lte->rspec->size_in_wim > end_offset)
2847                 return WIMLIB_ERR_RESOURCE_ORDER;
2848         return 0;
2849 }
2850
2851 /* Make sure no file or metadata resources are located after the XML data (or
2852  * integrity table if present)--- otherwise we can't safely overwrite the WIM in
2853  * place and we return WIMLIB_ERR_RESOURCE_ORDER.  */
2854 static int
2855 check_resource_offsets(WIMStruct *wim, off_t end_offset)
2856 {
2857         int ret;
2858         unsigned i;
2859
2860         wim->private = &end_offset;
2861         ret = for_lookup_table_entry(wim->lookup_table, check_resource_offset, wim);
2862         if (ret)
2863                 return ret;
2864
2865         for (i = 0; i < wim->hdr.image_count; i++) {
2866                 ret = check_resource_offset(wim->image_metadata[i]->metadata_lte, wim);
2867                 if (ret)
2868                         return ret;
2869         }
2870         return 0;
2871 }
2872
2873 /*
2874  * Overwrite a WIM, possibly appending streams to it.
2875  *
2876  * A WIM looks like (or is supposed to look like) the following:
2877  *
2878  *                   Header (212 bytes)
2879  *                   Streams and metadata resources (variable size)
2880  *                   Lookup table (variable size)
2881  *                   XML data (variable size)
2882  *                   Integrity table (optional) (variable size)
2883  *
2884  * If we are not adding any streams or metadata resources, the lookup table is
2885  * unchanged--- so we only need to overwrite the XML data, integrity table, and
2886  * header.  This operation is potentially unsafe if the program is abruptly
2887  * terminated while the XML data or integrity table are being overwritten, but
2888  * before the new header has been written.  To partially alleviate this problem,
2889  * a special flag (WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML) is passed to
2890  * finish_write() to cause a temporary WIM header to be written after the XML
2891  * data has been written.  This may prevent the WIM from becoming corrupted if
2892  * the program is terminated while the integrity table is being calculated (but
2893  * no guarantees, due to write re-ordering...).
2894  *
2895  * If we are adding new streams or images (metadata resources), the lookup table
2896  * needs to be changed, and those streams need to be written.  In this case, we
2897  * try to perform a safe update of the WIM file by writing the streams *after*
2898  * the end of the previous WIM, then writing the new lookup table, XML data, and
2899  * (optionally) integrity table following the new streams.  This will produce a
2900  * layout like the following:
2901  *
2902  *                   Header (212 bytes)
2903  *                   (OLD) Streams and metadata resources (variable size)
2904  *                   (OLD) Lookup table (variable size)
2905  *                   (OLD) XML data (variable size)
2906  *                   (OLD) Integrity table (optional) (variable size)
2907  *                   (NEW) Streams and metadata resources (variable size)
2908  *                   (NEW) Lookup table (variable size)
2909  *                   (NEW) XML data (variable size)
2910  *                   (NEW) Integrity table (optional) (variable size)
2911  *
2912  * At all points, the WIM is valid as nothing points to the new data yet.  Then,
2913  * the header is overwritten to point to the new lookup table, XML data, and
2914  * integrity table, to produce the following layout:
2915  *
2916  *                   Header (212 bytes)
2917  *                   Streams and metadata resources (variable size)
2918  *                   Nothing (variable size)
2919  *                   More Streams and metadata resources (variable size)
2920  *                   Lookup table (variable size)
2921  *                   XML data (variable size)
2922  *                   Integrity table (optional) (variable size)
2923  *
2924  * This method allows an image to be appended to a large WIM very quickly, and
2925  * is is crash-safe except in the case of write re-ordering, but the
2926  * disadvantage is that a small hole is left in the WIM where the old lookup
2927  * table, xml data, and integrity table were.  (These usually only take up a
2928  * small amount of space compared to the streams, however.)
2929  */
2930 static int
2931 overwrite_wim_inplace(WIMStruct *wim, int write_flags,
2932                       unsigned num_threads,
2933                       wimlib_progress_func_t progress_func)
2934 {
2935         int ret;
2936         struct list_head stream_list;
2937         off_t old_wim_end;
2938         u64 old_lookup_table_end, old_xml_begin, old_xml_end;
2939         struct wim_header hdr_save;
2940
2941         DEBUG("Overwriting `%"TS"' in-place", wim->filename);
2942
2943         /* Set default integrity flag.  */
2944         if (!(write_flags & (WIMLIB_WRITE_FLAG_CHECK_INTEGRITY |
2945                              WIMLIB_WRITE_FLAG_NO_CHECK_INTEGRITY)))
2946                 if (wim_has_integrity_table(wim))
2947                         write_flags |= WIMLIB_WRITE_FLAG_CHECK_INTEGRITY;
2948
2949         /* Set additional flags for overwrite.  */
2950         write_flags |= WIMLIB_WRITE_FLAG_OVERWRITE |
2951                        WIMLIB_WRITE_FLAG_STREAMS_OK;
2952
2953         /* Make sure that the integrity table (if present) is after the XML
2954          * data, and that there are no stream resources, metadata resources, or
2955          * lookup tables after the XML data.  Otherwise, these data would be
2956          * overwritten. */
2957         old_xml_begin = wim->hdr.xml_data_reshdr.offset_in_wim;
2958         old_xml_end = old_xml_begin + wim->hdr.xml_data_reshdr.size_in_wim;
2959         old_lookup_table_end = wim->hdr.lookup_table_reshdr.offset_in_wim +
2960                                wim->hdr.lookup_table_reshdr.size_in_wim;
2961         if (wim->hdr.integrity_table_reshdr.offset_in_wim != 0 &&
2962             wim->hdr.integrity_table_reshdr.offset_in_wim < old_xml_end) {
2963                 WARNING("Didn't expect the integrity table to be before the XML data");
2964                 return WIMLIB_ERR_RESOURCE_ORDER;
2965         }
2966
2967         if (old_lookup_table_end > old_xml_begin) {
2968                 WARNING("Didn't expect the lookup table to be after the XML data");
2969                 return WIMLIB_ERR_RESOURCE_ORDER;
2970         }
2971
2972         /* Set @old_wim_end, which indicates the point beyond which we don't
2973          * allow any file and metadata resources to appear without returning
2974          * WIMLIB_ERR_RESOURCE_ORDER (due to the fact that we would otherwise
2975          * overwrite these resources). */
2976         if (!wim->deletion_occurred && !any_images_modified(wim)) {
2977                 /* If no images have been modified and no images have been
2978                  * deleted, a new lookup table does not need to be written.  We
2979                  * shall write the new XML data and optional integrity table
2980                  * immediately after the lookup table.  Note that this may
2981                  * overwrite an existing integrity table. */
2982                 DEBUG("Skipping writing lookup table "
2983                       "(no images modified or deleted)");
2984                 old_wim_end = old_lookup_table_end;
2985                 write_flags |= WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE |
2986                                WIMLIB_WRITE_FLAG_CHECKPOINT_AFTER_XML;
2987         } else if (wim->hdr.integrity_table_reshdr.offset_in_wim != 0) {
2988                 /* Old WIM has an integrity table; begin writing new streams
2989                  * after it. */
2990                 old_wim_end = wim->hdr.integrity_table_reshdr.offset_in_wim +
2991                               wim->hdr.integrity_table_reshdr.size_in_wim;
2992         } else {
2993                 /* No existing integrity table; begin writing new streams after
2994                  * the old XML data. */
2995                 old_wim_end = old_xml_end;
2996         }
2997
2998         ret = check_resource_offsets(wim, old_wim_end);
2999         if (ret)
3000                 return ret;
3001
3002         ret = prepare_stream_list(wim, WIMLIB_ALL_IMAGES, write_flags,
3003                                   &stream_list);
3004         if (ret)
3005                 return ret;
3006
3007         ret = open_wim_writable(wim, wim->filename, O_RDWR);
3008         if (ret)
3009                 return ret;
3010
3011         ret = lock_wim(wim, wim->out_fd.fd);
3012         if (ret)
3013                 goto out_close_wim;
3014
3015         /* Save original header so it can be restored in case of error  */
3016         memcpy(&hdr_save, &wim->hdr, sizeof(struct wim_header));
3017
3018         /* Set WIM_HDR_FLAG_WRITE_IN_PROGRESS flag in header. */
3019         wim->hdr.flags |= WIM_HDR_FLAG_WRITE_IN_PROGRESS;
3020         ret = write_wim_header_flags(wim->hdr.flags, &wim->out_fd);
3021         if (ret) {
3022                 ERROR_WITH_ERRNO("Error updating WIM header flags");
3023                 goto out_restore_memory_hdr;
3024         }
3025
3026         if (filedes_seek(&wim->out_fd, old_wim_end) == -1) {
3027                 ERROR_WITH_ERRNO("Can't seek to end of WIM");
3028                 ret = WIMLIB_ERR_WRITE;
3029                 goto out_restore_physical_hdr;
3030         }
3031
3032         ret = write_stream_list(&stream_list,
3033                                 wim->lookup_table,
3034                                 &wim->out_fd,
3035                                 wim->compression_type,
3036                                 wim->chunk_size,
3037                                 &wim->lzx_context,
3038                                 write_flags,
3039                                 num_threads,
3040                                 progress_func);
3041         if (ret)
3042                 goto out_truncate;
3043
3044         ret = write_wim_metadata_resources(wim, WIMLIB_ALL_IMAGES,
3045                                            write_flags, progress_func);
3046         if (ret)
3047                 goto out_truncate;
3048
3049         write_flags |= WIMLIB_WRITE_FLAG_REUSE_INTEGRITY_TABLE;
3050         ret = finish_write(wim, WIMLIB_ALL_IMAGES, write_flags,
3051                            progress_func, NULL);
3052         if (ret)
3053                 goto out_truncate;
3054
3055         goto out_unlock_wim;
3056
3057 out_truncate:
3058         if (!(write_flags & WIMLIB_WRITE_FLAG_NO_LOOKUP_TABLE)) {
3059                 WARNING("Truncating `%"TS"' to its original size (%"PRIu64" bytes)",
3060                         wim->filename, old_wim_end);
3061                 /* Return value of ftruncate() is ignored because this is
3062                  * already an error path.  */
3063                 (void)ftruncate(wim->out_fd.fd, old_wim_end);
3064         }
3065 out_restore_physical_hdr:
3066         (void)write_wim_header_flags(hdr_save.flags, &wim->out_fd);
3067 out_restore_memory_hdr:
3068         memcpy(&wim->hdr, &hdr_save, sizeof(struct wim_header));
3069 out_close_wim:
3070         (void)close_wim_writable(wim, write_flags);
3071 out_unlock_wim:
3072         wim->wim_locked = 0;
3073         return ret;
3074 }
3075
3076 static int
3077 overwrite_wim_via_tmpfile(WIMStruct *wim, int write_flags,
3078                           unsigned num_threads,
3079                           wimlib_progress_func_t progress_func)
3080 {
3081         size_t wim_name_len;
3082         int ret;
3083
3084         DEBUG("Overwriting `%"TS"' via a temporary file", wim->filename);
3085
3086         /* Write the WIM to a temporary file in the same directory as the
3087          * original WIM. */
3088         wim_name_len = tstrlen(wim->filename);
3089         tchar tmpfile[wim_name_len + 10];
3090         tmemcpy(tmpfile, wim->filename, wim_name_len);
3091         randomize_char_array_with_alnum(tmpfile + wim_name_len, 9);
3092         tmpfile[wim_name_len + 9] = T('\0');
3093
3094         ret = wimlib_write(wim, tmpfile, WIMLIB_ALL_IMAGES,
3095                            write_flags | WIMLIB_WRITE_FLAG_FSYNC,
3096                            num_threads, progress_func);
3097         if (ret) {
3098                 tunlink(tmpfile);
3099                 return ret;
3100         }
3101
3102         close_wim(wim);
3103
3104         /* Rename the new WIM file to the original WIM file.  Note: on Windows
3105          * this actually calls win32_rename_replacement(), not _wrename(), so
3106          * that removing the existing destination file can be handled.  */
3107         DEBUG("Renaming `%"TS"' to `%"TS"'", tmpfile, wim->filename);
3108         ret = trename(tmpfile, wim->filename);
3109         if (ret) {
3110                 ERROR_WITH_ERRNO("Failed to rename `%"TS"' to `%"TS"'",
3111                                  tmpfile, wim->filename);
3112         #ifdef __WIN32__
3113                 if (ret < 0)
3114         #endif
3115                 {
3116                         tunlink(tmpfile);
3117                 }
3118                 return WIMLIB_ERR_RENAME;
3119         }
3120
3121         if (progress_func) {
3122                 union wimlib_progress_info progress;
3123                 progress.rename.from = tmpfile;
3124                 progress.rename.to = wim->filename;
3125                 progress_func(WIMLIB_PROGRESS_MSG_RENAME, &progress);
3126         }
3127         return 0;
3128 }
3129
3130 /* API function documented in wimlib.h  */
3131 WIMLIBAPI int
3132 wimlib_overwrite(WIMStruct *wim, int write_flags,
3133                  unsigned num_threads,
3134                  wimlib_progress_func_t progress_func)
3135 {
3136         int ret;
3137         u32 orig_hdr_flags;
3138
3139         write_flags &= WIMLIB_WRITE_MASK_PUBLIC;
3140
3141         if (write_flags & WIMLIB_WRITE_FLAG_FILE_DESCRIPTOR)
3142                 return WIMLIB_ERR_INVALID_PARAM;
3143
3144         if (!wim->filename)
3145                 return WIMLIB_ERR_NO_FILENAME;
3146
3147         orig_hdr_flags = wim->hdr.flags;
3148         if (write_flags & WIMLIB_WRITE_FLAG_IGNORE_READONLY_FLAG)
3149                 wim->hdr.flags &= ~WIM_HDR_FLAG_READONLY;
3150         ret = can_modify_wim(wim);
3151         wim->hdr.flags = orig_hdr_flags;
3152         if (ret)
3153                 return ret;
3154
3155         if ((!wim->deletion_occurred || (write_flags & WIMLIB_WRITE_FLAG_SOFT_DELETE))
3156             && !(write_flags & (WIMLIB_WRITE_FLAG_REBUILD |
3157                                 WIMLIB_WRITE_FLAG_PIPABLE))
3158             && !(wim_is_pipable(wim))
3159             && wim->compression_type == wim->out_compression_type
3160             && wim->chunk_size == wim->out_chunk_size)
3161         {
3162                 ret = overwrite_wim_inplace(wim, write_flags, num_threads,
3163                                             progress_func);
3164                 if (ret != WIMLIB_ERR_RESOURCE_ORDER)
3165                         return ret;
3166                 WARNING("Falling back to re-building entire WIM");
3167         }
3168         return overwrite_wim_via_tmpfile(wim, write_flags, num_threads,
3169                                          progress_func);
3170 }