1 /////////////////////////////////////////////////////////////////////////////// 2 // 3 /// \file file_io.c 4 /// \brief File opening, unlinking, and closing 5 // 6 // Author: Lasse Collin 7 // 8 // This file has been put into the public domain. 9 // You can do whatever you want with this file. 10 // 11 /////////////////////////////////////////////////////////////////////////////// 12 13 #include "private.h" 14 15 #include <fcntl.h> 16 17 #ifdef TUKLIB_DOSLIKE 18 # include <io.h> 19 #else 20 static bool warn_fchown; 21 #endif 22 23 #if defined(HAVE_FUTIMES) || defined(HAVE_FUTIMESAT) || defined(HAVE_UTIMES) 24 # include <sys/time.h> 25 #elif defined(HAVE_UTIME) 26 # include <utime.h> 27 #endif 28 29 #include "tuklib_open_stdxxx.h" 30 31 #ifndef O_BINARY 32 # define O_BINARY 0 33 #endif 34 35 #ifndef O_NOCTTY 36 # define O_NOCTTY 0 37 #endif 38 39 40 /// If true, try to create sparse files when decompressing. 41 static bool try_sparse = true; 42 43 #ifndef TUKLIB_DOSLIKE 44 /// File status flags of standard output. This is used by io_open_dest() 45 /// and io_close_dest(). 46 static int stdout_flags = 0; 47 #endif 48 49 50 static bool io_write_buf(file_pair *pair, const uint8_t *buf, size_t size); 51 52 53 extern void 54 io_init(void) 55 { 56 // Make sure that stdin, stdout, and and stderr are connected to 57 // a valid file descriptor. Exit immediately with exit code ERROR 58 // if we cannot make the file descriptors valid. Maybe we should 59 // print an error message, but our stderr could be screwed anyway. 60 tuklib_open_stdxxx(E_ERROR); 61 62 #ifndef TUKLIB_DOSLIKE 63 // If fchown() fails setting the owner, we warn about it only if 64 // we are root. 65 warn_fchown = geteuid() == 0; 66 #endif 67 68 #ifdef __DJGPP__ 69 // Avoid doing useless things when statting files. 70 // This isn't important but doesn't hurt. 71 _djstat_flags = _STAT_INODE | _STAT_EXEC_EXT 72 | _STAT_EXEC_MAGIC | _STAT_DIRSIZE; 73 #endif 74 75 return; 76 } 77 78 79 extern void 80 io_no_sparse(void) 81 { 82 try_sparse = false; 83 return; 84 } 85 86 87 /// \brief Unlink a file 88 /// 89 /// This tries to verify that the file being unlinked really is the file that 90 /// we want to unlink by verifying device and inode numbers. There's still 91 /// a small unavoidable race, but this is much better than nothing (the file 92 /// could have been moved/replaced even hours earlier). 93 static void 94 io_unlink(const char *name, const struct stat *known_st) 95 { 96 #if defined(TUKLIB_DOSLIKE) 97 // On DOS-like systems, st_ino is meaningless, so don't bother 98 // testing it. Just silence a compiler warning. 99 (void)known_st; 100 #else 101 struct stat new_st; 102 103 // If --force was used, use stat() instead of lstat(). This way 104 // (de)compressing symlinks works correctly. However, it also means 105 // that xz cannot detect if a regular file foo is renamed to bar 106 // and then a symlink foo -> bar is created. Because of stat() 107 // instead of lstat(), xz will think that foo hasn't been replaced 108 // with another file. Thus, xz will remove foo even though it no 109 // longer is the same file that xz used when it started compressing. 110 // Probably it's not too bad though, so this doesn't need a more 111 // complex fix. 112 const int stat_ret = opt_force 113 ? stat(name, &new_st) : lstat(name, &new_st); 114 115 if (stat_ret 116 # ifdef __VMS 117 // st_ino is an array, and we don't want to 118 // compare st_dev at all. 119 || memcmp(&new_st.st_ino, &known_st->st_ino, 120 sizeof(new_st.st_ino)) != 0 121 # else 122 // Typical POSIX-like system 123 || new_st.st_dev != known_st->st_dev 124 || new_st.st_ino != known_st->st_ino 125 # endif 126 ) 127 // TRANSLATORS: When compression or decompression finishes, 128 // and xz is going to remove the source file, xz first checks 129 // if the source file still exists, and if it does, does its 130 // device and inode numbers match what xz saw when it opened 131 // the source file. If these checks fail, this message is 132 // shown, %s being the filename, and the file is not deleted. 133 // The check for device and inode numbers is there, because 134 // it is possible that the user has put a new file in place 135 // of the original file, and in that case it obviously 136 // shouldn't be removed. 137 message_error(_("%s: File seems to have been moved, " 138 "not removing"), name); 139 else 140 #endif 141 // There's a race condition between lstat() and unlink() 142 // but at least we have tried to avoid removing wrong file. 143 if (unlink(name)) 144 message_error(_("%s: Cannot remove: %s"), 145 name, strerror(errno)); 146 147 return; 148 } 149 150 151 /// \brief Copies owner/group and permissions 152 /// 153 /// \todo ACL and EA support 154 /// 155 static void 156 io_copy_attrs(const file_pair *pair) 157 { 158 // Skip chown and chmod on Windows. 159 #ifndef TUKLIB_DOSLIKE 160 // This function is more tricky than you may think at first. 161 // Blindly copying permissions may permit users to access the 162 // destination file who didn't have permission to access the 163 // source file. 164 165 // Try changing the owner of the file. If we aren't root or the owner 166 // isn't already us, fchown() probably doesn't succeed. We warn 167 // about failing fchown() only if we are root. 168 if (fchown(pair->dest_fd, pair->src_st.st_uid, -1) && warn_fchown) 169 message_warning(_("%s: Cannot set the file owner: %s"), 170 pair->dest_name, strerror(errno)); 171 172 mode_t mode; 173 174 if (fchown(pair->dest_fd, -1, pair->src_st.st_gid)) { 175 message_warning(_("%s: Cannot set the file group: %s"), 176 pair->dest_name, strerror(errno)); 177 // We can still safely copy some additional permissions: 178 // `group' must be at least as strict as `other' and 179 // also vice versa. 180 // 181 // NOTE: After this, the owner of the source file may 182 // get additional permissions. This shouldn't be too bad, 183 // because the owner would have had permission to chmod 184 // the original file anyway. 185 mode = ((pair->src_st.st_mode & 0070) >> 3) 186 & (pair->src_st.st_mode & 0007); 187 mode = (pair->src_st.st_mode & 0700) | (mode << 3) | mode; 188 } else { 189 // Drop the setuid, setgid, and sticky bits. 190 mode = pair->src_st.st_mode & 0777; 191 } 192 193 if (fchmod(pair->dest_fd, mode)) 194 message_warning(_("%s: Cannot set the file permissions: %s"), 195 pair->dest_name, strerror(errno)); 196 #endif 197 198 // Copy the timestamps. We have several possible ways to do this, of 199 // which some are better in both security and precision. 200 // 201 // First, get the nanosecond part of the timestamps. As of writing, 202 // it's not standardized by POSIX, and there are several names for 203 // the same thing in struct stat. 204 long atime_nsec; 205 long mtime_nsec; 206 207 # if defined(HAVE_STRUCT_STAT_ST_ATIM_TV_NSEC) 208 // GNU and Solaris 209 atime_nsec = pair->src_st.st_atim.tv_nsec; 210 mtime_nsec = pair->src_st.st_mtim.tv_nsec; 211 212 # elif defined(HAVE_STRUCT_STAT_ST_ATIMESPEC_TV_NSEC) 213 // BSD 214 atime_nsec = pair->src_st.st_atimespec.tv_nsec; 215 mtime_nsec = pair->src_st.st_mtimespec.tv_nsec; 216 217 # elif defined(HAVE_STRUCT_STAT_ST_ATIMENSEC) 218 // GNU and BSD without extensions 219 atime_nsec = pair->src_st.st_atimensec; 220 mtime_nsec = pair->src_st.st_mtimensec; 221 222 # elif defined(HAVE_STRUCT_STAT_ST_UATIME) 223 // Tru64 224 atime_nsec = pair->src_st.st_uatime * 1000; 225 mtime_nsec = pair->src_st.st_umtime * 1000; 226 227 # elif defined(HAVE_STRUCT_STAT_ST_ATIM_ST__TIM_TV_NSEC) 228 // UnixWare 229 atime_nsec = pair->src_st.st_atim.st__tim.tv_nsec; 230 mtime_nsec = pair->src_st.st_mtim.st__tim.tv_nsec; 231 232 # else 233 // Safe fallback 234 atime_nsec = 0; 235 mtime_nsec = 0; 236 # endif 237 238 // Construct a structure to hold the timestamps and call appropriate 239 // function to set the timestamps. 240 #if defined(HAVE_FUTIMENS) 241 // Use nanosecond precision. 242 struct timespec tv[2]; 243 tv[0].tv_sec = pair->src_st.st_atime; 244 tv[0].tv_nsec = atime_nsec; 245 tv[1].tv_sec = pair->src_st.st_mtime; 246 tv[1].tv_nsec = mtime_nsec; 247 248 (void)futimens(pair->dest_fd, tv); 249 250 #elif defined(HAVE_FUTIMES) || defined(HAVE_FUTIMESAT) || defined(HAVE_UTIMES) 251 // Use microsecond precision. 252 struct timeval tv[2]; 253 tv[0].tv_sec = pair->src_st.st_atime; 254 tv[0].tv_usec = atime_nsec / 1000; 255 tv[1].tv_sec = pair->src_st.st_mtime; 256 tv[1].tv_usec = mtime_nsec / 1000; 257 258 # if defined(HAVE_FUTIMES) 259 (void)futimes(pair->dest_fd, tv); 260 # elif defined(HAVE_FUTIMESAT) 261 (void)futimesat(pair->dest_fd, NULL, tv); 262 # else 263 // Argh, no function to use a file descriptor to set the timestamp. 264 (void)utimes(pair->dest_name, tv); 265 # endif 266 267 #elif defined(HAVE_UTIME) 268 // Use one-second precision. utime() doesn't support using file 269 // descriptor either. Some systems have broken utime() prototype 270 // so don't make this const. 271 struct utimbuf buf = { 272 .actime = pair->src_st.st_atime, 273 .modtime = pair->src_st.st_mtime, 274 }; 275 276 // Avoid warnings. 277 (void)atime_nsec; 278 (void)mtime_nsec; 279 280 (void)utime(pair->dest_name, &buf); 281 #endif 282 283 return; 284 } 285 286 287 /// Opens the source file. Returns false on success, true on error. 288 static bool 289 io_open_src_real(file_pair *pair) 290 { 291 // There's nothing to open when reading from stdin. 292 if (pair->src_name == stdin_filename) { 293 pair->src_fd = STDIN_FILENO; 294 #ifdef TUKLIB_DOSLIKE 295 setmode(STDIN_FILENO, O_BINARY); 296 #endif 297 return false; 298 } 299 300 // Symlinks are not followed unless writing to stdout or --force 301 // was used. 302 const bool follow_symlinks = opt_stdout || opt_force; 303 304 // We accept only regular files if we are writing the output 305 // to disk too. bzip2 allows overriding this with --force but 306 // gzip and xz don't. 307 const bool reg_files_only = !opt_stdout; 308 309 // Flags for open() 310 int flags = O_RDONLY | O_BINARY | O_NOCTTY; 311 312 #ifndef TUKLIB_DOSLIKE 313 // If we accept only regular files, we need to be careful to avoid 314 // problems with special files like devices and FIFOs. O_NONBLOCK 315 // prevents blocking when opening such files. When we want to accept 316 // special files, we must not use O_NONBLOCK, or otherwise we won't 317 // block waiting e.g. FIFOs to become readable. 318 if (reg_files_only) 319 flags |= O_NONBLOCK; 320 #endif 321 322 #if defined(O_NOFOLLOW) 323 if (!follow_symlinks) 324 flags |= O_NOFOLLOW; 325 #elif !defined(TUKLIB_DOSLIKE) 326 // Some POSIX-like systems lack O_NOFOLLOW (it's not required 327 // by POSIX). Check for symlinks with a separate lstat() on 328 // these systems. 329 if (!follow_symlinks) { 330 struct stat st; 331 if (lstat(pair->src_name, &st)) { 332 message_error("%s: %s", pair->src_name, 333 strerror(errno)); 334 return true; 335 336 } else if (S_ISLNK(st.st_mode)) { 337 message_warning(_("%s: Is a symbolic link, " 338 "skipping"), pair->src_name); 339 return true; 340 } 341 } 342 #else 343 // Avoid warnings. 344 (void)follow_symlinks; 345 #endif 346 347 // Try to open the file. If we are accepting non-regular files, 348 // unblock the caught signals so that open() can be interrupted 349 // if it blocks e.g. due to a FIFO file. 350 if (!reg_files_only) 351 signals_unblock(); 352 353 // Maybe this wouldn't need a loop, since all the signal handlers for 354 // which we don't use SA_RESTART set user_abort to true. But it 355 // doesn't hurt to have it just in case. 356 do { 357 pair->src_fd = open(pair->src_name, flags); 358 } while (pair->src_fd == -1 && errno == EINTR && !user_abort); 359 360 if (!reg_files_only) 361 signals_block(); 362 363 if (pair->src_fd == -1) { 364 // If we were interrupted, don't display any error message. 365 if (errno == EINTR) { 366 // All the signals that don't have SA_RESTART 367 // set user_abort. 368 assert(user_abort); 369 return true; 370 } 371 372 #ifdef O_NOFOLLOW 373 // Give an understandable error message if the reason 374 // for failing was that the file was a symbolic link. 375 // 376 // Note that at least Linux, OpenBSD, Solaris, and Darwin 377 // use ELOOP to indicate that O_NOFOLLOW was the reason 378 // that open() failed. Because there may be 379 // directories in the pathname, ELOOP may occur also 380 // because of a symlink loop in the directory part. 381 // So ELOOP doesn't tell us what actually went wrong, 382 // and this stupidity went into POSIX-1.2008 too. 383 // 384 // FreeBSD associates EMLINK with O_NOFOLLOW and 385 // Tru64 uses ENOTSUP. We use these directly here 386 // and skip the lstat() call and the associated race. 387 // I want to hear if there are other kernels that 388 // fail with something else than ELOOP with O_NOFOLLOW. 389 bool was_symlink = false; 390 391 # if defined(__FreeBSD__) || defined(__DragonFly__) 392 if (errno == EMLINK) 393 was_symlink = true; 394 395 # elif defined(__digital__) && defined(__unix__) 396 if (errno == ENOTSUP) 397 was_symlink = true; 398 399 # elif defined(__NetBSD__) 400 // As of 2010-09-05, NetBSD doesn't document what errno is 401 // used with O_NOFOLLOW. It is EFTYPE though, and I 402 // understood that is very unlikely to change even though 403 // it is undocumented. 404 if (errno == EFTYPE) 405 was_symlink = true; 406 407 # else 408 if (errno == ELOOP && !follow_symlinks) { 409 const int saved_errno = errno; 410 struct stat st; 411 if (lstat(pair->src_name, &st) == 0 412 && S_ISLNK(st.st_mode)) 413 was_symlink = true; 414 415 errno = saved_errno; 416 } 417 # endif 418 419 if (was_symlink) 420 message_warning(_("%s: Is a symbolic link, " 421 "skipping"), pair->src_name); 422 else 423 #endif 424 // Something else than O_NOFOLLOW failing 425 // (assuming that the race conditions didn't 426 // confuse us). 427 message_error("%s: %s", pair->src_name, 428 strerror(errno)); 429 430 return true; 431 } 432 433 #ifndef TUKLIB_DOSLIKE 434 // Drop O_NONBLOCK, which is used only when we are accepting only 435 // regular files. After the open() call, we want things to block 436 // instead of giving EAGAIN. 437 if (reg_files_only) { 438 flags = fcntl(pair->src_fd, F_GETFL); 439 if (flags == -1) 440 goto error_msg; 441 442 flags &= ~O_NONBLOCK; 443 444 if (fcntl(pair->src_fd, F_SETFL, flags)) 445 goto error_msg; 446 } 447 #endif 448 449 // Stat the source file. We need the result also when we copy 450 // the permissions, and when unlinking. 451 if (fstat(pair->src_fd, &pair->src_st)) 452 goto error_msg; 453 454 if (S_ISDIR(pair->src_st.st_mode)) { 455 message_warning(_("%s: Is a directory, skipping"), 456 pair->src_name); 457 goto error; 458 } 459 460 if (reg_files_only) { 461 if (!S_ISREG(pair->src_st.st_mode)) { 462 message_warning(_("%s: Not a regular file, " 463 "skipping"), pair->src_name); 464 goto error; 465 } 466 467 // These are meaningless on Windows. 468 #ifndef TUKLIB_DOSLIKE 469 if (pair->src_st.st_mode & (S_ISUID | S_ISGID)) { 470 // gzip rejects setuid and setgid files even 471 // when --force was used. bzip2 doesn't check 472 // for them, but calls fchown() after fchmod(), 473 // and many systems automatically drop setuid 474 // and setgid bits there. 475 // 476 // We accept setuid and setgid files if 477 // --force was used. We drop these bits 478 // explicitly in io_copy_attr(). 479 message_warning(_("%s: File has setuid or " 480 "setgid bit set, skipping"), 481 pair->src_name); 482 goto error; 483 } 484 485 if (pair->src_st.st_mode & S_ISVTX) { 486 message_warning(_("%s: File has sticky bit " 487 "set, skipping"), 488 pair->src_name); 489 goto error; 490 } 491 492 if (pair->src_st.st_nlink > 1) { 493 message_warning(_("%s: Input file has more " 494 "than one hard link, " 495 "skipping"), pair->src_name); 496 goto error; 497 } 498 #endif 499 } 500 501 return false; 502 503 error_msg: 504 message_error("%s: %s", pair->src_name, strerror(errno)); 505 error: 506 (void)close(pair->src_fd); 507 return true; 508 } 509 510 511 extern file_pair * 512 io_open_src(const char *src_name) 513 { 514 if (is_empty_filename(src_name)) 515 return NULL; 516 517 // Since we have only one file open at a time, we can use 518 // a statically allocated structure. 519 static file_pair pair; 520 521 pair = (file_pair){ 522 .src_name = src_name, 523 .dest_name = NULL, 524 .src_fd = -1, 525 .dest_fd = -1, 526 .src_eof = false, 527 .dest_try_sparse = false, 528 .dest_pending_sparse = 0, 529 }; 530 531 // Block the signals, for which we have a custom signal handler, so 532 // that we don't need to worry about EINTR. 533 signals_block(); 534 const bool error = io_open_src_real(&pair); 535 signals_unblock(); 536 537 return error ? NULL : &pair; 538 } 539 540 541 /// \brief Closes source file of the file_pair structure 542 /// 543 /// \param pair File whose src_fd should be closed 544 /// \param success If true, the file will be removed from the disk if 545 /// closing succeeds and --keep hasn't been used. 546 static void 547 io_close_src(file_pair *pair, bool success) 548 { 549 if (pair->src_fd != STDIN_FILENO && pair->src_fd != -1) { 550 #ifdef TUKLIB_DOSLIKE 551 (void)close(pair->src_fd); 552 #endif 553 554 // If we are going to unlink(), do it before closing the file. 555 // This way there's no risk that someone replaces the file and 556 // happens to get same inode number, which would make us 557 // unlink() wrong file. 558 // 559 // NOTE: DOS-like systems are an exception to this, because 560 // they don't allow unlinking files that are open. *sigh* 561 if (success && !opt_keep_original) 562 io_unlink(pair->src_name, &pair->src_st); 563 564 #ifndef TUKLIB_DOSLIKE 565 (void)close(pair->src_fd); 566 #endif 567 } 568 569 return; 570 } 571 572 573 static bool 574 io_open_dest_real(file_pair *pair) 575 { 576 if (opt_stdout || pair->src_fd == STDIN_FILENO) { 577 // We don't modify or free() this. 578 pair->dest_name = (char *)"(stdout)"; 579 pair->dest_fd = STDOUT_FILENO; 580 #ifdef TUKLIB_DOSLIKE 581 setmode(STDOUT_FILENO, O_BINARY); 582 #endif 583 } else { 584 pair->dest_name = suffix_get_dest_name(pair->src_name); 585 if (pair->dest_name == NULL) 586 return true; 587 588 // If --force was used, unlink the target file first. 589 if (opt_force && unlink(pair->dest_name) && errno != ENOENT) { 590 message_error(_("%s: Cannot remove: %s"), 591 pair->dest_name, strerror(errno)); 592 free(pair->dest_name); 593 return true; 594 } 595 596 // Open the file. 597 const int flags = O_WRONLY | O_BINARY | O_NOCTTY 598 | O_CREAT | O_EXCL; 599 const mode_t mode = S_IRUSR | S_IWUSR; 600 pair->dest_fd = open(pair->dest_name, flags, mode); 601 602 if (pair->dest_fd == -1) { 603 message_error("%s: %s", pair->dest_name, 604 strerror(errno)); 605 free(pair->dest_name); 606 return true; 607 } 608 } 609 610 // If this really fails... well, we have a safe fallback. 611 if (fstat(pair->dest_fd, &pair->dest_st)) { 612 #if defined(__VMS) 613 pair->dest_st.st_ino[0] = 0; 614 pair->dest_st.st_ino[1] = 0; 615 pair->dest_st.st_ino[2] = 0; 616 #elif !defined(TUKLIB_DOSLIKE) 617 pair->dest_st.st_dev = 0; 618 pair->dest_st.st_ino = 0; 619 #endif 620 #ifndef TUKLIB_DOSLIKE 621 } else if (try_sparse && opt_mode == MODE_DECOMPRESS) { 622 // When writing to standard output, we need to be extra 623 // careful: 624 // - It may be connected to something else than 625 // a regular file. 626 // - We aren't necessarily writing to a new empty file 627 // or to the end of an existing file. 628 // - O_APPEND may be active. 629 // 630 // TODO: I'm keeping this disabled for DOS-like systems 631 // for now. FAT doesn't support sparse files, but NTFS 632 // does, so maybe this should be enabled on Windows after 633 // some testing. 634 if (pair->dest_fd == STDOUT_FILENO) { 635 if (!S_ISREG(pair->dest_st.st_mode)) 636 return false; 637 638 const int flags = fcntl(STDOUT_FILENO, F_GETFL); 639 if (flags == -1) 640 return false; 641 642 if (flags & O_APPEND) { 643 // Creating a sparse file is not possible 644 // when O_APPEND is active (it's used by 645 // shell's >> redirection). As I understand 646 // it, it is safe to temporarily disable 647 // O_APPEND in xz, because if someone 648 // happened to write to the same file at the 649 // same time, results would be bad anyway 650 // (users shouldn't assume that xz uses any 651 // specific block size when writing data). 652 // 653 // The write position may be something else 654 // than the end of the file, so we must fix 655 // it to start writing at the end of the file 656 // to imitate O_APPEND. 657 if (lseek(STDOUT_FILENO, 0, SEEK_END) == -1) 658 return false; 659 660 if (fcntl(STDOUT_FILENO, F_SETFL, 661 stdout_flags & ~O_APPEND)) 662 return false; 663 664 // Remember the flags so that io_close_dest() 665 // can restore them. 666 stdout_flags = flags; 667 668 } else if (lseek(STDOUT_FILENO, 0, SEEK_CUR) 669 != pair->dest_st.st_size) { 670 // Writing won't start exactly at the end 671 // of the file. We cannot use sparse output, 672 // because it would probably corrupt the file. 673 return false; 674 } 675 } 676 677 pair->dest_try_sparse = true; 678 #endif 679 } 680 681 return false; 682 } 683 684 685 extern bool 686 io_open_dest(file_pair *pair) 687 { 688 signals_block(); 689 const bool ret = io_open_dest_real(pair); 690 signals_unblock(); 691 return ret; 692 } 693 694 695 /// \brief Closes destination file of the file_pair structure 696 /// 697 /// \param pair File whose dest_fd should be closed 698 /// \param success If false, the file will be removed from the disk. 699 /// 700 /// \return Zero if closing succeeds. On error, -1 is returned and 701 /// error message printed. 702 static bool 703 io_close_dest(file_pair *pair, bool success) 704 { 705 #ifndef TUKLIB_DOSLIKE 706 // If io_open_dest() has disabled O_APPEND, restore it here. 707 if (stdout_flags != 0) { 708 assert(pair->dest_fd == STDOUT_FILENO); 709 710 const int fail = fcntl(STDOUT_FILENO, F_SETFL, stdout_flags); 711 stdout_flags = 0; 712 713 if (fail) { 714 message_error(_("Error restoring the O_APPEND flag " 715 "to standard output: %s"), 716 strerror(errno)); 717 return true; 718 } 719 } 720 #endif 721 722 if (pair->dest_fd == -1 || pair->dest_fd == STDOUT_FILENO) 723 return false; 724 725 if (close(pair->dest_fd)) { 726 message_error(_("%s: Closing the file failed: %s"), 727 pair->dest_name, strerror(errno)); 728 729 // Closing destination file failed, so we cannot trust its 730 // contents. Get rid of junk: 731 io_unlink(pair->dest_name, &pair->dest_st); 732 free(pair->dest_name); 733 return true; 734 } 735 736 // If the operation using this file wasn't successful, we git rid 737 // of the junk file. 738 if (!success) 739 io_unlink(pair->dest_name, &pair->dest_st); 740 741 free(pair->dest_name); 742 743 return false; 744 } 745 746 747 extern void 748 io_close(file_pair *pair, bool success) 749 { 750 // Take care of sparseness at the end of the output file. 751 if (success && pair->dest_try_sparse 752 && pair->dest_pending_sparse > 0) { 753 // Seek forward one byte less than the size of the pending 754 // hole, then write one zero-byte. This way the file grows 755 // to its correct size. An alternative would be to use 756 // ftruncate() but that isn't portable enough (e.g. it 757 // doesn't work with FAT on Linux; FAT isn't that important 758 // since it doesn't support sparse files anyway, but we don't 759 // want to create corrupt files on it). 760 if (lseek(pair->dest_fd, pair->dest_pending_sparse - 1, 761 SEEK_CUR) == -1) { 762 message_error(_("%s: Seeking failed when trying " 763 "to create a sparse file: %s"), 764 pair->dest_name, strerror(errno)); 765 success = false; 766 } else { 767 const uint8_t zero[1] = { '\0' }; 768 if (io_write_buf(pair, zero, 1)) 769 success = false; 770 } 771 } 772 773 signals_block(); 774 775 // Copy the file attributes. We need to skip this if destination 776 // file isn't open or it is standard output. 777 if (success && pair->dest_fd != -1 && pair->dest_fd != STDOUT_FILENO) 778 io_copy_attrs(pair); 779 780 // Close the destination first. If it fails, we must not remove 781 // the source file! 782 if (io_close_dest(pair, success)) 783 success = false; 784 785 // Close the source file, and unlink it if the operation using this 786 // file pair was successful and we haven't requested to keep the 787 // source file. 788 io_close_src(pair, success); 789 790 signals_unblock(); 791 792 return; 793 } 794 795 796 extern size_t 797 io_read(file_pair *pair, io_buf *buf_union, size_t size) 798 { 799 // We use small buffers here. 800 assert(size < SSIZE_MAX); 801 802 uint8_t *buf = buf_union->u8; 803 size_t left = size; 804 805 while (left > 0) { 806 const ssize_t amount = read(pair->src_fd, buf, left); 807 808 if (amount == 0) { 809 pair->src_eof = true; 810 break; 811 } 812 813 if (amount == -1) { 814 if (errno == EINTR) { 815 if (user_abort) 816 return SIZE_MAX; 817 818 continue; 819 } 820 821 message_error(_("%s: Read error: %s"), 822 pair->src_name, strerror(errno)); 823 824 // FIXME Is this needed? 825 pair->src_eof = true; 826 827 return SIZE_MAX; 828 } 829 830 buf += (size_t)(amount); 831 left -= (size_t)(amount); 832 } 833 834 return size - left; 835 } 836 837 838 extern bool 839 io_pread(file_pair *pair, io_buf *buf, size_t size, off_t pos) 840 { 841 // Using lseek() and read() is more portable than pread() and 842 // for us it is as good as real pread(). 843 if (lseek(pair->src_fd, pos, SEEK_SET) != pos) { 844 message_error(_("%s: Error seeking the file: %s"), 845 pair->src_name, strerror(errno)); 846 return true; 847 } 848 849 const size_t amount = io_read(pair, buf, size); 850 if (amount == SIZE_MAX) 851 return true; 852 853 if (amount != size) { 854 message_error(_("%s: Unexpected end of file"), 855 pair->src_name); 856 return true; 857 } 858 859 return false; 860 } 861 862 863 static bool 864 is_sparse(const io_buf *buf) 865 { 866 assert(IO_BUFFER_SIZE % sizeof(uint64_t) == 0); 867 868 for (size_t i = 0; i < ARRAY_SIZE(buf->u64); ++i) 869 if (buf->u64[i] != 0) 870 return false; 871 872 return true; 873 } 874 875 876 static bool 877 io_write_buf(file_pair *pair, const uint8_t *buf, size_t size) 878 { 879 assert(size < SSIZE_MAX); 880 881 while (size > 0) { 882 const ssize_t amount = write(pair->dest_fd, buf, size); 883 if (amount == -1) { 884 if (errno == EINTR) { 885 if (user_abort) 886 return -1; 887 888 continue; 889 } 890 891 // Handle broken pipe specially. gzip and bzip2 892 // don't print anything on SIGPIPE. In addition, 893 // gzip --quiet uses exit status 2 (warning) on 894 // broken pipe instead of whatever raise(SIGPIPE) 895 // would make it return. It is there to hide "Broken 896 // pipe" message on some old shells (probably old 897 // GNU bash). 898 // 899 // We don't do anything special with --quiet, which 900 // is what bzip2 does too. If we get SIGPIPE, we 901 // will handle it like other signals by setting 902 // user_abort, and get EPIPE here. 903 if (errno != EPIPE) 904 message_error(_("%s: Write error: %s"), 905 pair->dest_name, strerror(errno)); 906 907 return true; 908 } 909 910 buf += (size_t)(amount); 911 size -= (size_t)(amount); 912 } 913 914 return false; 915 } 916 917 918 extern bool 919 io_write(file_pair *pair, const io_buf *buf, size_t size) 920 { 921 assert(size <= IO_BUFFER_SIZE); 922 923 if (pair->dest_try_sparse) { 924 // Check if the block is sparse (contains only zeros). If it 925 // sparse, we just store the amount and return. We will take 926 // care of actually skipping over the hole when we hit the 927 // next data block or close the file. 928 // 929 // Since io_close() requires that dest_pending_sparse > 0 930 // if the file ends with sparse block, we must also return 931 // if size == 0 to avoid doing the lseek(). 932 if (size == IO_BUFFER_SIZE) { 933 if (is_sparse(buf)) { 934 pair->dest_pending_sparse += size; 935 return false; 936 } 937 } else if (size == 0) { 938 return false; 939 } 940 941 // This is not a sparse block. If we have a pending hole, 942 // skip it now. 943 if (pair->dest_pending_sparse > 0) { 944 if (lseek(pair->dest_fd, pair->dest_pending_sparse, 945 SEEK_CUR) == -1) { 946 message_error(_("%s: Seeking failed when " 947 "trying to create a sparse " 948 "file: %s"), pair->dest_name, 949 strerror(errno)); 950 return true; 951 } 952 953 pair->dest_pending_sparse = 0; 954 } 955 } 956 957 return io_write_buf(pair, buf->u8, size); 958 } 959