src/basic/fs-util.c

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   2
   3 #include <errno.h>
   4 #include <stddef.h>
   5 #include <stdlib.h>
   6 #include <sys/file.h>
   7 #include <linux/falloc.h>
   8 #include <linux/magic.h>
   9 #include <unistd.h>
  10
  11 #include "alloc-util.h"
  12 #include "dirent-util.h"
  13 #include "fd-util.h"
  14 #include "fileio.h"
  15 #include "fs-util.h"
  16 #include "hostname-util.h"
  17 #include "lock-util.h"
  18 #include "log.h"
  19 #include "macro.h"
  20 #include "missing_fcntl.h"
  21 #include "missing_fs.h"
  22 #include "missing_syscall.h"
  23 #include "mkdir.h"
  24 #include "parse-util.h"
  25 #include "path-util.h"
  26 #include "process-util.h"
  27 #include "random-util.h"
  28 #include "ratelimit.h"
  29 #include "stat-util.h"
  30 #include "stdio-util.h"
  31 #include "string-util.h"
  32 #include "strv.h"
  33 #include "time-util.h"
  34 #include "tmpfile-util.h"
  35 #include "umask-util.h"
  36 #include "user-util.h"
  37
  38 int rmdir_parents(const char *path, const char *stop) {
  39         char *p;
  40         int r;
  41
  42         assert(path);
  43         assert(stop);
  44
  45         if (!path_is_safe(path))
  46                 return -EINVAL;
  47
  48         if (!path_is_safe(stop))
  49                 return -EINVAL;
  50
  51         p = strdupa_safe(path);
  52
  53         for (;;) {
  54                 char *slash = NULL;
  55
  56                 /* skip the last component. */
  57                 r = path_find_last_component(p, /* accept_dot_dot= */ false, (const char **) &slash, NULL);
  58                 if (r <= 0)
  59                         return r;
  60                 if (slash == p)
  61                         return 0;
  62
  63                 assert(*slash == '/');
  64                 *slash = '\0';
  65
  66                 if (path_startswith_full(stop, p, /* accept_dot_dot= */ false))
  67                         return 0;
  68
  69                 if (rmdir(p) < 0 && errno != ENOENT)
  70                         return -errno;
  71         }
  72 }
  73
  74 int rename_noreplace(int olddirfd, const char *oldpath, int newdirfd, const char *newpath) {
  75         int r;
  76
  77         /* Try the ideal approach first */
  78         if (renameat2(olddirfd, oldpath, newdirfd, newpath, RENAME_NOREPLACE) >= 0)
  79                 return 0;
  80
  81         /* renameat2() exists since Linux 3.15, btrfs and FAT added support for it later. If it is not implemented,
  82          * fall back to a different method. */
  83         if (!ERRNO_IS_NOT_SUPPORTED(errno) && errno != EINVAL)
  84                 return -errno;
  85
  86         /* Let's try to use linkat()+unlinkat() as fallback. This doesn't work on directories and on some file systems
  87          * that do not support hard links (such as FAT, most prominently), but for files it's pretty close to what we
  88          * want — though not atomic (i.e. for a short period both the new and the old filename will exist). */
  89         if (linkat(olddirfd, oldpath, newdirfd, newpath, 0) >= 0) {
  90
  91                 r = RET_NERRNO(unlinkat(olddirfd, oldpath, 0));
  92                 if (r < 0) {
  93                         (void) unlinkat(newdirfd, newpath, 0);
  94                         return r;
  95                 }
  96
  97                 return 0;
  98         }
  99
 100         if (!ERRNO_IS_NOT_SUPPORTED(errno) && !IN_SET(errno, EINVAL, EPERM)) /* FAT returns EPERM on link()… */
 101                 return -errno;
 102
 103         /* OK, neither RENAME_NOREPLACE nor linkat()+unlinkat() worked. Let's then fall back to the racy TOCTOU
 104          * vulnerable accessat(F_OK) check followed by classic, replacing renameat(), we have nothing better. */
 105
 106         if (faccessat(newdirfd, newpath, F_OK, AT_SYMLINK_NOFOLLOW) >= 0)
 107                 return -EEXIST;
 108         if (errno != ENOENT)
 109                 return -errno;
 110
 111         return RET_NERRNO(renameat(olddirfd, oldpath, newdirfd, newpath));
 112 }
 113
 114 int readlinkat_malloc(int fd, const char *p, char **ret) {
 115         size_t l = PATH_MAX;
 116
 117         assert(p);
 118
 119         for (;;) {
 120                 _cleanup_free_ char *c = NULL;
 121                 ssize_t n;
 122
 123                 c = new(char, l+1);
 124                 if (!c)
 125                         return -ENOMEM;
 126
 127                 n = readlinkat(fd, p, c, l);
 128                 if (n < 0)
 129                         return -errno;
 130
 131                 if ((size_t) n < l) {
 132                         c[n] = 0;
 133
 134                         if (ret)
 135                                 *ret = TAKE_PTR(c);
 136
 137                         return 0;
 138                 }
 139
 140                 if (l > (SSIZE_MAX-1)/2) /* readlinkat() returns an ssize_t, and we want an extra byte for a
 141                                           * trailing NUL, hence do an overflow check relative to SSIZE_MAX-1
 142                                           * here */
 143                         return -EFBIG;
 144
 145                 l *= 2;
 146         }
 147 }
 148
 149 int readlink_malloc(const char *p, char **ret) {
 150         return readlinkat_malloc(AT_FDCWD, p, ret);
 151 }
 152
 153 int readlink_value(const char *p, char **ret) {
 154         _cleanup_free_ char *link = NULL, *name = NULL;
 155         int r;
 156
 157         assert(p);
 158         assert(ret);
 159
 160         r = readlink_malloc(p, &link);
 161         if (r < 0)
 162                 return r;
 163
 164         r = path_extract_filename(link, &name);
 165         if (r < 0)
 166                 return r;
 167         if (r == O_DIRECTORY)
 168                 return -EINVAL;
 169
 170         *ret = TAKE_PTR(name);
 171         return 0;
 172 }
 173
 174 int readlink_and_make_absolute(const char *p, char **ret) {
 175         _cleanup_free_ char *target = NULL;
 176         int r;
 177
 178         assert(p);
 179         assert(ret);
 180
 181         r = readlink_malloc(p, &target);
 182         if (r < 0)
 183                 return r;
 184
 185         return file_in_same_dir(p, target, ret);
 186 }
 187
 188 int chmod_and_chown_at(int dir_fd, const char *path, mode_t mode, uid_t uid, gid_t gid) {
 189         _cleanup_close_ int fd = -EBADF;
 190
 191         assert(dir_fd >= 0 || dir_fd == AT_FDCWD);
 192
 193         if (path) {
 194                 /* Let's acquire an O_PATH fd, as precaution to change mode/owner on the same file */
 195                 fd = openat(dir_fd, path, O_PATH|O_CLOEXEC|O_NOFOLLOW);
 196                 if (fd < 0)
 197                         return -errno;
 198                 dir_fd = fd;
 199
 200         } else if (dir_fd == AT_FDCWD) {
 201                 /* Let's acquire an O_PATH fd of the current directory */
 202                 fd = openat(dir_fd, ".", O_PATH|O_CLOEXEC|O_NOFOLLOW|O_DIRECTORY);
 203                 if (fd < 0)
 204                         return -errno;
 205                 dir_fd = fd;
 206         }
 207
 208         return fchmod_and_chown(dir_fd, mode, uid, gid);
 209 }
 210
 211 int fchmod_and_chown_with_fallback(int fd, const char *path, mode_t mode, uid_t uid, gid_t gid) {
 212         bool do_chown, do_chmod;
 213         struct stat st;
 214         int r;
 215
 216         /* Change ownership and access mode of the specified fd. Tries to do so safely, ensuring that at no
 217          * point in time the access mode is above the old access mode under the old ownership or the new
 218          * access mode under the new ownership. Note: this call tries hard to leave the access mode
 219          * unaffected if the uid/gid is changed, i.e. it undoes implicit suid/sgid dropping the kernel does
 220          * on chown().
 221          *
 222          * This call is happy with O_PATH fds.
 223          *
 224          * If path is given, allow a fallback path which does not use /proc/self/fd/. On any normal system
 225          * /proc will be mounted, but in certain improperly assembled environments it might not be. This is
 226          * less secure (potential TOCTOU), so should only be used after consideration. */
 227
 228         if (fstat(fd, &st) < 0)
 229                 return -errno;
 230
 231         do_chown =
 232                 (uid != UID_INVALID && st.st_uid != uid) ||
 233                 (gid != GID_INVALID && st.st_gid != gid);
 234
 235         do_chmod =
 236                 !S_ISLNK(st.st_mode) && /* chmod is not defined on symlinks */
 237                 ((mode != MODE_INVALID && ((st.st_mode ^ mode) & 07777) != 0) ||
 238                  do_chown); /* If we change ownership, make sure we reset the mode afterwards, since chown()
 239                              * modifies the access mode too */
 240
 241         if (mode == MODE_INVALID)
 242                 mode = st.st_mode; /* If we only shall do a chown(), save original mode, since chown() might break it. */
 243         else if ((mode & S_IFMT) != 0 && ((mode ^ st.st_mode) & S_IFMT) != 0)
 244                 return -EINVAL; /* insist on the right file type if it was specified */
 245
 246         if (do_chown && do_chmod) {
 247                 mode_t minimal = st.st_mode & mode; /* the subset of the old and the new mask */
 248
 249                 if (((minimal ^ st.st_mode) & 07777) != 0) {
 250                         r = fchmod_opath(fd, minimal & 07777);
 251                         if (r < 0) {
 252                                 if (!path || r != -ENOSYS)
 253                                         return r;
 254
 255                                 /* Fallback path which doesn't use /proc/self/fd/. */
 256                                 if (chmod(path, minimal & 07777) < 0)
 257                                         return -errno;
 258                         }
 259                 }
 260         }
 261
 262         if (do_chown)
 263                 if (fchownat(fd, "", uid, gid, AT_EMPTY_PATH) < 0)
 264                         return -errno;
 265
 266         if (do_chmod) {
 267                 r = fchmod_opath(fd, mode & 07777);
 268                 if (r < 0) {
 269                         if (!path || r != -ENOSYS)
 270                                 return r;
 271
 272                         /* Fallback path which doesn't use /proc/self/fd/. */
 273                         if (chmod(path, mode & 07777) < 0)
 274                                 return -errno;
 275                 }
 276         }
 277
 278         return do_chown || do_chmod;
 279 }
 280
 281 int fchmod_umask(int fd, mode_t m) {
 282         _cleanup_umask_ mode_t u = umask(0777);
 283
 284         return RET_NERRNO(fchmod(fd, m & (~u)));
 285 }
 286
 287 int fchmod_opath(int fd, mode_t m) {
 288         /* This function operates also on fd that might have been opened with
 289          * O_PATH. Indeed fchmodat() doesn't have the AT_EMPTY_PATH flag like
 290          * fchownat() does. */
 291
 292         if (chmod(FORMAT_PROC_FD_PATH(fd), m) < 0) {
 293                 if (errno != ENOENT)
 294                         return -errno;
 295
 296                 if (proc_mounted() == 0)
 297                         return -ENOSYS; /* if we have no /proc/, the concept is not implementable */
 298
 299                 return -ENOENT;
 300         }
 301
 302         return 0;
 303 }
 304
 305 int futimens_opath(int fd, const struct timespec ts[2]) {
 306         /* Similar to fchmod_path() but for futimens() */
 307
 308         if (utimensat(AT_FDCWD, FORMAT_PROC_FD_PATH(fd), ts, 0) < 0) {
 309                 if (errno != ENOENT)
 310                         return -errno;
 311
 312                 if (proc_mounted() == 0)
 313                         return -ENOSYS; /* if we have no /proc/, the concept is not implementable */
 314
 315                 return -ENOENT;
 316         }
 317
 318         return 0;
 319 }
 320
 321 int stat_warn_permissions(const char *path, const struct stat *st) {
 322         assert(path);
 323         assert(st);
 324
 325         /* Don't complain if we are reading something that is not a file, for example /dev/null */
 326         if (!S_ISREG(st->st_mode))
 327                 return 0;
 328
 329         if (st->st_mode & 0111)
 330                 log_warning("Configuration file %s is marked executable. Please remove executable permission bits. Proceeding anyway.", path);
 331
 332         if (st->st_mode & 0002)
 333                 log_warning("Configuration file %s is marked world-writable. Please remove world writability permission bits. Proceeding anyway.", path);
 334
 335         if (getpid_cached() == 1 && (st->st_mode & 0044) != 0044)
 336                 log_warning("Configuration file %s is marked world-inaccessible. This has no effect as configuration data is accessible via APIs without restrictions. Proceeding anyway.", path);
 337
 338         return 0;
 339 }
 340
 341 int fd_warn_permissions(const char *path, int fd) {
 342         struct stat st;
 343
 344         assert(path);
 345         assert(fd >= 0);
 346
 347         if (fstat(fd, &st) < 0)
 348                 return -errno;
 349
 350         return stat_warn_permissions(path, &st);
 351 }
 352
 353 int touch_file(const char *path, bool parents, usec_t stamp, uid_t uid, gid_t gid, mode_t mode) {
 354         _cleanup_close_ int fd = -EBADF;
 355         int r, ret;
 356
 357         assert(path);
 358
 359         /* Note that touch_file() does not follow symlinks: if invoked on an existing symlink, then it is the symlink
 360          * itself which is updated, not its target
 361          *
 362          * Returns the first error we encounter, but tries to apply as much as possible. */
 363
 364         if (parents)
 365                 (void) mkdir_parents(path, 0755);
 366
 367         /* Initially, we try to open the node with O_PATH, so that we get a reference to the node. This is useful in
 368          * case the path refers to an existing device or socket node, as we can open it successfully in all cases, and
 369          * won't trigger any driver magic or so. */
 370         fd = open(path, O_PATH|O_CLOEXEC|O_NOFOLLOW);
 371         if (fd < 0) {
 372                 if (errno != ENOENT)
 373                         return -errno;
 374
 375                 /* if the node doesn't exist yet, we create it, but with O_EXCL, so that we only create a regular file
 376                  * here, and nothing else */
 377                 fd = open(path, O_WRONLY|O_CREAT|O_EXCL|O_CLOEXEC, IN_SET(mode, 0, MODE_INVALID) ? 0644 : mode);
 378                 if (fd < 0)
 379                         return -errno;
 380         }
 381
 382         /* Let's make a path from the fd, and operate on that. With this logic, we can adjust the access mode,
 383          * ownership and time of the file node in all cases, even if the fd refers to an O_PATH object — which is
 384          * something fchown(), fchmod(), futimensat() don't allow. */
 385         ret = fchmod_and_chown(fd, mode, uid, gid);
 386
 387         if (stamp != USEC_INFINITY) {
 388                 struct timespec ts[2];
 389
 390                 timespec_store(&ts[0], stamp);
 391                 ts[1] = ts[0];
 392                 r = futimens_opath(fd, ts);
 393         } else
 394                 r = futimens_opath(fd, NULL);
 395         if (r < 0 && ret >= 0)
 396                 return r;
 397
 398         return ret;
 399 }
 400
 401 int symlink_idempotent(const char *from, const char *to, bool make_relative) {
 402         _cleanup_free_ char *relpath = NULL;
 403         int r;
 404
 405         assert(from);
 406         assert(to);
 407
 408         if (make_relative) {
 409                 r = path_make_relative_parent(to, from, &relpath);
 410                 if (r < 0)
 411                         return r;
 412
 413                 from = relpath;
 414         }
 415
 416         if (symlink(from, to) < 0) {
 417                 _cleanup_free_ char *p = NULL;
 418
 419                 if (errno != EEXIST)
 420                         return -errno;
 421
 422                 r = readlink_malloc(to, &p);
 423                 if (r == -EINVAL) /* Not a symlink? In that case return the original error we encountered: -EEXIST */
 424                         return -EEXIST;
 425                 if (r < 0) /* Any other error? In that case propagate it as is */
 426                         return r;
 427
 428                 if (!streq(p, from)) /* Not the symlink we want it to be? In that case, propagate the original -EEXIST */
 429                         return -EEXIST;
 430         }
 431
 432         return 0;
 433 }
 434
 435 int symlinkat_atomic_full(const char *from, int atfd, const char *to, bool make_relative) {
 436         _cleanup_free_ char *relpath = NULL, *t = NULL;
 437         int r;
 438
 439         assert(from);
 440         assert(to);
 441
 442         if (make_relative) {
 443                 r = path_make_relative_parent(to, from, &relpath);
 444                 if (r < 0)
 445                         return r;
 446
 447                 from = relpath;
 448         }
 449
 450         r = tempfn_random(to, NULL, &t);
 451         if (r < 0)
 452                 return r;
 453
 454         if (symlinkat(from, atfd, t) < 0)
 455                 return -errno;
 456
 457         r = RET_NERRNO(renameat(atfd, t, atfd, to));
 458         if (r < 0) {
 459                 (void) unlinkat(atfd, t, 0);
 460                 return r;
 461         }
 462
 463         return 0;
 464 }
 465
 466 int mknodat_atomic(int atfd, const char *path, mode_t mode, dev_t dev) {
 467         _cleanup_free_ char *t = NULL;
 468         int r;
 469
 470         assert(path);
 471
 472         r = tempfn_random(path, NULL, &t);
 473         if (r < 0)
 474                 return r;
 475
 476         if (mknodat(atfd, t, mode, dev) < 0)
 477                 return -errno;
 478
 479         r = RET_NERRNO(renameat(atfd, t, atfd, path));
 480         if (r < 0) {
 481                 (void) unlinkat(atfd, t, 0);
 482                 return r;
 483         }
 484
 485         return 0;
 486 }
 487
 488 int mkfifoat_atomic(int atfd, const char *path, mode_t mode) {
 489         _cleanup_free_ char *t = NULL;
 490         int r;
 491
 492         assert(path);
 493
 494         /* We're only interested in the (random) filename.  */
 495         r = tempfn_random(path, NULL, &t);
 496         if (r < 0)
 497                 return r;
 498
 499         if (mkfifoat(atfd, t, mode) < 0)
 500                 return -errno;
 501
 502         r = RET_NERRNO(renameat(atfd, t, atfd, path));
 503         if (r < 0) {
 504                 (void) unlinkat(atfd, t, 0);
 505                 return r;
 506         }
 507
 508         return 0;
 509 }
 510
 511 int get_files_in_directory(const char *path, char ***list) {
 512         _cleanup_strv_free_ char **l = NULL;
 513         _cleanup_closedir_ DIR *d = NULL;
 514         size_t n = 0;
 515
 516         assert(path);
 517
 518         /* Returns all files in a directory in *list, and the number
 519          * of files as return value. If list is NULL returns only the
 520          * number. */
 521
 522         d = opendir(path);
 523         if (!d)
 524                 return -errno;
 525
 526         FOREACH_DIRENT_ALL(de, d, return -errno) {
 527                 if (!dirent_is_file(de))
 528                         continue;
 529
 530                 if (list) {
 531                         /* one extra slot is needed for the terminating NULL */
 532                         if (!GREEDY_REALLOC(l, n + 2))
 533                                 return -ENOMEM;
 534
 535                         l[n] = strdup(de->d_name);
 536                         if (!l[n])
 537                                 return -ENOMEM;
 538
 539                         l[++n] = NULL;
 540                 } else
 541                         n++;
 542         }
 543
 544         if (list)
 545                 *list = TAKE_PTR(l);
 546
 547         return n;
 548 }
 549
 550 static int getenv_tmp_dir(const char **ret_path) {
 551         int r, ret = 0;
 552
 553         assert(ret_path);
 554
 555         /* We use the same order of environment variables python uses in tempfile.gettempdir():
 556          * https://docs.python.org/3/library/tempfile.html#tempfile.gettempdir */
 557         FOREACH_STRING(n, "TMPDIR", "TEMP", "TMP") {
 558                 const char *e;
 559
 560                 e = secure_getenv(n);
 561                 if (!e)
 562                         continue;
 563                 if (!path_is_absolute(e)) {
 564                         r = -ENOTDIR;
 565                         goto next;
 566                 }
 567                 if (!path_is_normalized(e)) {
 568                         r = -EPERM;
 569                         goto next;
 570                 }
 571
 572                 r = is_dir(e, true);
 573                 if (r < 0)
 574                         goto next;
 575                 if (r == 0) {
 576                         r = -ENOTDIR;
 577                         goto next;
 578                 }
 579
 580                 *ret_path = e;
 581                 return 1;
 582
 583         next:
 584                 /* Remember first error, to make this more debuggable */
 585                 if (ret >= 0)
 586                         ret = r;
 587         }
 588
 589         if (ret < 0)
 590                 return ret;
 591
 592         *ret_path = NULL;
 593         return ret;
 594 }
 595
 596 static int tmp_dir_internal(const char *def, const char **ret) {
 597         const char *e;
 598         int r, k;
 599
 600         assert(def);
 601         assert(ret);
 602
 603         r = getenv_tmp_dir(&e);
 604         if (r > 0) {
 605                 *ret = e;
 606                 return 0;
 607         }
 608
 609         k = is_dir(def, true);
 610         if (k == 0)
 611                 k = -ENOTDIR;
 612         if (k < 0)
 613                 return r < 0 ? r : k;
 614
 615         *ret = def;
 616         return 0;
 617 }
 618
 619 int var_tmp_dir(const char **ret) {
 620
 621         /* Returns the location for "larger" temporary files, that is backed by physical storage if available, and thus
 622          * even might survive a boot: /var/tmp. If $TMPDIR (or related environment variables) are set, its value is
 623          * returned preferably however. Note that both this function and tmp_dir() below are affected by $TMPDIR,
 624          * making it a variable that overrides all temporary file storage locations. */
 625
 626         return tmp_dir_internal("/var/tmp", ret);
 627 }
 628
 629 int tmp_dir(const char **ret) {
 630
 631         /* Similar to var_tmp_dir() above, but returns the location for "smaller" temporary files, which is usually
 632          * backed by an in-memory file system: /tmp. */
 633
 634         return tmp_dir_internal("/tmp", ret);
 635 }
 636
 637 int unlink_or_warn(const char *filename) {
 638         if (unlink(filename) < 0 && errno != ENOENT)
 639                 /* If the file doesn't exist and the fs simply was read-only (in which
 640                  * case unlink() returns EROFS even if the file doesn't exist), don't
 641                  * complain */
 642                 if (errno != EROFS || access(filename, F_OK) >= 0)
 643                         return log_error_errno(errno, "Failed to remove \"%s\": %m", filename);
 644
 645         return 0;
 646 }
 647
 648 int access_fd(int fd, int mode) {
 649         /* Like access() but operates on an already open fd */
 650
 651         if (access(FORMAT_PROC_FD_PATH(fd), mode) < 0) {
 652                 if (errno != ENOENT)
 653                         return -errno;
 654
 655                 /* ENOENT can mean two things: that the fd does not exist or that /proc is not mounted. Let's
 656                  * make things debuggable and distinguish the two. */
 657
 658                 if (proc_mounted() == 0)
 659                         return -ENOSYS;  /* /proc is not available or not set up properly, we're most likely in some chroot
 660                                           * environment. */
 661
 662                 return -EBADF; /* The directory exists, hence it's the fd that doesn't. */
 663         }
 664
 665         return 0;
 666 }
 667
 668 void unlink_tempfilep(char (*p)[]) {
 669         /* If the file is created with mkstemp(), it will (almost always)
 670          * change the suffix. Treat this as a sign that the file was
 671          * successfully created. We ignore both the rare case where the
 672          * original suffix is used and unlink failures. */
 673         if (!endswith(*p, ".XXXXXX"))
 674                 (void) unlink(*p);
 675 }
 676
 677 int unlinkat_deallocate(int fd, const char *name, UnlinkDeallocateFlags flags) {
 678         _cleanup_close_ int truncate_fd = -EBADF;
 679         struct stat st;
 680         off_t l, bs;
 681
 682         assert((flags & ~(UNLINK_REMOVEDIR|UNLINK_ERASE)) == 0);
 683
 684         /* Operates like unlinkat() but also deallocates the file contents if it is a regular file and there's no other
 685          * link to it. This is useful to ensure that other processes that might have the file open for reading won't be
 686          * able to keep the data pinned on disk forever. This call is particular useful whenever we execute clean-up
 687          * jobs ("vacuuming"), where we want to make sure the data is really gone and the disk space released and
 688          * returned to the free pool.
 689          *
 690          * Deallocation is preferably done by FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE (👊) if supported, which means
 691          * the file won't change size. That's a good thing since we shouldn't needlessly trigger SIGBUS in other
 692          * programs that have mmap()ed the file. (The assumption here is that changing file contents to all zeroes
 693          * underneath those programs is the better choice than simply triggering SIGBUS in them which truncation does.)
 694          * However if hole punching is not implemented in the kernel or file system we'll fall back to normal file
 695          * truncation (🔪), as our goal of deallocating the data space trumps our goal of being nice to readers (💐).
 696          *
 697          * Note that we attempt deallocation, but failure to succeed with that is not considered fatal, as long as the
 698          * primary job – to delete the file – is accomplished. */
 699
 700         if (!FLAGS_SET(flags, UNLINK_REMOVEDIR)) {
 701                 truncate_fd = openat(fd, name, O_WRONLY|O_CLOEXEC|O_NOCTTY|O_NOFOLLOW|O_NONBLOCK);
 702                 if (truncate_fd < 0) {
 703
 704                         /* If this failed because the file doesn't exist propagate the error right-away. Also,
 705                          * AT_REMOVEDIR wasn't set, and we tried to open the file for writing, which means EISDIR is
 706                          * returned when this is a directory but we are not supposed to delete those, hence propagate
 707                          * the error right-away too. */
 708                         if (IN_SET(errno, ENOENT, EISDIR))
 709                                 return -errno;
 710
 711                         if (errno != ELOOP) /* don't complain if this is a symlink */
 712                                 log_debug_errno(errno, "Failed to open file '%s' for deallocation, ignoring: %m", name);
 713                 }
 714         }
 715
 716         if (unlinkat(fd, name, FLAGS_SET(flags, UNLINK_REMOVEDIR) ? AT_REMOVEDIR : 0) < 0)
 717                 return -errno;
 718
 719         if (truncate_fd < 0) /* Don't have a file handle, can't do more ☹️ */
 720                 return 0;
 721
 722         if (fstat(truncate_fd, &st) < 0) {
 723                 log_debug_errno(errno, "Failed to stat file '%s' for deallocation, ignoring: %m", name);
 724                 return 0;
 725         }
 726
 727         if (!S_ISREG(st.st_mode))
 728                 return 0;
 729
 730         if (FLAGS_SET(flags, UNLINK_ERASE) && st.st_size > 0 && st.st_nlink == 0) {
 731                 uint64_t left = st.st_size;
 732                 char buffer[64 * 1024];
 733
 734                 /* If erasing is requested, let's overwrite the file with random data once before deleting
 735                  * it. This isn't going to give you shred(1) semantics, but hopefully should be good enough
 736                  * for stuff backed by tmpfs at least.
 737                  *
 738                  * Note that we only erase like this if the link count of the file is zero. If it is higher it
 739                  * is still linked by someone else and we'll leave it to them to remove it securely
 740                  * eventually! */
 741
 742                 random_bytes(buffer, sizeof(buffer));
 743
 744                 while (left > 0) {
 745                         ssize_t n;
 746
 747                         n = write(truncate_fd, buffer, MIN(sizeof(buffer), left));
 748                         if (n < 0) {
 749                                 log_debug_errno(errno, "Failed to erase data in file '%s', ignoring.", name);
 750                                 break;
 751                         }
 752
 753                         assert(left >= (size_t) n);
 754                         left -= n;
 755                 }
 756
 757                 /* Let's refresh metadata */
 758                 if (fstat(truncate_fd, &st) < 0) {
 759                         log_debug_errno(errno, "Failed to stat file '%s' for deallocation, ignoring: %m", name);
 760                         return 0;
 761                 }
 762         }
 763
 764         /* Don't dallocate if there's nothing to deallocate or if the file is linked elsewhere */
 765         if (st.st_blocks == 0 || st.st_nlink > 0)
 766                 return 0;
 767
 768         /* If this is a regular file, it actually took up space on disk and there are no other links it's time to
 769          * punch-hole/truncate this to release the disk space. */
 770
 771         bs = MAX(st.st_blksize, 512);
 772         l = DIV_ROUND_UP(st.st_size, bs) * bs; /* Round up to next block size */
 773
 774         if (fallocate(truncate_fd, FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE, 0, l) >= 0)
 775                 return 0; /* Successfully punched a hole! 😊 */
 776
 777         /* Fall back to truncation */
 778         if (ftruncate(truncate_fd, 0) < 0) {
 779                 log_debug_errno(errno, "Failed to truncate file to 0, ignoring: %m");
 780                 return 0;
 781         }
 782
 783         return 0;
 784 }
 785
 786 int open_parent_at(int dir_fd, const char *path, int flags, mode_t mode) {
 787         _cleanup_free_ char *parent = NULL;
 788         int r;
 789
 790         assert(dir_fd >= 0 || dir_fd == AT_FDCWD);
 791         assert(path);
 792
 793         r = path_extract_directory(path, &parent);
 794         if (r == -EDESTADDRREQ) {
 795                 parent = strdup(".");
 796                 if (!parent)
 797                         return -ENOMEM;
 798         } else if (r == -EADDRNOTAVAIL) {
 799                 parent = strdup(path);
 800                 if (!parent)
 801                         return -ENOMEM;
 802         } else if (r < 0)
 803                 return r;
 804
 805         /* Let's insist on O_DIRECTORY since the parent of a file or directory is a directory. Except if we open an
 806          * O_TMPFILE file, because in that case we are actually create a regular file below the parent directory. */
 807
 808         if (FLAGS_SET(flags, O_PATH))
 809                 flags |= O_DIRECTORY;
 810         else if (!FLAGS_SET(flags, O_TMPFILE))
 811                 flags |= O_DIRECTORY|O_RDONLY;
 812
 813         return RET_NERRNO(openat(dir_fd, parent, flags, mode));
 814 }
 815
 816 int conservative_renameat(
 817                 int olddirfd, const char *oldpath,
 818                 int newdirfd, const char *newpath) {
 819
 820         _cleanup_close_ int old_fd = -EBADF, new_fd = -EBADF;
 821         struct stat old_stat, new_stat;
 822
 823         /* Renames the old path to thew new path, much like renameat() — except if both are regular files and
 824          * have the exact same contents and basic file attributes already. In that case remove the new file
 825          * instead. This call is useful for reducing inotify wakeups on files that are updated but don't
 826          * actually change. This function is written in a style that we rather rename too often than suppress
 827          * too much. i.e. whenever we are in doubt we rather rename than fail. After all reducing inotify
 828          * events is an optimization only, not more. */
 829
 830         old_fd = openat(olddirfd, oldpath, O_CLOEXEC|O_RDONLY|O_NOCTTY|O_NOFOLLOW);
 831         if (old_fd < 0)
 832                 goto do_rename;
 833
 834         new_fd = openat(newdirfd, newpath, O_CLOEXEC|O_RDONLY|O_NOCTTY|O_NOFOLLOW);
 835         if (new_fd < 0)
 836                 goto do_rename;
 837
 838         if (fstat(old_fd, &old_stat) < 0)
 839                 goto do_rename;
 840
 841         if (!S_ISREG(old_stat.st_mode))
 842                 goto do_rename;
 843
 844         if (fstat(new_fd, &new_stat) < 0)
 845                 goto do_rename;
 846
 847         if (stat_inode_same(&new_stat, &old_stat))
 848                 goto is_same;
 849
 850         if (old_stat.st_mode != new_stat.st_mode ||
 851             old_stat.st_size != new_stat.st_size ||
 852             old_stat.st_uid != new_stat.st_uid ||
 853             old_stat.st_gid != new_stat.st_gid)
 854                 goto do_rename;
 855
 856         for (;;) {
 857                 uint8_t buf1[16*1024];
 858                 uint8_t buf2[sizeof(buf1)];
 859                 ssize_t l1, l2;
 860
 861                 l1 = read(old_fd, buf1, sizeof(buf1));
 862                 if (l1 < 0)
 863                         goto do_rename;
 864
 865                 if (l1 == sizeof(buf1))
 866                         /* Read the full block, hence read a full block in the other file too */
 867
 868                         l2 = read(new_fd, buf2, l1);
 869                 else {
 870                         assert((size_t) l1 < sizeof(buf1));
 871
 872                         /* Short read. This hence was the last block in the first file, and then came
 873                          * EOF. Read one byte more in the second file, so that we can verify we hit EOF there
 874                          * too. */
 875
 876                         assert((size_t) (l1 + 1) <= sizeof(buf2));
 877                         l2 = read(new_fd, buf2, l1 + 1);
 878                 }
 879                 if (l2 != l1)
 880                         goto do_rename;
 881
 882                 if (memcmp(buf1, buf2, l1) != 0)
 883                         goto do_rename;
 884
 885                 if ((size_t) l1 < sizeof(buf1)) /* We hit EOF on the first file, and the second file too, hence exit
 886                                                  * now. */
 887                         break;
 888         }
 889
 890 is_same:
 891         /* Everything matches? Then don't rename, instead remove the source file, and leave the existing
 892          * destination in place */
 893
 894         if (unlinkat(olddirfd, oldpath, 0) < 0)
 895                 goto do_rename;
 896
 897         return 0;
 898
 899 do_rename:
 900         if (renameat(olddirfd, oldpath, newdirfd, newpath) < 0)
 901                 return -errno;
 902
 903         return 1;
 904 }
 905
 906 int posix_fallocate_loop(int fd, uint64_t offset, uint64_t size) {
 907         RateLimit rl;
 908         int r;
 909
 910         r = posix_fallocate(fd, offset, size); /* returns positive errnos on error */
 911         if (r != EINTR)
 912                 return -r; /* Let's return negative errnos, like common in our codebase */
 913
 914         /* On EINTR try a couple of times more, but protect against busy looping
 915          * (not more than 16 times per 10s) */
 916         rl = (const RateLimit) { 10 * USEC_PER_SEC, 16 };
 917         while (ratelimit_below(&rl)) {
 918                 r = posix_fallocate(fd, offset, size);
 919                 if (r != EINTR)
 920                         return -r;
 921         }
 922
 923         return -EINTR;
 924 }
 925
 926 int parse_cifs_service(
 927                 const char *s,
 928                 char **ret_host,
 929                 char **ret_service,
 930                 char **ret_path) {
 931
 932         _cleanup_free_ char *h = NULL, *ss = NULL, *x = NULL;
 933         const char *p, *e, *d;
 934         char delimiter;
 935
 936         /* Parses a CIFS service in form of //host/service/path… and splitting it in three parts. The last
 937          * part is optional, in which case NULL is returned there. To maximize compatibility syntax with
 938          * backslashes instead of slashes is accepted too. */
 939
 940         if (!s)
 941                 return -EINVAL;
 942
 943         p = startswith(s, "//");
 944         if (!p) {
 945                 p = startswith(s, "\\\\");
 946                 if (!p)
 947                         return -EINVAL;
 948         }
 949
 950         delimiter = s[0];
 951         e = strchr(p, delimiter);
 952         if (!e)
 953                 return -EINVAL;
 954
 955         h = strndup(p, e - p);
 956         if (!h)
 957                 return -ENOMEM;
 958
 959         if (!hostname_is_valid(h, 0))
 960                 return -EINVAL;
 961
 962         e++;
 963
 964         d = strchrnul(e, delimiter);
 965
 966         ss = strndup(e, d - e);
 967         if (!ss)
 968                 return -ENOMEM;
 969
 970         if (!filename_is_valid(ss))
 971                 return -EINVAL;
 972
 973         if (!isempty(d)) {
 974                 x = strdup(skip_leading_chars(d, CHAR_TO_STR(delimiter)));
 975                 if (!x)
 976                         return -EINVAL;
 977
 978                 /* Make sure to convert Windows-style "\" → Unix-style / */
 979                 for (char *i = x; *i; i++)
 980                         if (*i == delimiter)
 981                                 *i = '/';
 982
 983                 if (!path_is_valid(x))
 984                         return -EINVAL;
 985
 986                 path_simplify(x);
 987                 if (!path_is_normalized(x))
 988                         return -EINVAL;
 989         }
 990
 991         if (ret_host)
 992                 *ret_host = TAKE_PTR(h);
 993         if (ret_service)
 994                 *ret_service = TAKE_PTR(ss);
 995         if (ret_path)
 996                 *ret_path = TAKE_PTR(x);
 997
 998         return 0;
 999 }
1000
1001 int open_mkdir_at(int dirfd, const char *path, int flags, mode_t mode) {
1002         _cleanup_close_ int fd = -EBADF, parent_fd = -EBADF;
1003         _cleanup_free_ char *fname = NULL;
1004         int r;
1005
1006         /* Creates a directory with mkdirat() and then opens it, in the "most atomic" fashion we can
1007          * do. Guarantees that the returned fd refers to a directory. If O_EXCL is specified will fail if the
1008          * dir already exists. Otherwise will open an existing dir, but only if it is one.  */
1009
1010         if (flags & ~(O_RDONLY|O_CLOEXEC|O_DIRECTORY|O_EXCL|O_NOATIME|O_NOFOLLOW|O_PATH))
1011                 return -EINVAL;
1012         if ((flags & O_ACCMODE) != O_RDONLY)
1013                 return -EINVAL;
1014
1015         /* Note that O_DIRECTORY|O_NOFOLLOW is implied, but we allow specifying it anyway. The following
1016          * flags actually make sense to specify: O_CLOEXEC, O_EXCL, O_NOATIME, O_PATH */
1017
1018         if (isempty(path))
1019                 return -EINVAL;
1020
1021         if (!filename_is_valid(path)) {
1022                 _cleanup_free_ char *parent = NULL;
1023
1024                 /* If this is not a valid filename, it's a path. Let's open the parent directory then, so
1025                  * that we can pin it, and operate below it. */
1026
1027                 r = path_extract_directory(path, &parent);
1028                 if (r < 0)
1029                         return r;
1030
1031                 r = path_extract_filename(path, &fname);
1032                 if (r < 0)
1033                         return r;
1034
1035                 parent_fd = openat(dirfd, parent, O_PATH|O_DIRECTORY|O_CLOEXEC);
1036                 if (parent_fd < 0)
1037                         return -errno;
1038
1039                 dirfd = parent_fd;
1040                 path = fname;
1041         }
1042
1043         fd = xopenat(dirfd, path, flags|O_CREAT|O_DIRECTORY|O_NOFOLLOW, mode);
1044         if (IN_SET(fd, -ELOOP, -ENOTDIR))
1045                 return -EEXIST;
1046         if (fd < 0)
1047                 return fd;
1048
1049         return TAKE_FD(fd);
1050 }
1051
1052 int openat_report_new(int dirfd, const char *pathname, int flags, mode_t mode, bool *ret_newly_created) {
1053         unsigned attempts = 7;
1054         int fd;
1055
1056         /* Just like openat(), but adds one thing: optionally returns whether we created the file anew or if
1057          * it already existed before. This is only relevant if O_CREAT is set without O_EXCL, and thus will
1058          * shortcut to openat() otherwise */
1059
1060         if (!ret_newly_created)
1061                 return RET_NERRNO(openat(dirfd, pathname, flags, mode));
1062
1063         if (!FLAGS_SET(flags, O_CREAT) || FLAGS_SET(flags, O_EXCL)) {
1064                 fd = openat(dirfd, pathname, flags, mode);
1065                 if (fd < 0)
1066                         return -errno;
1067
1068                 *ret_newly_created = FLAGS_SET(flags, O_CREAT);
1069                 return fd;
1070         }
1071
1072         for (;;) {
1073                 /* First, attempt to open without O_CREAT/O_EXCL, i.e. open existing file */
1074                 fd = openat(dirfd, pathname, flags & ~(O_CREAT | O_EXCL), mode);
1075                 if (fd >= 0) {
1076                         *ret_newly_created = false;
1077                         return fd;
1078                 }
1079                 if (errno != ENOENT)
1080                         return -errno;
1081
1082                 /* So the file didn't exist yet, hence create it with O_CREAT/O_EXCL. */
1083                 fd = openat(dirfd, pathname, flags | O_CREAT | O_EXCL, mode);
1084                 if (fd >= 0) {
1085                         *ret_newly_created = true;
1086                         return fd;
1087                 }
1088                 if (errno != EEXIST)
1089                         return -errno;
1090
1091                 /* Hmm, so now we got EEXIST? So it apparently exists now? If so, let's try to open again
1092                  * without the two flags. But let's not spin forever, hence put a limit on things */
1093
1094                 if (--attempts == 0) /* Give up eventually, somebody is playing with us */
1095                         return -EEXIST;
1096         }
1097 }
1098
1099 int xopenat(int dir_fd, const char *path, int flags, mode_t mode) {
1100         _cleanup_close_ int fd = -EBADF;
1101         bool made = false;
1102         int r;
1103
1104         assert(dir_fd >= 0 || dir_fd == AT_FDCWD);
1105         assert(path);
1106
1107         if (isempty(path)) {
1108                 assert(!FLAGS_SET(flags, O_CREAT|O_EXCL));
1109                 return fd_reopen(dir_fd, flags & ~O_NOFOLLOW);
1110         }
1111
1112         if (FLAGS_SET(flags, O_DIRECTORY|O_CREAT)) {
1113                 r = RET_NERRNO(mkdirat(dir_fd, path, mode));
1114                 if (r == -EEXIST) {
1115                         if (FLAGS_SET(flags, O_EXCL))
1116                                 return -EEXIST;
1117
1118                         made = false;
1119                 } else if (r < 0)
1120                         return r;
1121                 else
1122                         made = true;
1123
1124                 flags &= ~(O_EXCL|O_CREAT);
1125         }
1126
1127         fd = RET_NERRNO(openat(dir_fd, path, flags, mode));
1128         if (fd < 0) {
1129                 if (IN_SET(fd,
1130                            /* We got ENOENT? then someone else immediately removed it after we
1131                            * created it. In that case let's return immediately without unlinking
1132                            * anything, because there simply isn't anything to unlink anymore. */
1133                            -ENOENT,
1134                            /* is a symlink? exists already → created by someone else, don't unlink */
1135                            -ELOOP,
1136                            /* not a directory? exists already → created by someone else, don't unlink */
1137                            -ENOTDIR))
1138                         return fd;
1139
1140                 if (made)
1141                         (void) unlinkat(dir_fd, path, AT_REMOVEDIR);
1142
1143                 return fd;
1144         }
1145
1146         return TAKE_FD(fd);
1147 }
1148
1149 int xopenat_lock(int dir_fd, const char *path, int flags, mode_t mode, LockType locktype, int operation) {
1150         _cleanup_close_ int fd = -EBADF;
1151         int r;
1152
1153         assert(dir_fd >= 0 || dir_fd == AT_FDCWD);
1154         assert(path);
1155         assert(IN_SET(operation & ~LOCK_NB, LOCK_EX, LOCK_SH));
1156
1157         /* POSIX/UNPOSIX locks don't work on directories (errno is set to -EBADF so let's return early with
1158          * the same error here). */
1159         if (FLAGS_SET(flags, O_DIRECTORY) && locktype != LOCK_BSD)
1160                 return -EBADF;
1161
1162         for (;;) {
1163                 struct stat st;
1164
1165                 fd = xopenat(dir_fd, path, flags, mode);
1166                 if (fd < 0)
1167                         return fd;
1168
1169                 r = lock_generic(fd, locktype, operation);
1170                 if (r < 0)
1171                         return r;
1172
1173                 /* If we acquired the lock, let's check if the file/directory still exists in the file
1174                  * system. If not, then the previous exclusive owner removed it and then closed it. In such a
1175                  * case our acquired lock is worthless, hence try again. */
1176
1177                 if (fstat(fd, &st) < 0)
1178                         return -errno;
1179                 if (st.st_nlink > 0)
1180                         break;
1181
1182                 fd = safe_close(fd);
1183         }
1184
1185         return TAKE_FD(fd);
1186 }