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