]>
Commit | Line | Data |
---|---|---|
db9ecf05 | 1 | /* SPDX-License-Identifier: LGPL-2.1-or-later */ |
3ffd4af2 | 2 | |
11c3a366 TA |
3 | #include <errno.h> |
4 | #include <fcntl.h> | |
5 | #include <sys/resource.h> | |
11c3a366 TA |
6 | #include <sys/stat.h> |
7 | #include <unistd.h> | |
8 | ||
4960ce43 LP |
9 | #include "alloc-util.h" |
10 | #include "copy.h" | |
8fb3f009 | 11 | #include "dirent-util.h" |
3ffd4af2 | 12 | #include "fd-util.h" |
a548e14d | 13 | #include "fileio.h" |
4aeb20f5 | 14 | #include "fs-util.h" |
4960ce43 | 15 | #include "io-util.h" |
11c3a366 | 16 | #include "macro.h" |
a548e14d | 17 | #include "memfd-util.h" |
0499585f | 18 | #include "missing_fcntl.h" |
f5947a5e | 19 | #include "missing_syscall.h" |
93cc7779 | 20 | #include "parse-util.h" |
11c3a366 | 21 | #include "path-util.h" |
df0ff127 | 22 | #include "process-util.h" |
93cc7779 | 23 | #include "socket-util.h" |
b8cfa2da | 24 | #include "sort-util.h" |
f8606626 | 25 | #include "stat-util.h" |
4aeb20f5 | 26 | #include "stdio-util.h" |
e4de7287 | 27 | #include "tmpfile-util.h" |
f8606626 | 28 | #include "util.h" |
3ffd4af2 | 29 | |
6a461d1f ZJS |
30 | /* The maximum number of iterations in the loop to close descriptors in the fallback case |
31 | * when /proc/self/fd/ is inaccessible. */ | |
32 | #define MAX_FD_LOOP_LIMIT (1024*1024) | |
33 | ||
3ffd4af2 LP |
34 | int close_nointr(int fd) { |
35 | assert(fd >= 0); | |
36 | ||
37 | if (close(fd) >= 0) | |
38 | return 0; | |
39 | ||
40 | /* | |
41 | * Just ignore EINTR; a retry loop is the wrong thing to do on | |
42 | * Linux. | |
43 | * | |
44 | * http://lkml.indiana.edu/hypermail/linux/kernel/0509.1/0877.html | |
45 | * https://bugzilla.gnome.org/show_bug.cgi?id=682819 | |
46 | * http://utcc.utoronto.ca/~cks/space/blog/unix/CloseEINTR | |
47 | * https://sites.google.com/site/michaelsafyan/software-engineering/checkforeintrwheninvokingclosethinkagain | |
48 | */ | |
49 | if (errno == EINTR) | |
50 | return 0; | |
51 | ||
52 | return -errno; | |
53 | } | |
54 | ||
55 | int safe_close(int fd) { | |
56 | ||
57 | /* | |
58 | * Like close_nointr() but cannot fail. Guarantees errno is | |
59 | * unchanged. Is a NOP with negative fds passed, and returns | |
60 | * -1, so that it can be used in this syntax: | |
61 | * | |
62 | * fd = safe_close(fd); | |
63 | */ | |
64 | ||
65 | if (fd >= 0) { | |
66 | PROTECT_ERRNO; | |
67 | ||
68 | /* The kernel might return pretty much any error code | |
69 | * via close(), but the fd will be closed anyway. The | |
70 | * only condition we want to check for here is whether | |
71 | * the fd was invalid at all... */ | |
72 | ||
73 | assert_se(close_nointr(fd) != -EBADF); | |
74 | } | |
75 | ||
76 | return -1; | |
77 | } | |
78 | ||
3042bbeb | 79 | void safe_close_pair(int p[static 2]) { |
3ffd4af2 LP |
80 | assert(p); |
81 | ||
82 | if (p[0] == p[1]) { | |
83 | /* Special case pairs which use the same fd in both | |
84 | * directions... */ | |
85 | p[0] = p[1] = safe_close(p[0]); | |
86 | return; | |
87 | } | |
88 | ||
89 | p[0] = safe_close(p[0]); | |
90 | p[1] = safe_close(p[1]); | |
91 | } | |
92 | ||
da6053d0 | 93 | void close_many(const int fds[], size_t n_fd) { |
3ffd4af2 LP |
94 | assert(fds || n_fd <= 0); |
95 | ||
fe96c0f8 | 96 | for (size_t i = 0; i < n_fd; i++) |
3ffd4af2 LP |
97 | safe_close(fds[i]); |
98 | } | |
99 | ||
100 | int fclose_nointr(FILE *f) { | |
101 | assert(f); | |
102 | ||
103 | /* Same as close_nointr(), but for fclose() */ | |
104 | ||
75f6d5d8 LP |
105 | errno = 0; /* Extra safety: if the FILE* object is not encapsulating an fd, it might not set errno |
106 | * correctly. Let's hence initialize it to zero first, so that we aren't confused by any | |
107 | * prior errno here */ | |
3ffd4af2 LP |
108 | if (fclose(f) == 0) |
109 | return 0; | |
110 | ||
111 | if (errno == EINTR) | |
112 | return 0; | |
113 | ||
75f6d5d8 | 114 | return errno_or_else(EIO); |
3ffd4af2 LP |
115 | } |
116 | ||
117 | FILE* safe_fclose(FILE *f) { | |
118 | ||
119 | /* Same as safe_close(), but for fclose() */ | |
120 | ||
121 | if (f) { | |
122 | PROTECT_ERRNO; | |
123 | ||
6dce3bb4 | 124 | assert_se(fclose_nointr(f) != -EBADF); |
3ffd4af2 LP |
125 | } |
126 | ||
127 | return NULL; | |
128 | } | |
129 | ||
130 | DIR* safe_closedir(DIR *d) { | |
131 | ||
132 | if (d) { | |
133 | PROTECT_ERRNO; | |
134 | ||
135 | assert_se(closedir(d) >= 0 || errno != EBADF); | |
136 | } | |
137 | ||
138 | return NULL; | |
139 | } | |
140 | ||
141 | int fd_nonblock(int fd, bool nonblock) { | |
142 | int flags, nflags; | |
143 | ||
144 | assert(fd >= 0); | |
145 | ||
146 | flags = fcntl(fd, F_GETFL, 0); | |
147 | if (flags < 0) | |
148 | return -errno; | |
149 | ||
0da96503 | 150 | nflags = UPDATE_FLAG(flags, O_NONBLOCK, nonblock); |
3ffd4af2 LP |
151 | if (nflags == flags) |
152 | return 0; | |
153 | ||
154 | if (fcntl(fd, F_SETFL, nflags) < 0) | |
155 | return -errno; | |
156 | ||
157 | return 0; | |
158 | } | |
159 | ||
160 | int fd_cloexec(int fd, bool cloexec) { | |
161 | int flags, nflags; | |
162 | ||
163 | assert(fd >= 0); | |
164 | ||
165 | flags = fcntl(fd, F_GETFD, 0); | |
166 | if (flags < 0) | |
167 | return -errno; | |
168 | ||
0da96503 | 169 | nflags = UPDATE_FLAG(flags, FD_CLOEXEC, cloexec); |
3ffd4af2 LP |
170 | if (nflags == flags) |
171 | return 0; | |
172 | ||
173 | if (fcntl(fd, F_SETFD, nflags) < 0) | |
174 | return -errno; | |
175 | ||
176 | return 0; | |
177 | } | |
178 | ||
da6053d0 | 179 | _pure_ static bool fd_in_set(int fd, const int fdset[], size_t n_fdset) { |
3ffd4af2 LP |
180 | assert(n_fdset == 0 || fdset); |
181 | ||
fe96c0f8 | 182 | for (size_t i = 0; i < n_fdset; i++) |
3ffd4af2 LP |
183 | if (fdset[i] == fd) |
184 | return true; | |
185 | ||
186 | return false; | |
187 | } | |
188 | ||
498e265d LP |
189 | static int get_max_fd(void) { |
190 | struct rlimit rl; | |
191 | rlim_t m; | |
192 | ||
193 | /* Return the highest possible fd, based RLIMIT_NOFILE, but enforcing FD_SETSIZE-1 as lower boundary | |
194 | * and INT_MAX as upper boundary. */ | |
195 | ||
196 | if (getrlimit(RLIMIT_NOFILE, &rl) < 0) | |
197 | return -errno; | |
198 | ||
199 | m = MAX(rl.rlim_cur, rl.rlim_max); | |
200 | if (m < FD_SETSIZE) /* Let's always cover at least 1024 fds */ | |
201 | return FD_SETSIZE-1; | |
202 | ||
203 | if (m == RLIM_INFINITY || m > INT_MAX) /* Saturate on overflow. After all fds are "int", hence can | |
204 | * never be above INT_MAX */ | |
205 | return INT_MAX; | |
206 | ||
207 | return (int) (m - 1); | |
208 | } | |
209 | ||
da6053d0 | 210 | int close_all_fds(const int except[], size_t n_except) { |
b8cfa2da | 211 | static bool have_close_range = true; /* Assume we live in the future */ |
3ffd4af2 LP |
212 | _cleanup_closedir_ DIR *d = NULL; |
213 | struct dirent *de; | |
214 | int r = 0; | |
215 | ||
216 | assert(n_except == 0 || except); | |
217 | ||
b8cfa2da LP |
218 | if (have_close_range) { |
219 | /* In the best case we have close_range() to close all fds between a start and an end fd, | |
220 | * which we can use on the "inverted" exception array, i.e. all intervals between all | |
221 | * adjacent pairs from the sorted exception array. This changes loop complexity from O(n) | |
222 | * where n is number of open fds to O(m⋅log(m)) where m is the number of fds to keep | |
223 | * open. Given that we assume n ≫ m that's preferable to us. */ | |
224 | ||
225 | if (n_except == 0) { | |
226 | /* Close everything. Yay! */ | |
227 | ||
228 | if (close_range(3, -1, 0) >= 0) | |
229 | return 1; | |
230 | ||
231 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && !ERRNO_IS_PRIVILEGE(errno)) | |
232 | return -errno; | |
233 | ||
234 | have_close_range = false; | |
235 | } else { | |
236 | _cleanup_free_ int *sorted_malloc = NULL; | |
237 | size_t n_sorted; | |
238 | int *sorted; | |
239 | ||
240 | assert(n_except < SIZE_MAX); | |
241 | n_sorted = n_except + 1; | |
242 | ||
243 | if (n_sorted > 64) /* Use heap for large numbers of fds, stack otherwise */ | |
244 | sorted = sorted_malloc = new(int, n_sorted); | |
245 | else | |
246 | sorted = newa(int, n_sorted); | |
247 | ||
248 | if (sorted) { | |
249 | int c = 0; | |
250 | ||
251 | memcpy(sorted, except, n_except * sizeof(int)); | |
252 | ||
253 | /* Let's add fd 2 to the list of fds, to simplify the loop below, as this | |
254 | * allows us to cover the head of the array the same way as the body */ | |
255 | sorted[n_sorted-1] = 2; | |
256 | ||
257 | typesafe_qsort(sorted, n_sorted, cmp_int); | |
258 | ||
259 | for (size_t i = 0; i < n_sorted-1; i++) { | |
260 | int start, end; | |
261 | ||
262 | start = MAX(sorted[i], 2); /* The first three fds shall always remain open */ | |
263 | end = MAX(sorted[i+1], 2); | |
264 | ||
265 | assert(end >= start); | |
266 | ||
267 | if (end - start <= 1) | |
268 | continue; | |
269 | ||
270 | /* Close everything between the start and end fds (both of which shall stay open) */ | |
271 | if (close_range(start + 1, end - 1, 0) < 0) { | |
272 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && !ERRNO_IS_PRIVILEGE(errno)) | |
273 | return -errno; | |
274 | ||
275 | have_close_range = false; | |
276 | break; | |
277 | } | |
278 | ||
279 | c += end - start - 1; | |
280 | } | |
281 | ||
282 | if (have_close_range) { | |
283 | /* The loop succeeded. Let's now close everything beyond the end */ | |
284 | ||
285 | if (sorted[n_sorted-1] >= INT_MAX) /* Dont let the addition below overflow */ | |
286 | return c; | |
287 | ||
288 | if (close_range(sorted[n_sorted-1] + 1, -1, 0) >= 0) | |
289 | return c + 1; | |
290 | ||
291 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && !ERRNO_IS_PRIVILEGE(errno)) | |
292 | return -errno; | |
293 | ||
294 | have_close_range = false; | |
295 | } | |
296 | } | |
297 | } | |
298 | ||
299 | /* Fallback on OOM or if close_range() is not supported */ | |
300 | } | |
301 | ||
3ffd4af2 LP |
302 | d = opendir("/proc/self/fd"); |
303 | if (!d) { | |
37bc14de | 304 | int fd, max_fd; |
3ffd4af2 | 305 | |
498e265d LP |
306 | /* When /proc isn't available (for example in chroots) the fallback is brute forcing through |
307 | * the fd table */ | |
37bc14de | 308 | |
498e265d LP |
309 | max_fd = get_max_fd(); |
310 | if (max_fd < 0) | |
311 | return max_fd; | |
37bc14de | 312 | |
6a461d1f ZJS |
313 | /* Refuse to do the loop over more too many elements. It's better to fail immediately than to |
314 | * spin the CPU for a long time. */ | |
315 | if (max_fd > MAX_FD_LOOP_LIMIT) | |
316 | return log_debug_errno(SYNTHETIC_ERRNO(EPERM), | |
317 | "/proc/self/fd is inaccessible. Refusing to loop over %d potential fds.", | |
318 | max_fd); | |
319 | ||
37bc14de | 320 | for (fd = 3; fd >= 0; fd = fd < max_fd ? fd + 1 : -1) { |
e43bc9f5 | 321 | int q; |
3ffd4af2 LP |
322 | |
323 | if (fd_in_set(fd, except, n_except)) | |
324 | continue; | |
325 | ||
e43bc9f5 LP |
326 | q = close_nointr(fd); |
327 | if (q < 0 && q != -EBADF && r >= 0) | |
328 | r = q; | |
3ffd4af2 LP |
329 | } |
330 | ||
331 | return r; | |
332 | } | |
333 | ||
8fb3f009 | 334 | FOREACH_DIRENT(de, d, return -errno) { |
e43bc9f5 | 335 | int fd = -1, q; |
3ffd4af2 | 336 | |
3ffd4af2 LP |
337 | if (safe_atoi(de->d_name, &fd) < 0) |
338 | /* Let's better ignore this, just in case */ | |
339 | continue; | |
340 | ||
341 | if (fd < 3) | |
342 | continue; | |
343 | ||
344 | if (fd == dirfd(d)) | |
345 | continue; | |
346 | ||
347 | if (fd_in_set(fd, except, n_except)) | |
348 | continue; | |
349 | ||
e43bc9f5 LP |
350 | q = close_nointr(fd); |
351 | if (q < 0 && q != -EBADF && r >= 0) /* Valgrind has its own FD and doesn't want to have it closed */ | |
352 | r = q; | |
3ffd4af2 LP |
353 | } |
354 | ||
355 | return r; | |
356 | } | |
357 | ||
358 | int same_fd(int a, int b) { | |
359 | struct stat sta, stb; | |
360 | pid_t pid; | |
361 | int r, fa, fb; | |
362 | ||
363 | assert(a >= 0); | |
364 | assert(b >= 0); | |
365 | ||
366 | /* Compares two file descriptors. Note that semantics are | |
367 | * quite different depending on whether we have kcmp() or we | |
368 | * don't. If we have kcmp() this will only return true for | |
369 | * dup()ed file descriptors, but not otherwise. If we don't | |
370 | * have kcmp() this will also return true for two fds of the same | |
371 | * file, created by separate open() calls. Since we use this | |
372 | * call mostly for filtering out duplicates in the fd store | |
373 | * this difference hopefully doesn't matter too much. */ | |
374 | ||
375 | if (a == b) | |
376 | return true; | |
377 | ||
378 | /* Try to use kcmp() if we have it. */ | |
df0ff127 | 379 | pid = getpid_cached(); |
3ffd4af2 LP |
380 | r = kcmp(pid, pid, KCMP_FILE, a, b); |
381 | if (r == 0) | |
382 | return true; | |
383 | if (r > 0) | |
384 | return false; | |
9e2acd1d | 385 | if (!IN_SET(errno, ENOSYS, EACCES, EPERM)) |
3ffd4af2 LP |
386 | return -errno; |
387 | ||
388 | /* We don't have kcmp(), use fstat() instead. */ | |
389 | if (fstat(a, &sta) < 0) | |
390 | return -errno; | |
391 | ||
392 | if (fstat(b, &stb) < 0) | |
393 | return -errno; | |
394 | ||
395 | if ((sta.st_mode & S_IFMT) != (stb.st_mode & S_IFMT)) | |
396 | return false; | |
397 | ||
398 | /* We consider all device fds different, since two device fds | |
399 | * might refer to quite different device contexts even though | |
400 | * they share the same inode and backing dev_t. */ | |
401 | ||
402 | if (S_ISCHR(sta.st_mode) || S_ISBLK(sta.st_mode)) | |
403 | return false; | |
404 | ||
405 | if (sta.st_dev != stb.st_dev || sta.st_ino != stb.st_ino) | |
406 | return false; | |
407 | ||
408 | /* The fds refer to the same inode on disk, let's also check | |
409 | * if they have the same fd flags. This is useful to | |
410 | * distinguish the read and write side of a pipe created with | |
411 | * pipe(). */ | |
412 | fa = fcntl(a, F_GETFL); | |
413 | if (fa < 0) | |
414 | return -errno; | |
415 | ||
416 | fb = fcntl(b, F_GETFL); | |
417 | if (fb < 0) | |
418 | return -errno; | |
419 | ||
420 | return fa == fb; | |
421 | } | |
422 | ||
423 | void cmsg_close_all(struct msghdr *mh) { | |
424 | struct cmsghdr *cmsg; | |
425 | ||
426 | assert(mh); | |
427 | ||
428 | CMSG_FOREACH(cmsg, mh) | |
429 | if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS) | |
430 | close_many((int*) CMSG_DATA(cmsg), (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int)); | |
431 | } | |
4fee3975 LP |
432 | |
433 | bool fdname_is_valid(const char *s) { | |
434 | const char *p; | |
435 | ||
436 | /* Validates a name for $LISTEN_FDNAMES. We basically allow | |
437 | * everything ASCII that's not a control character. Also, as | |
438 | * special exception the ":" character is not allowed, as we | |
439 | * use that as field separator in $LISTEN_FDNAMES. | |
440 | * | |
441 | * Note that the empty string is explicitly allowed | |
442 | * here. However, we limit the length of the names to 255 | |
443 | * characters. */ | |
444 | ||
445 | if (!s) | |
446 | return false; | |
447 | ||
448 | for (p = s; *p; p++) { | |
449 | if (*p < ' ') | |
450 | return false; | |
451 | if (*p >= 127) | |
452 | return false; | |
453 | if (*p == ':') | |
454 | return false; | |
455 | } | |
456 | ||
457 | return p - s < 256; | |
458 | } | |
4aeb20f5 LP |
459 | |
460 | int fd_get_path(int fd, char **ret) { | |
f267719c | 461 | char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int)]; |
a0fe2a2d | 462 | int r; |
4aeb20f5 | 463 | |
f267719c LP |
464 | xsprintf(procfs_path, "/proc/self/fd/%i", fd); |
465 | r = readlink_malloc(procfs_path, ret); | |
466 | if (r == -ENOENT) { | |
467 | /* ENOENT can mean two things: that the fd does not exist or that /proc is not mounted. Let's make | |
5238e957 | 468 | * things debuggable and distinguish the two. */ |
4aeb20f5 | 469 | |
8fe8f3aa LP |
470 | if (proc_mounted() == 0) |
471 | return -ENOSYS; /* /proc is not available or not set up properly, we're most likely in some chroot | |
472 | * environment. */ | |
f267719c LP |
473 | return -EBADF; /* The directory exists, hence it's the fd that doesn't. */ |
474 | } | |
a0fe2a2d LP |
475 | |
476 | return r; | |
4aeb20f5 | 477 | } |
046a82c1 LP |
478 | |
479 | int move_fd(int from, int to, int cloexec) { | |
480 | int r; | |
481 | ||
482 | /* Move fd 'from' to 'to', make sure FD_CLOEXEC remains equal if requested, and release the old fd. If | |
483 | * 'cloexec' is passed as -1, the original FD_CLOEXEC is inherited for the new fd. If it is 0, it is turned | |
484 | * off, if it is > 0 it is turned on. */ | |
485 | ||
486 | if (from < 0) | |
487 | return -EBADF; | |
488 | if (to < 0) | |
489 | return -EBADF; | |
490 | ||
491 | if (from == to) { | |
492 | ||
493 | if (cloexec >= 0) { | |
494 | r = fd_cloexec(to, cloexec); | |
495 | if (r < 0) | |
496 | return r; | |
497 | } | |
498 | ||
499 | return to; | |
500 | } | |
501 | ||
502 | if (cloexec < 0) { | |
503 | int fl; | |
504 | ||
505 | fl = fcntl(from, F_GETFD, 0); | |
506 | if (fl < 0) | |
507 | return -errno; | |
508 | ||
509 | cloexec = !!(fl & FD_CLOEXEC); | |
510 | } | |
511 | ||
512 | r = dup3(from, to, cloexec ? O_CLOEXEC : 0); | |
513 | if (r < 0) | |
514 | return -errno; | |
515 | ||
516 | assert(r == to); | |
517 | ||
518 | safe_close(from); | |
519 | ||
520 | return to; | |
521 | } | |
a548e14d LP |
522 | |
523 | int acquire_data_fd(const void *data, size_t size, unsigned flags) { | |
524 | ||
a548e14d LP |
525 | _cleanup_close_pair_ int pipefds[2] = { -1, -1 }; |
526 | char pattern[] = "/dev/shm/data-fd-XXXXXX"; | |
527 | _cleanup_close_ int fd = -1; | |
528 | int isz = 0, r; | |
529 | ssize_t n; | |
530 | off_t f; | |
531 | ||
532 | assert(data || size == 0); | |
533 | ||
534 | /* Acquire a read-only file descriptor that when read from returns the specified data. This is much more | |
535 | * complex than I wish it was. But here's why: | |
536 | * | |
537 | * a) First we try to use memfds. They are the best option, as we can seal them nicely to make them | |
538 | * read-only. Unfortunately they require kernel 3.17, and – at the time of writing – we still support 3.14. | |
539 | * | |
540 | * b) Then, we try classic pipes. They are the second best options, as we can close the writing side, retaining | |
541 | * a nicely read-only fd in the reading side. However, they are by default quite small, and unprivileged | |
542 | * clients can only bump their size to a system-wide limit, which might be quite low. | |
543 | * | |
544 | * c) Then, we try an O_TMPFILE file in /dev/shm (that dir is the only suitable one known to exist from | |
545 | * earliest boot on). To make it read-only we open the fd a second time with O_RDONLY via | |
546 | * /proc/self/<fd>. Unfortunately O_TMPFILE is not available on older kernels on tmpfs. | |
547 | * | |
548 | * d) Finally, we try creating a regular file in /dev/shm, which we then delete. | |
549 | * | |
550 | * It sucks a bit that depending on the situation we return very different objects here, but that's Linux I | |
551 | * figure. */ | |
552 | ||
553 | if (size == 0 && ((flags & ACQUIRE_NO_DEV_NULL) == 0)) { | |
554 | /* As a special case, return /dev/null if we have been called for an empty data block */ | |
555 | r = open("/dev/null", O_RDONLY|O_CLOEXEC|O_NOCTTY); | |
556 | if (r < 0) | |
557 | return -errno; | |
558 | ||
559 | return r; | |
560 | } | |
561 | ||
562 | if ((flags & ACQUIRE_NO_MEMFD) == 0) { | |
563 | fd = memfd_new("data-fd"); | |
564 | if (fd < 0) | |
565 | goto try_pipe; | |
566 | ||
567 | n = write(fd, data, size); | |
568 | if (n < 0) | |
569 | return -errno; | |
570 | if ((size_t) n != size) | |
571 | return -EIO; | |
572 | ||
573 | f = lseek(fd, 0, SEEK_SET); | |
574 | if (f != 0) | |
575 | return -errno; | |
576 | ||
577 | r = memfd_set_sealed(fd); | |
578 | if (r < 0) | |
579 | return r; | |
580 | ||
c10d6bdb | 581 | return TAKE_FD(fd); |
a548e14d LP |
582 | } |
583 | ||
584 | try_pipe: | |
585 | if ((flags & ACQUIRE_NO_PIPE) == 0) { | |
586 | if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0) | |
587 | return -errno; | |
588 | ||
589 | isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0); | |
590 | if (isz < 0) | |
591 | return -errno; | |
592 | ||
593 | if ((size_t) isz < size) { | |
594 | isz = (int) size; | |
595 | if (isz < 0 || (size_t) isz != size) | |
596 | return -E2BIG; | |
597 | ||
598 | /* Try to bump the pipe size */ | |
599 | (void) fcntl(pipefds[1], F_SETPIPE_SZ, isz); | |
600 | ||
601 | /* See if that worked */ | |
602 | isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0); | |
603 | if (isz < 0) | |
604 | return -errno; | |
605 | ||
606 | if ((size_t) isz < size) | |
607 | goto try_dev_shm; | |
608 | } | |
609 | ||
610 | n = write(pipefds[1], data, size); | |
611 | if (n < 0) | |
612 | return -errno; | |
613 | if ((size_t) n != size) | |
614 | return -EIO; | |
615 | ||
616 | (void) fd_nonblock(pipefds[0], false); | |
617 | ||
c10d6bdb | 618 | return TAKE_FD(pipefds[0]); |
a548e14d LP |
619 | } |
620 | ||
621 | try_dev_shm: | |
622 | if ((flags & ACQUIRE_NO_TMPFILE) == 0) { | |
623 | fd = open("/dev/shm", O_RDWR|O_TMPFILE|O_CLOEXEC, 0500); | |
624 | if (fd < 0) | |
625 | goto try_dev_shm_without_o_tmpfile; | |
626 | ||
627 | n = write(fd, data, size); | |
628 | if (n < 0) | |
629 | return -errno; | |
630 | if ((size_t) n != size) | |
631 | return -EIO; | |
632 | ||
633 | /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */ | |
f2324783 | 634 | return fd_reopen(fd, O_RDONLY|O_CLOEXEC); |
a548e14d LP |
635 | } |
636 | ||
637 | try_dev_shm_without_o_tmpfile: | |
638 | if ((flags & ACQUIRE_NO_REGULAR) == 0) { | |
639 | fd = mkostemp_safe(pattern); | |
640 | if (fd < 0) | |
641 | return fd; | |
642 | ||
643 | n = write(fd, data, size); | |
644 | if (n < 0) { | |
645 | r = -errno; | |
646 | goto unlink_and_return; | |
647 | } | |
648 | if ((size_t) n != size) { | |
649 | r = -EIO; | |
650 | goto unlink_and_return; | |
651 | } | |
652 | ||
653 | /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */ | |
654 | r = open(pattern, O_RDONLY|O_CLOEXEC); | |
655 | if (r < 0) | |
656 | r = -errno; | |
657 | ||
658 | unlink_and_return: | |
659 | (void) unlink(pattern); | |
660 | return r; | |
661 | } | |
662 | ||
663 | return -EOPNOTSUPP; | |
664 | } | |
7fe2903c | 665 | |
4960ce43 LP |
666 | /* When the data is smaller or equal to 64K, try to place the copy in a memfd/pipe */ |
667 | #define DATA_FD_MEMORY_LIMIT (64U*1024U) | |
668 | ||
669 | /* If memfd/pipe didn't work out, then let's use a file in /tmp up to a size of 1M. If it's large than that use /var/tmp instead. */ | |
670 | #define DATA_FD_TMP_LIMIT (1024U*1024U) | |
671 | ||
672 | int fd_duplicate_data_fd(int fd) { | |
673 | ||
674 | _cleanup_close_ int copy_fd = -1, tmp_fd = -1; | |
675 | _cleanup_free_ void *remains = NULL; | |
4960ce43 LP |
676 | size_t remains_size = 0; |
677 | const char *td; | |
678 | struct stat st; | |
679 | int r; | |
680 | ||
681 | /* Creates a 'data' fd from the specified source fd, containing all the same data in a read-only fashion, but | |
682 | * independent of it (i.e. the source fd can be closed and unmounted after this call succeeded). Tries to be | |
683 | * somewhat smart about where to place the data. In the best case uses a memfd(). If memfd() are not supported | |
684 | * uses a pipe instead. For larger data will use an unlinked file in /tmp, and for even larger data one in | |
685 | * /var/tmp. */ | |
686 | ||
687 | if (fstat(fd, &st) < 0) | |
688 | return -errno; | |
689 | ||
690 | /* For now, let's only accept regular files, sockets, pipes and char devices */ | |
691 | if (S_ISDIR(st.st_mode)) | |
692 | return -EISDIR; | |
693 | if (S_ISLNK(st.st_mode)) | |
694 | return -ELOOP; | |
695 | if (!S_ISREG(st.st_mode) && !S_ISSOCK(st.st_mode) && !S_ISFIFO(st.st_mode) && !S_ISCHR(st.st_mode)) | |
696 | return -EBADFD; | |
697 | ||
698 | /* If we have reason to believe the data is bounded in size, then let's use memfds or pipes as backing fd. Note | |
699 | * that we use the reported regular file size only as a hint, given that there are plenty special files in | |
700 | * /proc and /sys which report a zero file size but can be read from. */ | |
701 | ||
702 | if (!S_ISREG(st.st_mode) || st.st_size < DATA_FD_MEMORY_LIMIT) { | |
703 | ||
704 | /* Try a memfd first */ | |
705 | copy_fd = memfd_new("data-fd"); | |
706 | if (copy_fd >= 0) { | |
707 | off_t f; | |
708 | ||
709 | r = copy_bytes(fd, copy_fd, DATA_FD_MEMORY_LIMIT, 0); | |
710 | if (r < 0) | |
711 | return r; | |
712 | ||
713 | f = lseek(copy_fd, 0, SEEK_SET); | |
714 | if (f != 0) | |
715 | return -errno; | |
716 | ||
717 | if (r == 0) { | |
718 | /* Did it fit into the limit? If so, we are done. */ | |
719 | r = memfd_set_sealed(copy_fd); | |
720 | if (r < 0) | |
721 | return r; | |
722 | ||
723 | return TAKE_FD(copy_fd); | |
724 | } | |
725 | ||
726 | /* Hmm, pity, this didn't fit. Let's fall back to /tmp then, see below */ | |
727 | ||
728 | } else { | |
729 | _cleanup_(close_pairp) int pipefds[2] = { -1, -1 }; | |
730 | int isz; | |
731 | ||
732 | /* If memfds aren't available, use a pipe. Set O_NONBLOCK so that we will get EAGAIN rather | |
733 | * then block indefinitely when we hit the pipe size limit */ | |
734 | ||
735 | if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0) | |
736 | return -errno; | |
737 | ||
738 | isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0); | |
739 | if (isz < 0) | |
740 | return -errno; | |
741 | ||
742 | /* Try to enlarge the pipe size if necessary */ | |
743 | if ((size_t) isz < DATA_FD_MEMORY_LIMIT) { | |
744 | ||
745 | (void) fcntl(pipefds[1], F_SETPIPE_SZ, DATA_FD_MEMORY_LIMIT); | |
746 | ||
747 | isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0); | |
748 | if (isz < 0) | |
749 | return -errno; | |
750 | } | |
751 | ||
752 | if ((size_t) isz >= DATA_FD_MEMORY_LIMIT) { | |
753 | ||
b3cade0c | 754 | r = copy_bytes_full(fd, pipefds[1], DATA_FD_MEMORY_LIMIT, 0, &remains, &remains_size, NULL, NULL); |
4960ce43 LP |
755 | if (r < 0 && r != -EAGAIN) |
756 | return r; /* If we get EAGAIN it could be because of the source or because of | |
757 | * the destination fd, we can't know, as sendfile() and friends won't | |
758 | * tell us. Hence, treat this as reason to fall back, just to be | |
759 | * sure. */ | |
760 | if (r == 0) { | |
761 | /* Everything fit in, yay! */ | |
762 | (void) fd_nonblock(pipefds[0], false); | |
763 | ||
764 | return TAKE_FD(pipefds[0]); | |
765 | } | |
766 | ||
767 | /* Things didn't fit in. But we read data into the pipe, let's remember that, so that | |
768 | * when writing the new file we incorporate this first. */ | |
769 | copy_fd = TAKE_FD(pipefds[0]); | |
770 | } | |
771 | } | |
772 | } | |
773 | ||
774 | /* If we have reason to believe this will fit fine in /tmp, then use that as first fallback. */ | |
775 | if ((!S_ISREG(st.st_mode) || st.st_size < DATA_FD_TMP_LIMIT) && | |
776 | (DATA_FD_MEMORY_LIMIT + remains_size) < DATA_FD_TMP_LIMIT) { | |
777 | off_t f; | |
778 | ||
779 | tmp_fd = open_tmpfile_unlinkable(NULL /* NULL as directory means /tmp */, O_RDWR|O_CLOEXEC); | |
780 | if (tmp_fd < 0) | |
781 | return tmp_fd; | |
782 | ||
783 | if (copy_fd >= 0) { | |
784 | /* If we tried a memfd/pipe first and it ended up being too large, then copy this into the | |
785 | * temporary file first. */ | |
786 | ||
787 | r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, 0); | |
788 | if (r < 0) | |
789 | return r; | |
790 | ||
791 | assert(r == 0); | |
792 | } | |
793 | ||
794 | if (remains_size > 0) { | |
795 | /* If there were remaining bytes (i.e. read into memory, but not written out yet) from the | |
796 | * failed copy operation, let's flush them out next. */ | |
797 | ||
798 | r = loop_write(tmp_fd, remains, remains_size, false); | |
799 | if (r < 0) | |
800 | return r; | |
801 | } | |
802 | ||
803 | r = copy_bytes(fd, tmp_fd, DATA_FD_TMP_LIMIT - DATA_FD_MEMORY_LIMIT - remains_size, COPY_REFLINK); | |
804 | if (r < 0) | |
805 | return r; | |
806 | if (r == 0) | |
807 | goto finish; /* Yay, it fit in */ | |
808 | ||
809 | /* It didn't fit in. Let's not forget to use what we already used */ | |
810 | f = lseek(tmp_fd, 0, SEEK_SET); | |
811 | if (f != 0) | |
812 | return -errno; | |
813 | ||
0706c012 | 814 | CLOSE_AND_REPLACE(copy_fd, tmp_fd); |
4960ce43 LP |
815 | |
816 | remains = mfree(remains); | |
817 | remains_size = 0; | |
818 | } | |
819 | ||
820 | /* As last fallback use /var/tmp */ | |
821 | r = var_tmp_dir(&td); | |
822 | if (r < 0) | |
823 | return r; | |
824 | ||
825 | tmp_fd = open_tmpfile_unlinkable(td, O_RDWR|O_CLOEXEC); | |
826 | if (tmp_fd < 0) | |
827 | return tmp_fd; | |
828 | ||
829 | if (copy_fd >= 0) { | |
830 | /* If we tried a memfd/pipe first, or a file in /tmp, and it ended up being too large, than copy this | |
831 | * into the temporary file first. */ | |
832 | r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, COPY_REFLINK); | |
833 | if (r < 0) | |
834 | return r; | |
835 | ||
836 | assert(r == 0); | |
837 | } | |
838 | ||
839 | if (remains_size > 0) { | |
840 | /* Then, copy in any read but not yet written bytes. */ | |
841 | r = loop_write(tmp_fd, remains, remains_size, false); | |
842 | if (r < 0) | |
843 | return r; | |
844 | } | |
845 | ||
846 | /* Copy in the rest */ | |
847 | r = copy_bytes(fd, tmp_fd, UINT64_MAX, COPY_REFLINK); | |
848 | if (r < 0) | |
849 | return r; | |
850 | ||
851 | assert(r == 0); | |
852 | ||
853 | finish: | |
854 | /* Now convert the O_RDWR file descriptor into an O_RDONLY one (and as side effect seek to the beginning of the | |
855 | * file again */ | |
856 | ||
857 | return fd_reopen(tmp_fd, O_RDONLY|O_CLOEXEC); | |
858 | } | |
859 | ||
7fe2903c LP |
860 | int fd_move_above_stdio(int fd) { |
861 | int flags, copy; | |
862 | PROTECT_ERRNO; | |
863 | ||
864 | /* Moves the specified file descriptor if possible out of the range [0…2], i.e. the range of | |
865 | * stdin/stdout/stderr. If it can't be moved outside of this range the original file descriptor is | |
866 | * returned. This call is supposed to be used for long-lasting file descriptors we allocate in our code that | |
867 | * might get loaded into foreign code, and where we want ensure our fds are unlikely used accidentally as | |
868 | * stdin/stdout/stderr of unrelated code. | |
869 | * | |
870 | * Note that this doesn't fix any real bugs, it just makes it less likely that our code will be affected by | |
871 | * buggy code from others that mindlessly invokes 'fprintf(stderr, …' or similar in places where stderr has | |
872 | * been closed before. | |
873 | * | |
874 | * This function is written in a "best-effort" and "least-impact" style. This means whenever we encounter an | |
875 | * error we simply return the original file descriptor, and we do not touch errno. */ | |
876 | ||
877 | if (fd < 0 || fd > 2) | |
878 | return fd; | |
879 | ||
880 | flags = fcntl(fd, F_GETFD, 0); | |
881 | if (flags < 0) | |
882 | return fd; | |
883 | ||
884 | if (flags & FD_CLOEXEC) | |
885 | copy = fcntl(fd, F_DUPFD_CLOEXEC, 3); | |
886 | else | |
887 | copy = fcntl(fd, F_DUPFD, 3); | |
888 | if (copy < 0) | |
889 | return fd; | |
890 | ||
891 | assert(copy > 2); | |
892 | ||
893 | (void) close(fd); | |
894 | return copy; | |
895 | } | |
aa11e28b LP |
896 | |
897 | int rearrange_stdio(int original_input_fd, int original_output_fd, int original_error_fd) { | |
898 | ||
899 | int fd[3] = { /* Put together an array of fds we work on */ | |
900 | original_input_fd, | |
901 | original_output_fd, | |
902 | original_error_fd | |
903 | }; | |
904 | ||
905 | int r, i, | |
906 | null_fd = -1, /* if we open /dev/null, we store the fd to it here */ | |
907 | copy_fd[3] = { -1, -1, -1 }; /* This contains all fds we duplicate here temporarily, and hence need to close at the end */ | |
908 | bool null_readable, null_writable; | |
909 | ||
910 | /* Sets up stdin, stdout, stderr with the three file descriptors passed in. If any of the descriptors is | |
911 | * specified as -1 it will be connected with /dev/null instead. If any of the file descriptors is passed as | |
912 | * itself (e.g. stdin as STDIN_FILENO) it is left unmodified, but the O_CLOEXEC bit is turned off should it be | |
913 | * on. | |
914 | * | |
915 | * Note that if any of the passed file descriptors are > 2 they will be closed — both on success and on | |
916 | * failure! Thus, callers should assume that when this function returns the input fds are invalidated. | |
917 | * | |
918 | * Note that when this function fails stdin/stdout/stderr might remain half set up! | |
919 | * | |
920 | * O_CLOEXEC is turned off for all three file descriptors (which is how it should be for | |
921 | * stdin/stdout/stderr). */ | |
922 | ||
923 | null_readable = original_input_fd < 0; | |
924 | null_writable = original_output_fd < 0 || original_error_fd < 0; | |
925 | ||
926 | /* First step, open /dev/null once, if we need it */ | |
927 | if (null_readable || null_writable) { | |
928 | ||
929 | /* Let's open this with O_CLOEXEC first, and convert it to non-O_CLOEXEC when we move the fd to the final position. */ | |
930 | null_fd = open("/dev/null", (null_readable && null_writable ? O_RDWR : | |
931 | null_readable ? O_RDONLY : O_WRONLY) | O_CLOEXEC); | |
932 | if (null_fd < 0) { | |
933 | r = -errno; | |
934 | goto finish; | |
935 | } | |
936 | ||
937 | /* If this fd is in the 0…2 range, let's move it out of it */ | |
938 | if (null_fd < 3) { | |
939 | int copy; | |
940 | ||
941 | copy = fcntl(null_fd, F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */ | |
942 | if (copy < 0) { | |
943 | r = -errno; | |
944 | goto finish; | |
945 | } | |
946 | ||
0706c012 | 947 | CLOSE_AND_REPLACE(null_fd, copy); |
aa11e28b LP |
948 | } |
949 | } | |
950 | ||
951 | /* Let's assemble fd[] with the fds to install in place of stdin/stdout/stderr */ | |
952 | for (i = 0; i < 3; i++) { | |
953 | ||
954 | if (fd[i] < 0) | |
955 | fd[i] = null_fd; /* A negative parameter means: connect this one to /dev/null */ | |
956 | else if (fd[i] != i && fd[i] < 3) { | |
957 | /* This fd is in the 0…2 territory, but not at its intended place, move it out of there, so that we can work there. */ | |
958 | copy_fd[i] = fcntl(fd[i], F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */ | |
959 | if (copy_fd[i] < 0) { | |
960 | r = -errno; | |
961 | goto finish; | |
962 | } | |
963 | ||
964 | fd[i] = copy_fd[i]; | |
965 | } | |
966 | } | |
967 | ||
968 | /* At this point we now have the fds to use in fd[], and they are all above the stdio range, so that we | |
969 | * have freedom to move them around. If the fds already were at the right places then the specific fds are | |
970 | * -1. Let's now move them to the right places. This is the point of no return. */ | |
971 | for (i = 0; i < 3; i++) { | |
972 | ||
973 | if (fd[i] == i) { | |
974 | ||
975 | /* fd is already in place, but let's make sure O_CLOEXEC is off */ | |
976 | r = fd_cloexec(i, false); | |
977 | if (r < 0) | |
978 | goto finish; | |
979 | ||
980 | } else { | |
981 | assert(fd[i] > 2); | |
982 | ||
983 | if (dup2(fd[i], i) < 0) { /* Turns off O_CLOEXEC on the new fd. */ | |
984 | r = -errno; | |
985 | goto finish; | |
986 | } | |
987 | } | |
988 | } | |
989 | ||
990 | r = 0; | |
991 | ||
992 | finish: | |
993 | /* Close the original fds, but only if they were outside of the stdio range. Also, properly check for the same | |
994 | * fd passed in multiple times. */ | |
995 | safe_close_above_stdio(original_input_fd); | |
996 | if (original_output_fd != original_input_fd) | |
997 | safe_close_above_stdio(original_output_fd); | |
998 | if (original_error_fd != original_input_fd && original_error_fd != original_output_fd) | |
999 | safe_close_above_stdio(original_error_fd); | |
1000 | ||
1001 | /* Close the copies we moved > 2 */ | |
1002 | for (i = 0; i < 3; i++) | |
1003 | safe_close(copy_fd[i]); | |
1004 | ||
1005 | /* Close our null fd, if it's > 2 */ | |
1006 | safe_close_above_stdio(null_fd); | |
1007 | ||
1008 | return r; | |
1009 | } | |
f2324783 LP |
1010 | |
1011 | int fd_reopen(int fd, int flags) { | |
1012 | char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int)]; | |
1013 | int new_fd; | |
1014 | ||
1015 | /* Reopens the specified fd with new flags. This is useful for convert an O_PATH fd into a regular one, or to | |
1016 | * turn O_RDWR fds into O_RDONLY fds. | |
1017 | * | |
1018 | * This doesn't work on sockets (since they cannot be open()ed, ever). | |
1019 | * | |
1020 | * This implicitly resets the file read index to 0. */ | |
1021 | ||
1022 | xsprintf(procfs_path, "/proc/self/fd/%i", fd); | |
1023 | new_fd = open(procfs_path, flags); | |
f8606626 LP |
1024 | if (new_fd < 0) { |
1025 | if (errno != ENOENT) | |
1026 | return -errno; | |
1027 | ||
1028 | if (proc_mounted() == 0) | |
1029 | return -ENOSYS; /* if we have no /proc/, the concept is not implementable */ | |
1030 | ||
1031 | return -ENOENT; | |
1032 | } | |
f2324783 LP |
1033 | |
1034 | return new_fd; | |
1035 | } | |
9264cc39 LP |
1036 | |
1037 | int read_nr_open(void) { | |
1038 | _cleanup_free_ char *nr_open = NULL; | |
1039 | int r; | |
1040 | ||
1041 | /* Returns the kernel's current fd limit, either by reading it of /proc/sys if that works, or using the | |
1042 | * hard-coded default compiled-in value of current kernels (1M) if not. This call will never fail. */ | |
1043 | ||
1044 | r = read_one_line_file("/proc/sys/fs/nr_open", &nr_open); | |
1045 | if (r < 0) | |
1046 | log_debug_errno(r, "Failed to read /proc/sys/fs/nr_open, ignoring: %m"); | |
1047 | else { | |
1048 | int v; | |
1049 | ||
1050 | r = safe_atoi(nr_open, &v); | |
1051 | if (r < 0) | |
1052 | log_debug_errno(r, "Failed to parse /proc/sys/fs/nr_open value '%s', ignoring: %m", nr_open); | |
1053 | else | |
1054 | return v; | |
1055 | } | |
1056 | ||
2aed63f4 | 1057 | /* If we fail, fall back to the hard-coded kernel limit of 1024 * 1024. */ |
9264cc39 LP |
1058 | return 1024 * 1024; |
1059 | } |