]>
Commit | Line | Data |
---|---|---|
53e1b683 | 1 | /* SPDX-License-Identifier: LGPL-2.1+ */ |
3ffd4af2 | 2 | |
11c3a366 TA |
3 | #include <errno.h> |
4 | #include <fcntl.h> | |
5 | #include <sys/resource.h> | |
6 | #include <sys/socket.h> | |
7 | #include <sys/stat.h> | |
8 | #include <unistd.h> | |
9 | ||
4960ce43 LP |
10 | #include "alloc-util.h" |
11 | #include "copy.h" | |
8fb3f009 | 12 | #include "dirent-util.h" |
3ffd4af2 | 13 | #include "fd-util.h" |
a548e14d | 14 | #include "fileio.h" |
4aeb20f5 | 15 | #include "fs-util.h" |
4960ce43 | 16 | #include "io-util.h" |
11c3a366 | 17 | #include "macro.h" |
a548e14d | 18 | #include "memfd-util.h" |
11c3a366 | 19 | #include "missing.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" |
4aeb20f5 | 24 | #include "stdio-util.h" |
3ffd4af2 LP |
25 | #include "util.h" |
26 | ||
27 | int close_nointr(int fd) { | |
28 | assert(fd >= 0); | |
29 | ||
30 | if (close(fd) >= 0) | |
31 | return 0; | |
32 | ||
33 | /* | |
34 | * Just ignore EINTR; a retry loop is the wrong thing to do on | |
35 | * Linux. | |
36 | * | |
37 | * http://lkml.indiana.edu/hypermail/linux/kernel/0509.1/0877.html | |
38 | * https://bugzilla.gnome.org/show_bug.cgi?id=682819 | |
39 | * http://utcc.utoronto.ca/~cks/space/blog/unix/CloseEINTR | |
40 | * https://sites.google.com/site/michaelsafyan/software-engineering/checkforeintrwheninvokingclosethinkagain | |
41 | */ | |
42 | if (errno == EINTR) | |
43 | return 0; | |
44 | ||
45 | return -errno; | |
46 | } | |
47 | ||
48 | int safe_close(int fd) { | |
49 | ||
50 | /* | |
51 | * Like close_nointr() but cannot fail. Guarantees errno is | |
52 | * unchanged. Is a NOP with negative fds passed, and returns | |
53 | * -1, so that it can be used in this syntax: | |
54 | * | |
55 | * fd = safe_close(fd); | |
56 | */ | |
57 | ||
58 | if (fd >= 0) { | |
59 | PROTECT_ERRNO; | |
60 | ||
61 | /* The kernel might return pretty much any error code | |
62 | * via close(), but the fd will be closed anyway. The | |
63 | * only condition we want to check for here is whether | |
64 | * the fd was invalid at all... */ | |
65 | ||
66 | assert_se(close_nointr(fd) != -EBADF); | |
67 | } | |
68 | ||
69 | return -1; | |
70 | } | |
71 | ||
72 | void safe_close_pair(int p[]) { | |
73 | assert(p); | |
74 | ||
75 | if (p[0] == p[1]) { | |
76 | /* Special case pairs which use the same fd in both | |
77 | * directions... */ | |
78 | p[0] = p[1] = safe_close(p[0]); | |
79 | return; | |
80 | } | |
81 | ||
82 | p[0] = safe_close(p[0]); | |
83 | p[1] = safe_close(p[1]); | |
84 | } | |
85 | ||
da6053d0 LP |
86 | void close_many(const int fds[], size_t n_fd) { |
87 | size_t i; | |
3ffd4af2 LP |
88 | |
89 | assert(fds || n_fd <= 0); | |
90 | ||
91 | for (i = 0; i < n_fd; i++) | |
92 | safe_close(fds[i]); | |
93 | } | |
94 | ||
95 | int fclose_nointr(FILE *f) { | |
96 | assert(f); | |
97 | ||
98 | /* Same as close_nointr(), but for fclose() */ | |
99 | ||
100 | if (fclose(f) == 0) | |
101 | return 0; | |
102 | ||
103 | if (errno == EINTR) | |
104 | return 0; | |
105 | ||
106 | return -errno; | |
107 | } | |
108 | ||
109 | FILE* safe_fclose(FILE *f) { | |
110 | ||
111 | /* Same as safe_close(), but for fclose() */ | |
112 | ||
113 | if (f) { | |
114 | PROTECT_ERRNO; | |
115 | ||
116 | assert_se(fclose_nointr(f) != EBADF); | |
117 | } | |
118 | ||
119 | return NULL; | |
120 | } | |
121 | ||
122 | DIR* safe_closedir(DIR *d) { | |
123 | ||
124 | if (d) { | |
125 | PROTECT_ERRNO; | |
126 | ||
127 | assert_se(closedir(d) >= 0 || errno != EBADF); | |
128 | } | |
129 | ||
130 | return NULL; | |
131 | } | |
132 | ||
133 | int fd_nonblock(int fd, bool nonblock) { | |
134 | int flags, nflags; | |
135 | ||
136 | assert(fd >= 0); | |
137 | ||
138 | flags = fcntl(fd, F_GETFL, 0); | |
139 | if (flags < 0) | |
140 | return -errno; | |
141 | ||
142 | if (nonblock) | |
143 | nflags = flags | O_NONBLOCK; | |
144 | else | |
145 | nflags = flags & ~O_NONBLOCK; | |
146 | ||
147 | if (nflags == flags) | |
148 | return 0; | |
149 | ||
150 | if (fcntl(fd, F_SETFL, nflags) < 0) | |
151 | return -errno; | |
152 | ||
153 | return 0; | |
154 | } | |
155 | ||
156 | int fd_cloexec(int fd, bool cloexec) { | |
157 | int flags, nflags; | |
158 | ||
159 | assert(fd >= 0); | |
160 | ||
161 | flags = fcntl(fd, F_GETFD, 0); | |
162 | if (flags < 0) | |
163 | return -errno; | |
164 | ||
165 | if (cloexec) | |
166 | nflags = flags | FD_CLOEXEC; | |
167 | else | |
168 | nflags = flags & ~FD_CLOEXEC; | |
169 | ||
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 LP |
179 | _pure_ static bool fd_in_set(int fd, const int fdset[], size_t n_fdset) { |
180 | size_t i; | |
3ffd4af2 LP |
181 | |
182 | assert(n_fdset == 0 || fdset); | |
183 | ||
184 | for (i = 0; i < n_fdset; i++) | |
185 | if (fdset[i] == fd) | |
186 | return true; | |
187 | ||
188 | return false; | |
189 | } | |
190 | ||
da6053d0 | 191 | int close_all_fds(const int except[], size_t n_except) { |
3ffd4af2 LP |
192 | _cleanup_closedir_ DIR *d = NULL; |
193 | struct dirent *de; | |
194 | int r = 0; | |
195 | ||
196 | assert(n_except == 0 || except); | |
197 | ||
198 | d = opendir("/proc/self/fd"); | |
199 | if (!d) { | |
3ffd4af2 | 200 | struct rlimit rl; |
37bc14de | 201 | int fd, max_fd; |
3ffd4af2 | 202 | |
37bc14de | 203 | /* When /proc isn't available (for example in chroots) the fallback is brute forcing through the fd |
3ffd4af2 LP |
204 | * table */ |
205 | ||
206 | assert_se(getrlimit(RLIMIT_NOFILE, &rl) >= 0); | |
37bc14de LP |
207 | |
208 | if (rl.rlim_max == 0) | |
209 | return -EINVAL; | |
210 | ||
211 | /* Let's take special care if the resource limit is set to unlimited, or actually larger than the range | |
212 | * of 'int'. Let's avoid implicit overflows. */ | |
213 | max_fd = (rl.rlim_max == RLIM_INFINITY || rl.rlim_max > INT_MAX) ? INT_MAX : (int) (rl.rlim_max - 1); | |
214 | ||
215 | for (fd = 3; fd >= 0; fd = fd < max_fd ? fd + 1 : -1) { | |
e43bc9f5 | 216 | int q; |
3ffd4af2 LP |
217 | |
218 | if (fd_in_set(fd, except, n_except)) | |
219 | continue; | |
220 | ||
e43bc9f5 LP |
221 | q = close_nointr(fd); |
222 | if (q < 0 && q != -EBADF && r >= 0) | |
223 | r = q; | |
3ffd4af2 LP |
224 | } |
225 | ||
226 | return r; | |
227 | } | |
228 | ||
8fb3f009 | 229 | FOREACH_DIRENT(de, d, return -errno) { |
e43bc9f5 | 230 | int fd = -1, q; |
3ffd4af2 | 231 | |
3ffd4af2 LP |
232 | if (safe_atoi(de->d_name, &fd) < 0) |
233 | /* Let's better ignore this, just in case */ | |
234 | continue; | |
235 | ||
236 | if (fd < 3) | |
237 | continue; | |
238 | ||
239 | if (fd == dirfd(d)) | |
240 | continue; | |
241 | ||
242 | if (fd_in_set(fd, except, n_except)) | |
243 | continue; | |
244 | ||
e43bc9f5 LP |
245 | q = close_nointr(fd); |
246 | if (q < 0 && q != -EBADF && r >= 0) /* Valgrind has its own FD and doesn't want to have it closed */ | |
247 | r = q; | |
3ffd4af2 LP |
248 | } |
249 | ||
250 | return r; | |
251 | } | |
252 | ||
253 | int same_fd(int a, int b) { | |
254 | struct stat sta, stb; | |
255 | pid_t pid; | |
256 | int r, fa, fb; | |
257 | ||
258 | assert(a >= 0); | |
259 | assert(b >= 0); | |
260 | ||
261 | /* Compares two file descriptors. Note that semantics are | |
262 | * quite different depending on whether we have kcmp() or we | |
263 | * don't. If we have kcmp() this will only return true for | |
264 | * dup()ed file descriptors, but not otherwise. If we don't | |
265 | * have kcmp() this will also return true for two fds of the same | |
266 | * file, created by separate open() calls. Since we use this | |
267 | * call mostly for filtering out duplicates in the fd store | |
268 | * this difference hopefully doesn't matter too much. */ | |
269 | ||
270 | if (a == b) | |
271 | return true; | |
272 | ||
273 | /* Try to use kcmp() if we have it. */ | |
df0ff127 | 274 | pid = getpid_cached(); |
3ffd4af2 LP |
275 | r = kcmp(pid, pid, KCMP_FILE, a, b); |
276 | if (r == 0) | |
277 | return true; | |
278 | if (r > 0) | |
279 | return false; | |
280 | if (errno != ENOSYS) | |
281 | return -errno; | |
282 | ||
283 | /* We don't have kcmp(), use fstat() instead. */ | |
284 | if (fstat(a, &sta) < 0) | |
285 | return -errno; | |
286 | ||
287 | if (fstat(b, &stb) < 0) | |
288 | return -errno; | |
289 | ||
290 | if ((sta.st_mode & S_IFMT) != (stb.st_mode & S_IFMT)) | |
291 | return false; | |
292 | ||
293 | /* We consider all device fds different, since two device fds | |
294 | * might refer to quite different device contexts even though | |
295 | * they share the same inode and backing dev_t. */ | |
296 | ||
297 | if (S_ISCHR(sta.st_mode) || S_ISBLK(sta.st_mode)) | |
298 | return false; | |
299 | ||
300 | if (sta.st_dev != stb.st_dev || sta.st_ino != stb.st_ino) | |
301 | return false; | |
302 | ||
303 | /* The fds refer to the same inode on disk, let's also check | |
304 | * if they have the same fd flags. This is useful to | |
305 | * distinguish the read and write side of a pipe created with | |
306 | * pipe(). */ | |
307 | fa = fcntl(a, F_GETFL); | |
308 | if (fa < 0) | |
309 | return -errno; | |
310 | ||
311 | fb = fcntl(b, F_GETFL); | |
312 | if (fb < 0) | |
313 | return -errno; | |
314 | ||
315 | return fa == fb; | |
316 | } | |
317 | ||
318 | void cmsg_close_all(struct msghdr *mh) { | |
319 | struct cmsghdr *cmsg; | |
320 | ||
321 | assert(mh); | |
322 | ||
323 | CMSG_FOREACH(cmsg, mh) | |
324 | if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS) | |
325 | close_many((int*) CMSG_DATA(cmsg), (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int)); | |
326 | } | |
4fee3975 LP |
327 | |
328 | bool fdname_is_valid(const char *s) { | |
329 | const char *p; | |
330 | ||
331 | /* Validates a name for $LISTEN_FDNAMES. We basically allow | |
332 | * everything ASCII that's not a control character. Also, as | |
333 | * special exception the ":" character is not allowed, as we | |
334 | * use that as field separator in $LISTEN_FDNAMES. | |
335 | * | |
336 | * Note that the empty string is explicitly allowed | |
337 | * here. However, we limit the length of the names to 255 | |
338 | * characters. */ | |
339 | ||
340 | if (!s) | |
341 | return false; | |
342 | ||
343 | for (p = s; *p; p++) { | |
344 | if (*p < ' ') | |
345 | return false; | |
346 | if (*p >= 127) | |
347 | return false; | |
348 | if (*p == ':') | |
349 | return false; | |
350 | } | |
351 | ||
352 | return p - s < 256; | |
353 | } | |
4aeb20f5 LP |
354 | |
355 | int fd_get_path(int fd, char **ret) { | |
3ceae1bc ZJS |
356 | _cleanup_close_ int dir = -1; |
357 | char fdname[DECIMAL_STR_MAX(int)]; | |
a0fe2a2d | 358 | int r; |
4aeb20f5 | 359 | |
3ceae1bc ZJS |
360 | dir = open("/proc/self/fd/", O_CLOEXEC | O_DIRECTORY | O_PATH); |
361 | if (dir < 0) | |
362 | /* /proc is not available or not set up properly, we're most likely | |
363 | * in some chroot environment. */ | |
364 | return errno == ENOENT ? -EOPNOTSUPP : -errno; | |
4aeb20f5 | 365 | |
3ceae1bc | 366 | xsprintf(fdname, "%i", fd); |
a0fe2a2d | 367 | |
3ceae1bc ZJS |
368 | r = readlinkat_malloc(dir, fdname, ret); |
369 | if (r == -ENOENT) | |
370 | /* If the file doesn't exist the fd is invalid */ | |
a0fe2a2d LP |
371 | return -EBADF; |
372 | ||
373 | return r; | |
4aeb20f5 | 374 | } |
046a82c1 LP |
375 | |
376 | int move_fd(int from, int to, int cloexec) { | |
377 | int r; | |
378 | ||
379 | /* Move fd 'from' to 'to', make sure FD_CLOEXEC remains equal if requested, and release the old fd. If | |
380 | * 'cloexec' is passed as -1, the original FD_CLOEXEC is inherited for the new fd. If it is 0, it is turned | |
381 | * off, if it is > 0 it is turned on. */ | |
382 | ||
383 | if (from < 0) | |
384 | return -EBADF; | |
385 | if (to < 0) | |
386 | return -EBADF; | |
387 | ||
388 | if (from == to) { | |
389 | ||
390 | if (cloexec >= 0) { | |
391 | r = fd_cloexec(to, cloexec); | |
392 | if (r < 0) | |
393 | return r; | |
394 | } | |
395 | ||
396 | return to; | |
397 | } | |
398 | ||
399 | if (cloexec < 0) { | |
400 | int fl; | |
401 | ||
402 | fl = fcntl(from, F_GETFD, 0); | |
403 | if (fl < 0) | |
404 | return -errno; | |
405 | ||
406 | cloexec = !!(fl & FD_CLOEXEC); | |
407 | } | |
408 | ||
409 | r = dup3(from, to, cloexec ? O_CLOEXEC : 0); | |
410 | if (r < 0) | |
411 | return -errno; | |
412 | ||
413 | assert(r == to); | |
414 | ||
415 | safe_close(from); | |
416 | ||
417 | return to; | |
418 | } | |
a548e14d LP |
419 | |
420 | int acquire_data_fd(const void *data, size_t size, unsigned flags) { | |
421 | ||
a548e14d LP |
422 | _cleanup_close_pair_ int pipefds[2] = { -1, -1 }; |
423 | char pattern[] = "/dev/shm/data-fd-XXXXXX"; | |
424 | _cleanup_close_ int fd = -1; | |
425 | int isz = 0, r; | |
426 | ssize_t n; | |
427 | off_t f; | |
428 | ||
429 | assert(data || size == 0); | |
430 | ||
431 | /* Acquire a read-only file descriptor that when read from returns the specified data. This is much more | |
432 | * complex than I wish it was. But here's why: | |
433 | * | |
434 | * a) First we try to use memfds. They are the best option, as we can seal them nicely to make them | |
435 | * read-only. Unfortunately they require kernel 3.17, and – at the time of writing – we still support 3.14. | |
436 | * | |
437 | * b) Then, we try classic pipes. They are the second best options, as we can close the writing side, retaining | |
438 | * a nicely read-only fd in the reading side. However, they are by default quite small, and unprivileged | |
439 | * clients can only bump their size to a system-wide limit, which might be quite low. | |
440 | * | |
441 | * c) Then, we try an O_TMPFILE file in /dev/shm (that dir is the only suitable one known to exist from | |
442 | * earliest boot on). To make it read-only we open the fd a second time with O_RDONLY via | |
443 | * /proc/self/<fd>. Unfortunately O_TMPFILE is not available on older kernels on tmpfs. | |
444 | * | |
445 | * d) Finally, we try creating a regular file in /dev/shm, which we then delete. | |
446 | * | |
447 | * It sucks a bit that depending on the situation we return very different objects here, but that's Linux I | |
448 | * figure. */ | |
449 | ||
450 | if (size == 0 && ((flags & ACQUIRE_NO_DEV_NULL) == 0)) { | |
451 | /* As a special case, return /dev/null if we have been called for an empty data block */ | |
452 | r = open("/dev/null", O_RDONLY|O_CLOEXEC|O_NOCTTY); | |
453 | if (r < 0) | |
454 | return -errno; | |
455 | ||
456 | return r; | |
457 | } | |
458 | ||
459 | if ((flags & ACQUIRE_NO_MEMFD) == 0) { | |
460 | fd = memfd_new("data-fd"); | |
461 | if (fd < 0) | |
462 | goto try_pipe; | |
463 | ||
464 | n = write(fd, data, size); | |
465 | if (n < 0) | |
466 | return -errno; | |
467 | if ((size_t) n != size) | |
468 | return -EIO; | |
469 | ||
470 | f = lseek(fd, 0, SEEK_SET); | |
471 | if (f != 0) | |
472 | return -errno; | |
473 | ||
474 | r = memfd_set_sealed(fd); | |
475 | if (r < 0) | |
476 | return r; | |
477 | ||
c10d6bdb | 478 | return TAKE_FD(fd); |
a548e14d LP |
479 | } |
480 | ||
481 | try_pipe: | |
482 | if ((flags & ACQUIRE_NO_PIPE) == 0) { | |
483 | if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0) | |
484 | return -errno; | |
485 | ||
486 | isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0); | |
487 | if (isz < 0) | |
488 | return -errno; | |
489 | ||
490 | if ((size_t) isz < size) { | |
491 | isz = (int) size; | |
492 | if (isz < 0 || (size_t) isz != size) | |
493 | return -E2BIG; | |
494 | ||
495 | /* Try to bump the pipe size */ | |
496 | (void) fcntl(pipefds[1], F_SETPIPE_SZ, isz); | |
497 | ||
498 | /* See if that worked */ | |
499 | isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0); | |
500 | if (isz < 0) | |
501 | return -errno; | |
502 | ||
503 | if ((size_t) isz < size) | |
504 | goto try_dev_shm; | |
505 | } | |
506 | ||
507 | n = write(pipefds[1], data, size); | |
508 | if (n < 0) | |
509 | return -errno; | |
510 | if ((size_t) n != size) | |
511 | return -EIO; | |
512 | ||
513 | (void) fd_nonblock(pipefds[0], false); | |
514 | ||
c10d6bdb | 515 | return TAKE_FD(pipefds[0]); |
a548e14d LP |
516 | } |
517 | ||
518 | try_dev_shm: | |
519 | if ((flags & ACQUIRE_NO_TMPFILE) == 0) { | |
520 | fd = open("/dev/shm", O_RDWR|O_TMPFILE|O_CLOEXEC, 0500); | |
521 | if (fd < 0) | |
522 | goto try_dev_shm_without_o_tmpfile; | |
523 | ||
524 | n = write(fd, data, size); | |
525 | if (n < 0) | |
526 | return -errno; | |
527 | if ((size_t) n != size) | |
528 | return -EIO; | |
529 | ||
530 | /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */ | |
f2324783 | 531 | return fd_reopen(fd, O_RDONLY|O_CLOEXEC); |
a548e14d LP |
532 | } |
533 | ||
534 | try_dev_shm_without_o_tmpfile: | |
535 | if ((flags & ACQUIRE_NO_REGULAR) == 0) { | |
536 | fd = mkostemp_safe(pattern); | |
537 | if (fd < 0) | |
538 | return fd; | |
539 | ||
540 | n = write(fd, data, size); | |
541 | if (n < 0) { | |
542 | r = -errno; | |
543 | goto unlink_and_return; | |
544 | } | |
545 | if ((size_t) n != size) { | |
546 | r = -EIO; | |
547 | goto unlink_and_return; | |
548 | } | |
549 | ||
550 | /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */ | |
551 | r = open(pattern, O_RDONLY|O_CLOEXEC); | |
552 | if (r < 0) | |
553 | r = -errno; | |
554 | ||
555 | unlink_and_return: | |
556 | (void) unlink(pattern); | |
557 | return r; | |
558 | } | |
559 | ||
560 | return -EOPNOTSUPP; | |
561 | } | |
7fe2903c | 562 | |
4960ce43 LP |
563 | /* When the data is smaller or equal to 64K, try to place the copy in a memfd/pipe */ |
564 | #define DATA_FD_MEMORY_LIMIT (64U*1024U) | |
565 | ||
566 | /* 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. */ | |
567 | #define DATA_FD_TMP_LIMIT (1024U*1024U) | |
568 | ||
569 | int fd_duplicate_data_fd(int fd) { | |
570 | ||
571 | _cleanup_close_ int copy_fd = -1, tmp_fd = -1; | |
572 | _cleanup_free_ void *remains = NULL; | |
4960ce43 LP |
573 | size_t remains_size = 0; |
574 | const char *td; | |
575 | struct stat st; | |
576 | int r; | |
577 | ||
578 | /* Creates a 'data' fd from the specified source fd, containing all the same data in a read-only fashion, but | |
579 | * independent of it (i.e. the source fd can be closed and unmounted after this call succeeded). Tries to be | |
580 | * somewhat smart about where to place the data. In the best case uses a memfd(). If memfd() are not supported | |
581 | * uses a pipe instead. For larger data will use an unlinked file in /tmp, and for even larger data one in | |
582 | * /var/tmp. */ | |
583 | ||
584 | if (fstat(fd, &st) < 0) | |
585 | return -errno; | |
586 | ||
587 | /* For now, let's only accept regular files, sockets, pipes and char devices */ | |
588 | if (S_ISDIR(st.st_mode)) | |
589 | return -EISDIR; | |
590 | if (S_ISLNK(st.st_mode)) | |
591 | return -ELOOP; | |
592 | if (!S_ISREG(st.st_mode) && !S_ISSOCK(st.st_mode) && !S_ISFIFO(st.st_mode) && !S_ISCHR(st.st_mode)) | |
593 | return -EBADFD; | |
594 | ||
595 | /* If we have reason to believe the data is bounded in size, then let's use memfds or pipes as backing fd. Note | |
596 | * that we use the reported regular file size only as a hint, given that there are plenty special files in | |
597 | * /proc and /sys which report a zero file size but can be read from. */ | |
598 | ||
599 | if (!S_ISREG(st.st_mode) || st.st_size < DATA_FD_MEMORY_LIMIT) { | |
600 | ||
601 | /* Try a memfd first */ | |
602 | copy_fd = memfd_new("data-fd"); | |
603 | if (copy_fd >= 0) { | |
604 | off_t f; | |
605 | ||
606 | r = copy_bytes(fd, copy_fd, DATA_FD_MEMORY_LIMIT, 0); | |
607 | if (r < 0) | |
608 | return r; | |
609 | ||
610 | f = lseek(copy_fd, 0, SEEK_SET); | |
611 | if (f != 0) | |
612 | return -errno; | |
613 | ||
614 | if (r == 0) { | |
615 | /* Did it fit into the limit? If so, we are done. */ | |
616 | r = memfd_set_sealed(copy_fd); | |
617 | if (r < 0) | |
618 | return r; | |
619 | ||
620 | return TAKE_FD(copy_fd); | |
621 | } | |
622 | ||
623 | /* Hmm, pity, this didn't fit. Let's fall back to /tmp then, see below */ | |
624 | ||
625 | } else { | |
626 | _cleanup_(close_pairp) int pipefds[2] = { -1, -1 }; | |
627 | int isz; | |
628 | ||
629 | /* If memfds aren't available, use a pipe. Set O_NONBLOCK so that we will get EAGAIN rather | |
630 | * then block indefinitely when we hit the pipe size limit */ | |
631 | ||
632 | if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0) | |
633 | return -errno; | |
634 | ||
635 | isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0); | |
636 | if (isz < 0) | |
637 | return -errno; | |
638 | ||
639 | /* Try to enlarge the pipe size if necessary */ | |
640 | if ((size_t) isz < DATA_FD_MEMORY_LIMIT) { | |
641 | ||
642 | (void) fcntl(pipefds[1], F_SETPIPE_SZ, DATA_FD_MEMORY_LIMIT); | |
643 | ||
644 | isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0); | |
645 | if (isz < 0) | |
646 | return -errno; | |
647 | } | |
648 | ||
649 | if ((size_t) isz >= DATA_FD_MEMORY_LIMIT) { | |
650 | ||
651 | r = copy_bytes_full(fd, pipefds[1], DATA_FD_MEMORY_LIMIT, 0, &remains, &remains_size); | |
652 | if (r < 0 && r != -EAGAIN) | |
653 | return r; /* If we get EAGAIN it could be because of the source or because of | |
654 | * the destination fd, we can't know, as sendfile() and friends won't | |
655 | * tell us. Hence, treat this as reason to fall back, just to be | |
656 | * sure. */ | |
657 | if (r == 0) { | |
658 | /* Everything fit in, yay! */ | |
659 | (void) fd_nonblock(pipefds[0], false); | |
660 | ||
661 | return TAKE_FD(pipefds[0]); | |
662 | } | |
663 | ||
664 | /* Things didn't fit in. But we read data into the pipe, let's remember that, so that | |
665 | * when writing the new file we incorporate this first. */ | |
666 | copy_fd = TAKE_FD(pipefds[0]); | |
667 | } | |
668 | } | |
669 | } | |
670 | ||
671 | /* If we have reason to believe this will fit fine in /tmp, then use that as first fallback. */ | |
672 | if ((!S_ISREG(st.st_mode) || st.st_size < DATA_FD_TMP_LIMIT) && | |
673 | (DATA_FD_MEMORY_LIMIT + remains_size) < DATA_FD_TMP_LIMIT) { | |
674 | off_t f; | |
675 | ||
676 | tmp_fd = open_tmpfile_unlinkable(NULL /* NULL as directory means /tmp */, O_RDWR|O_CLOEXEC); | |
677 | if (tmp_fd < 0) | |
678 | return tmp_fd; | |
679 | ||
680 | if (copy_fd >= 0) { | |
681 | /* If we tried a memfd/pipe first and it ended up being too large, then copy this into the | |
682 | * temporary file first. */ | |
683 | ||
684 | r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, 0); | |
685 | if (r < 0) | |
686 | return r; | |
687 | ||
688 | assert(r == 0); | |
689 | } | |
690 | ||
691 | if (remains_size > 0) { | |
692 | /* If there were remaining bytes (i.e. read into memory, but not written out yet) from the | |
693 | * failed copy operation, let's flush them out next. */ | |
694 | ||
695 | r = loop_write(tmp_fd, remains, remains_size, false); | |
696 | if (r < 0) | |
697 | return r; | |
698 | } | |
699 | ||
700 | r = copy_bytes(fd, tmp_fd, DATA_FD_TMP_LIMIT - DATA_FD_MEMORY_LIMIT - remains_size, COPY_REFLINK); | |
701 | if (r < 0) | |
702 | return r; | |
703 | if (r == 0) | |
704 | goto finish; /* Yay, it fit in */ | |
705 | ||
706 | /* It didn't fit in. Let's not forget to use what we already used */ | |
707 | f = lseek(tmp_fd, 0, SEEK_SET); | |
708 | if (f != 0) | |
709 | return -errno; | |
710 | ||
711 | safe_close(copy_fd); | |
712 | copy_fd = TAKE_FD(tmp_fd); | |
713 | ||
714 | remains = mfree(remains); | |
715 | remains_size = 0; | |
716 | } | |
717 | ||
718 | /* As last fallback use /var/tmp */ | |
719 | r = var_tmp_dir(&td); | |
720 | if (r < 0) | |
721 | return r; | |
722 | ||
723 | tmp_fd = open_tmpfile_unlinkable(td, O_RDWR|O_CLOEXEC); | |
724 | if (tmp_fd < 0) | |
725 | return tmp_fd; | |
726 | ||
727 | if (copy_fd >= 0) { | |
728 | /* If we tried a memfd/pipe first, or a file in /tmp, and it ended up being too large, than copy this | |
729 | * into the temporary file first. */ | |
730 | r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, COPY_REFLINK); | |
731 | if (r < 0) | |
732 | return r; | |
733 | ||
734 | assert(r == 0); | |
735 | } | |
736 | ||
737 | if (remains_size > 0) { | |
738 | /* Then, copy in any read but not yet written bytes. */ | |
739 | r = loop_write(tmp_fd, remains, remains_size, false); | |
740 | if (r < 0) | |
741 | return r; | |
742 | } | |
743 | ||
744 | /* Copy in the rest */ | |
745 | r = copy_bytes(fd, tmp_fd, UINT64_MAX, COPY_REFLINK); | |
746 | if (r < 0) | |
747 | return r; | |
748 | ||
749 | assert(r == 0); | |
750 | ||
751 | finish: | |
752 | /* Now convert the O_RDWR file descriptor into an O_RDONLY one (and as side effect seek to the beginning of the | |
753 | * file again */ | |
754 | ||
755 | return fd_reopen(tmp_fd, O_RDONLY|O_CLOEXEC); | |
756 | } | |
757 | ||
7fe2903c LP |
758 | int fd_move_above_stdio(int fd) { |
759 | int flags, copy; | |
760 | PROTECT_ERRNO; | |
761 | ||
762 | /* Moves the specified file descriptor if possible out of the range [0…2], i.e. the range of | |
763 | * stdin/stdout/stderr. If it can't be moved outside of this range the original file descriptor is | |
764 | * returned. This call is supposed to be used for long-lasting file descriptors we allocate in our code that | |
765 | * might get loaded into foreign code, and where we want ensure our fds are unlikely used accidentally as | |
766 | * stdin/stdout/stderr of unrelated code. | |
767 | * | |
768 | * Note that this doesn't fix any real bugs, it just makes it less likely that our code will be affected by | |
769 | * buggy code from others that mindlessly invokes 'fprintf(stderr, …' or similar in places where stderr has | |
770 | * been closed before. | |
771 | * | |
772 | * This function is written in a "best-effort" and "least-impact" style. This means whenever we encounter an | |
773 | * error we simply return the original file descriptor, and we do not touch errno. */ | |
774 | ||
775 | if (fd < 0 || fd > 2) | |
776 | return fd; | |
777 | ||
778 | flags = fcntl(fd, F_GETFD, 0); | |
779 | if (flags < 0) | |
780 | return fd; | |
781 | ||
782 | if (flags & FD_CLOEXEC) | |
783 | copy = fcntl(fd, F_DUPFD_CLOEXEC, 3); | |
784 | else | |
785 | copy = fcntl(fd, F_DUPFD, 3); | |
786 | if (copy < 0) | |
787 | return fd; | |
788 | ||
789 | assert(copy > 2); | |
790 | ||
791 | (void) close(fd); | |
792 | return copy; | |
793 | } | |
aa11e28b LP |
794 | |
795 | int rearrange_stdio(int original_input_fd, int original_output_fd, int original_error_fd) { | |
796 | ||
797 | int fd[3] = { /* Put together an array of fds we work on */ | |
798 | original_input_fd, | |
799 | original_output_fd, | |
800 | original_error_fd | |
801 | }; | |
802 | ||
803 | int r, i, | |
804 | null_fd = -1, /* if we open /dev/null, we store the fd to it here */ | |
805 | copy_fd[3] = { -1, -1, -1 }; /* This contains all fds we duplicate here temporarily, and hence need to close at the end */ | |
806 | bool null_readable, null_writable; | |
807 | ||
808 | /* Sets up stdin, stdout, stderr with the three file descriptors passed in. If any of the descriptors is | |
809 | * specified as -1 it will be connected with /dev/null instead. If any of the file descriptors is passed as | |
810 | * itself (e.g. stdin as STDIN_FILENO) it is left unmodified, but the O_CLOEXEC bit is turned off should it be | |
811 | * on. | |
812 | * | |
813 | * Note that if any of the passed file descriptors are > 2 they will be closed — both on success and on | |
814 | * failure! Thus, callers should assume that when this function returns the input fds are invalidated. | |
815 | * | |
816 | * Note that when this function fails stdin/stdout/stderr might remain half set up! | |
817 | * | |
818 | * O_CLOEXEC is turned off for all three file descriptors (which is how it should be for | |
819 | * stdin/stdout/stderr). */ | |
820 | ||
821 | null_readable = original_input_fd < 0; | |
822 | null_writable = original_output_fd < 0 || original_error_fd < 0; | |
823 | ||
824 | /* First step, open /dev/null once, if we need it */ | |
825 | if (null_readable || null_writable) { | |
826 | ||
827 | /* Let's open this with O_CLOEXEC first, and convert it to non-O_CLOEXEC when we move the fd to the final position. */ | |
828 | null_fd = open("/dev/null", (null_readable && null_writable ? O_RDWR : | |
829 | null_readable ? O_RDONLY : O_WRONLY) | O_CLOEXEC); | |
830 | if (null_fd < 0) { | |
831 | r = -errno; | |
832 | goto finish; | |
833 | } | |
834 | ||
835 | /* If this fd is in the 0…2 range, let's move it out of it */ | |
836 | if (null_fd < 3) { | |
837 | int copy; | |
838 | ||
839 | copy = fcntl(null_fd, F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */ | |
840 | if (copy < 0) { | |
841 | r = -errno; | |
842 | goto finish; | |
843 | } | |
844 | ||
845 | safe_close(null_fd); | |
846 | null_fd = copy; | |
847 | } | |
848 | } | |
849 | ||
850 | /* Let's assemble fd[] with the fds to install in place of stdin/stdout/stderr */ | |
851 | for (i = 0; i < 3; i++) { | |
852 | ||
853 | if (fd[i] < 0) | |
854 | fd[i] = null_fd; /* A negative parameter means: connect this one to /dev/null */ | |
855 | else if (fd[i] != i && fd[i] < 3) { | |
856 | /* 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. */ | |
857 | copy_fd[i] = fcntl(fd[i], F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */ | |
858 | if (copy_fd[i] < 0) { | |
859 | r = -errno; | |
860 | goto finish; | |
861 | } | |
862 | ||
863 | fd[i] = copy_fd[i]; | |
864 | } | |
865 | } | |
866 | ||
867 | /* At this point we now have the fds to use in fd[], and they are all above the stdio range, so that we | |
868 | * have freedom to move them around. If the fds already were at the right places then the specific fds are | |
869 | * -1. Let's now move them to the right places. This is the point of no return. */ | |
870 | for (i = 0; i < 3; i++) { | |
871 | ||
872 | if (fd[i] == i) { | |
873 | ||
874 | /* fd is already in place, but let's make sure O_CLOEXEC is off */ | |
875 | r = fd_cloexec(i, false); | |
876 | if (r < 0) | |
877 | goto finish; | |
878 | ||
879 | } else { | |
880 | assert(fd[i] > 2); | |
881 | ||
882 | if (dup2(fd[i], i) < 0) { /* Turns off O_CLOEXEC on the new fd. */ | |
883 | r = -errno; | |
884 | goto finish; | |
885 | } | |
886 | } | |
887 | } | |
888 | ||
889 | r = 0; | |
890 | ||
891 | finish: | |
892 | /* Close the original fds, but only if they were outside of the stdio range. Also, properly check for the same | |
893 | * fd passed in multiple times. */ | |
894 | safe_close_above_stdio(original_input_fd); | |
895 | if (original_output_fd != original_input_fd) | |
896 | safe_close_above_stdio(original_output_fd); | |
897 | if (original_error_fd != original_input_fd && original_error_fd != original_output_fd) | |
898 | safe_close_above_stdio(original_error_fd); | |
899 | ||
900 | /* Close the copies we moved > 2 */ | |
901 | for (i = 0; i < 3; i++) | |
902 | safe_close(copy_fd[i]); | |
903 | ||
904 | /* Close our null fd, if it's > 2 */ | |
905 | safe_close_above_stdio(null_fd); | |
906 | ||
907 | return r; | |
908 | } | |
f2324783 LP |
909 | |
910 | int fd_reopen(int fd, int flags) { | |
911 | char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int)]; | |
912 | int new_fd; | |
913 | ||
914 | /* Reopens the specified fd with new flags. This is useful for convert an O_PATH fd into a regular one, or to | |
915 | * turn O_RDWR fds into O_RDONLY fds. | |
916 | * | |
917 | * This doesn't work on sockets (since they cannot be open()ed, ever). | |
918 | * | |
919 | * This implicitly resets the file read index to 0. */ | |
920 | ||
921 | xsprintf(procfs_path, "/proc/self/fd/%i", fd); | |
922 | new_fd = open(procfs_path, flags); | |
923 | if (new_fd < 0) | |
924 | return -errno; | |
925 | ||
926 | return new_fd; | |
927 | } | |
9264cc39 LP |
928 | |
929 | int read_nr_open(void) { | |
930 | _cleanup_free_ char *nr_open = NULL; | |
931 | int r; | |
932 | ||
933 | /* Returns the kernel's current fd limit, either by reading it of /proc/sys if that works, or using the | |
934 | * hard-coded default compiled-in value of current kernels (1M) if not. This call will never fail. */ | |
935 | ||
936 | r = read_one_line_file("/proc/sys/fs/nr_open", &nr_open); | |
937 | if (r < 0) | |
938 | log_debug_errno(r, "Failed to read /proc/sys/fs/nr_open, ignoring: %m"); | |
939 | else { | |
940 | int v; | |
941 | ||
942 | r = safe_atoi(nr_open, &v); | |
943 | if (r < 0) | |
944 | log_debug_errno(r, "Failed to parse /proc/sys/fs/nr_open value '%s', ignoring: %m", nr_open); | |
945 | else | |
946 | return v; | |
947 | } | |
948 | ||
949 | /* If we fail, fallback to the hard-coded kernel limit of 1024 * 1024. */ | |
950 | return 1024 * 1024; | |
951 | } |