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