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