]>
Commit | Line | Data |
---|---|---|
db9ecf05 | 1 | /* SPDX-License-Identifier: LGPL-2.1-or-later */ |
3ffd4af2 | 2 | |
11c3a366 TA |
3 | #include <errno.h> |
4 | #include <fcntl.h> | |
3657d3a0 | 5 | #if WANT_LINUX_FS_H |
7e93a658 | 6 | #include <linux/fs.h> |
3657d3a0 | 7 | #endif |
65ddc2c5 ZJS |
8 | #include <linux/magic.h> |
9 | #include <sys/ioctl.h> | |
11c3a366 | 10 | #include <sys/resource.h> |
11c3a366 TA |
11 | #include <sys/stat.h> |
12 | #include <unistd.h> | |
13 | ||
4960ce43 | 14 | #include "alloc-util.h" |
8fb3f009 | 15 | #include "dirent-util.h" |
3ffd4af2 | 16 | #include "fd-util.h" |
a548e14d | 17 | #include "fileio.h" |
4aeb20f5 | 18 | #include "fs-util.h" |
4960ce43 | 19 | #include "io-util.h" |
11c3a366 | 20 | #include "macro.h" |
0499585f | 21 | #include "missing_fcntl.h" |
7e93a658 | 22 | #include "missing_fs.h" |
f5947a5e | 23 | #include "missing_syscall.h" |
93cc7779 | 24 | #include "parse-util.h" |
11c3a366 | 25 | #include "path-util.h" |
df0ff127 | 26 | #include "process-util.h" |
93cc7779 | 27 | #include "socket-util.h" |
b8cfa2da | 28 | #include "sort-util.h" |
f8606626 | 29 | #include "stat-util.h" |
4aeb20f5 | 30 | #include "stdio-util.h" |
e4de7287 | 31 | #include "tmpfile-util.h" |
3ffd4af2 | 32 | |
6a461d1f ZJS |
33 | /* The maximum number of iterations in the loop to close descriptors in the fallback case |
34 | * when /proc/self/fd/ is inaccessible. */ | |
35 | #define MAX_FD_LOOP_LIMIT (1024*1024) | |
36 | ||
3ffd4af2 LP |
37 | int close_nointr(int fd) { |
38 | assert(fd >= 0); | |
39 | ||
40 | if (close(fd) >= 0) | |
41 | return 0; | |
42 | ||
43 | /* | |
44 | * Just ignore EINTR; a retry loop is the wrong thing to do on | |
45 | * Linux. | |
46 | * | |
47 | * http://lkml.indiana.edu/hypermail/linux/kernel/0509.1/0877.html | |
48 | * https://bugzilla.gnome.org/show_bug.cgi?id=682819 | |
49 | * http://utcc.utoronto.ca/~cks/space/blog/unix/CloseEINTR | |
50 | * https://sites.google.com/site/michaelsafyan/software-engineering/checkforeintrwheninvokingclosethinkagain | |
51 | */ | |
52 | if (errno == EINTR) | |
53 | return 0; | |
54 | ||
55 | return -errno; | |
56 | } | |
57 | ||
58 | int safe_close(int fd) { | |
59 | ||
60 | /* | |
61 | * Like close_nointr() but cannot fail. Guarantees errno is | |
62 | * unchanged. Is a NOP with negative fds passed, and returns | |
63 | * -1, so that it can be used in this syntax: | |
64 | * | |
65 | * fd = safe_close(fd); | |
66 | */ | |
67 | ||
68 | if (fd >= 0) { | |
69 | PROTECT_ERRNO; | |
70 | ||
71 | /* The kernel might return pretty much any error code | |
72 | * via close(), but the fd will be closed anyway. The | |
73 | * only condition we want to check for here is whether | |
74 | * the fd was invalid at all... */ | |
75 | ||
76 | assert_se(close_nointr(fd) != -EBADF); | |
77 | } | |
78 | ||
79 | return -1; | |
80 | } | |
81 | ||
3042bbeb | 82 | void safe_close_pair(int p[static 2]) { |
3ffd4af2 LP |
83 | assert(p); |
84 | ||
85 | if (p[0] == p[1]) { | |
86 | /* Special case pairs which use the same fd in both | |
87 | * directions... */ | |
88 | p[0] = p[1] = safe_close(p[0]); | |
89 | return; | |
90 | } | |
91 | ||
92 | p[0] = safe_close(p[0]); | |
93 | p[1] = safe_close(p[1]); | |
94 | } | |
95 | ||
da6053d0 | 96 | void close_many(const int fds[], size_t n_fd) { |
3ffd4af2 LP |
97 | assert(fds || n_fd <= 0); |
98 | ||
fe96c0f8 | 99 | for (size_t i = 0; i < n_fd; i++) |
3ffd4af2 LP |
100 | safe_close(fds[i]); |
101 | } | |
102 | ||
103 | int fclose_nointr(FILE *f) { | |
104 | assert(f); | |
105 | ||
106 | /* Same as close_nointr(), but for fclose() */ | |
107 | ||
75f6d5d8 LP |
108 | errno = 0; /* Extra safety: if the FILE* object is not encapsulating an fd, it might not set errno |
109 | * correctly. Let's hence initialize it to zero first, so that we aren't confused by any | |
110 | * prior errno here */ | |
3ffd4af2 LP |
111 | if (fclose(f) == 0) |
112 | return 0; | |
113 | ||
114 | if (errno == EINTR) | |
115 | return 0; | |
116 | ||
75f6d5d8 | 117 | return errno_or_else(EIO); |
3ffd4af2 LP |
118 | } |
119 | ||
120 | FILE* safe_fclose(FILE *f) { | |
121 | ||
122 | /* Same as safe_close(), but for fclose() */ | |
123 | ||
124 | if (f) { | |
125 | PROTECT_ERRNO; | |
126 | ||
6dce3bb4 | 127 | assert_se(fclose_nointr(f) != -EBADF); |
3ffd4af2 LP |
128 | } |
129 | ||
130 | return NULL; | |
131 | } | |
132 | ||
133 | DIR* safe_closedir(DIR *d) { | |
134 | ||
135 | if (d) { | |
136 | PROTECT_ERRNO; | |
137 | ||
138 | assert_se(closedir(d) >= 0 || errno != EBADF); | |
139 | } | |
140 | ||
141 | return NULL; | |
142 | } | |
143 | ||
144 | int fd_nonblock(int fd, bool nonblock) { | |
145 | int flags, nflags; | |
146 | ||
147 | assert(fd >= 0); | |
148 | ||
149 | flags = fcntl(fd, F_GETFL, 0); | |
150 | if (flags < 0) | |
151 | return -errno; | |
152 | ||
0da96503 | 153 | nflags = UPDATE_FLAG(flags, O_NONBLOCK, nonblock); |
3ffd4af2 LP |
154 | if (nflags == flags) |
155 | return 0; | |
156 | ||
7c248223 | 157 | return RET_NERRNO(fcntl(fd, F_SETFL, nflags)); |
3ffd4af2 LP |
158 | } |
159 | ||
160 | int fd_cloexec(int fd, bool cloexec) { | |
161 | int flags, nflags; | |
162 | ||
163 | assert(fd >= 0); | |
164 | ||
165 | flags = fcntl(fd, F_GETFD, 0); | |
166 | if (flags < 0) | |
167 | return -errno; | |
168 | ||
0da96503 | 169 | nflags = UPDATE_FLAG(flags, FD_CLOEXEC, cloexec); |
3ffd4af2 LP |
170 | if (nflags == flags) |
171 | return 0; | |
172 | ||
7c248223 | 173 | return RET_NERRNO(fcntl(fd, F_SETFD, nflags)); |
3ffd4af2 LP |
174 | } |
175 | ||
ed18c22c LP |
176 | int fd_cloexec_many(const int fds[], size_t n_fds, bool cloexec) { |
177 | int ret = 0, r; | |
178 | ||
179 | assert(n_fds == 0 || fds); | |
180 | ||
181 | for (size_t i = 0; i < n_fds; i++) { | |
182 | if (fds[i] < 0) /* Skip gracefully over already invalidated fds */ | |
183 | continue; | |
184 | ||
185 | r = fd_cloexec(fds[i], cloexec); | |
186 | if (r < 0 && ret >= 0) /* Continue going, but return first error */ | |
187 | ret = r; | |
188 | else | |
189 | ret = 1; /* report if we did anything */ | |
190 | } | |
191 | ||
192 | return ret; | |
193 | } | |
194 | ||
da6053d0 | 195 | _pure_ static bool fd_in_set(int fd, const int fdset[], size_t n_fdset) { |
3ffd4af2 LP |
196 | assert(n_fdset == 0 || fdset); |
197 | ||
d11c14a9 LP |
198 | for (size_t i = 0; i < n_fdset; i++) { |
199 | if (fdset[i] < 0) | |
200 | continue; | |
201 | ||
3ffd4af2 LP |
202 | if (fdset[i] == fd) |
203 | return true; | |
d11c14a9 | 204 | } |
3ffd4af2 LP |
205 | |
206 | return false; | |
207 | } | |
208 | ||
73fc0cbc | 209 | int get_max_fd(void) { |
498e265d LP |
210 | struct rlimit rl; |
211 | rlim_t m; | |
212 | ||
213 | /* Return the highest possible fd, based RLIMIT_NOFILE, but enforcing FD_SETSIZE-1 as lower boundary | |
214 | * and INT_MAX as upper boundary. */ | |
215 | ||
216 | if (getrlimit(RLIMIT_NOFILE, &rl) < 0) | |
217 | return -errno; | |
218 | ||
219 | m = MAX(rl.rlim_cur, rl.rlim_max); | |
220 | if (m < FD_SETSIZE) /* Let's always cover at least 1024 fds */ | |
221 | return FD_SETSIZE-1; | |
222 | ||
223 | if (m == RLIM_INFINITY || m > INT_MAX) /* Saturate on overflow. After all fds are "int", hence can | |
224 | * never be above INT_MAX */ | |
225 | return INT_MAX; | |
226 | ||
227 | return (int) (m - 1); | |
228 | } | |
229 | ||
5cfa0798 | 230 | static int close_all_fds_frugal(const int except[], size_t n_except) { |
11966552 LP |
231 | int max_fd, r = 0; |
232 | ||
233 | assert(n_except == 0 || except); | |
234 | ||
235 | /* This is the inner fallback core of close_all_fds(). This never calls malloc() or opendir() or so | |
236 | * and hence is safe to be called in signal handler context. Most users should call close_all_fds(), | |
237 | * but when we assume we are called from signal handler context, then use this simpler call | |
238 | * instead. */ | |
239 | ||
240 | max_fd = get_max_fd(); | |
241 | if (max_fd < 0) | |
242 | return max_fd; | |
243 | ||
244 | /* Refuse to do the loop over more too many elements. It's better to fail immediately than to | |
245 | * spin the CPU for a long time. */ | |
246 | if (max_fd > MAX_FD_LOOP_LIMIT) | |
247 | return log_debug_errno(SYNTHETIC_ERRNO(EPERM), | |
248 | "Refusing to loop over %d potential fds.", | |
249 | max_fd); | |
250 | ||
251 | for (int fd = 3; fd >= 0; fd = fd < max_fd ? fd + 1 : -1) { | |
252 | int q; | |
253 | ||
254 | if (fd_in_set(fd, except, n_except)) | |
255 | continue; | |
256 | ||
257 | q = close_nointr(fd); | |
258 | if (q < 0 && q != -EBADF && r >= 0) | |
259 | r = q; | |
260 | } | |
261 | ||
262 | return r; | |
263 | } | |
264 | ||
5cfa0798 | 265 | static bool have_close_range = true; /* Assume we live in the future */ |
3ffd4af2 | 266 | |
5cfa0798 | 267 | static int close_all_fds_special_case(const int except[], size_t n_except) { |
3ffd4af2 LP |
268 | assert(n_except == 0 || except); |
269 | ||
5cfa0798 LP |
270 | /* Handles a few common special cases separately, since they are common and can be optimized really |
271 | * nicely, since we won't need sorting for them. Returns > 0 if the special casing worked, 0 | |
272 | * otherwise. */ | |
b8cfa2da | 273 | |
5cfa0798 LP |
274 | if (!have_close_range) |
275 | return 0; | |
b8cfa2da | 276 | |
d11c14a9 LP |
277 | if (n_except == 1 && except[0] < 0) /* Minor optimization: if we only got one fd, and it's invalid, |
278 | * we got none */ | |
279 | n_except = 0; | |
280 | ||
5cfa0798 | 281 | switch (n_except) { |
b8cfa2da | 282 | |
5cfa0798 LP |
283 | case 0: |
284 | /* Close everything. Yay! */ | |
b8cfa2da | 285 | |
5cfa0798 LP |
286 | if (close_range(3, -1, 0) >= 0) |
287 | return 1; | |
f498720a | 288 | |
5cfa0798 LP |
289 | if (ERRNO_IS_NOT_SUPPORTED(errno) || ERRNO_IS_PRIVILEGE(errno)) { |
290 | have_close_range = false; | |
291 | return 0; | |
292 | } | |
f498720a | 293 | |
5cfa0798 | 294 | return -errno; |
f498720a | 295 | |
5cfa0798 LP |
296 | case 1: |
297 | /* Close all but exactly one, then we don't need no sorting. This is a pretty common | |
298 | * case, hence let's handle it specially. */ | |
f498720a | 299 | |
5cfa0798 LP |
300 | if ((except[0] <= 3 || close_range(3, except[0]-1, 0) >= 0) && |
301 | (except[0] >= INT_MAX || close_range(MAX(3, except[0]+1), -1, 0) >= 0)) | |
302 | return 1; | |
f498720a | 303 | |
5cfa0798 | 304 | if (ERRNO_IS_NOT_SUPPORTED(errno) || ERRNO_IS_PRIVILEGE(errno)) { |
f498720a | 305 | have_close_range = false; |
5cfa0798 LP |
306 | return 0; |
307 | } | |
f498720a | 308 | |
5cfa0798 | 309 | return -errno; |
c85cb3bc | 310 | |
5cfa0798 LP |
311 | default: |
312 | return 0; | |
313 | } | |
314 | } | |
c85cb3bc | 315 | |
5cfa0798 LP |
316 | int close_all_fds_without_malloc(const int except[], size_t n_except) { |
317 | int r; | |
c85cb3bc | 318 | |
5cfa0798 | 319 | assert(n_except == 0 || except); |
c85cb3bc | 320 | |
5cfa0798 LP |
321 | r = close_all_fds_special_case(except, n_except); |
322 | if (r < 0) | |
323 | return r; | |
324 | if (r > 0) /* special case worked! */ | |
325 | return 0; | |
b8cfa2da | 326 | |
5cfa0798 LP |
327 | return close_all_fds_frugal(except, n_except); |
328 | } | |
b8cfa2da | 329 | |
5cfa0798 LP |
330 | int close_all_fds(const int except[], size_t n_except) { |
331 | _cleanup_closedir_ DIR *d = NULL; | |
5cfa0798 LP |
332 | int r = 0; |
333 | ||
334 | assert(n_except == 0 || except); | |
335 | ||
336 | r = close_all_fds_special_case(except, n_except); | |
337 | if (r < 0) | |
338 | return r; | |
339 | if (r > 0) /* special case worked! */ | |
340 | return 0; | |
341 | ||
342 | if (have_close_range) { | |
343 | _cleanup_free_ int *sorted_malloc = NULL; | |
344 | size_t n_sorted; | |
345 | int *sorted; | |
346 | ||
347 | /* In the best case we have close_range() to close all fds between a start and an end fd, | |
348 | * which we can use on the "inverted" exception array, i.e. all intervals between all | |
349 | * adjacent pairs from the sorted exception array. This changes loop complexity from O(n) | |
350 | * where n is number of open fds to O(m⋅log(m)) where m is the number of fds to keep | |
351 | * open. Given that we assume n ≫ m that's preferable to us. */ | |
b8cfa2da | 352 | |
5cfa0798 LP |
353 | assert(n_except < SIZE_MAX); |
354 | n_sorted = n_except + 1; | |
c85cb3bc | 355 | |
5cfa0798 LP |
356 | if (n_sorted > 64) /* Use heap for large numbers of fds, stack otherwise */ |
357 | sorted = sorted_malloc = new(int, n_sorted); | |
358 | else | |
359 | sorted = newa(int, n_sorted); | |
c85cb3bc | 360 | |
5cfa0798 LP |
361 | if (sorted) { |
362 | memcpy(sorted, except, n_except * sizeof(int)); | |
c85cb3bc | 363 | |
5cfa0798 LP |
364 | /* Let's add fd 2 to the list of fds, to simplify the loop below, as this |
365 | * allows us to cover the head of the array the same way as the body */ | |
366 | sorted[n_sorted-1] = 2; | |
b8cfa2da | 367 | |
5cfa0798 LP |
368 | typesafe_qsort(sorted, n_sorted, cmp_int); |
369 | ||
370 | for (size_t i = 0; i < n_sorted-1; i++) { | |
371 | int start, end; | |
b8cfa2da | 372 | |
5cfa0798 LP |
373 | start = MAX(sorted[i], 2); /* The first three fds shall always remain open */ |
374 | end = MAX(sorted[i+1], 2); | |
b8cfa2da | 375 | |
5cfa0798 | 376 | assert(end >= start); |
b8cfa2da | 377 | |
5cfa0798 LP |
378 | if (end - start <= 1) |
379 | continue; | |
b8cfa2da | 380 | |
5cfa0798 LP |
381 | /* Close everything between the start and end fds (both of which shall stay open) */ |
382 | if (close_range(start + 1, end - 1, 0) < 0) { | |
c85cb3bc LP |
383 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && !ERRNO_IS_PRIVILEGE(errno)) |
384 | return -errno; | |
b8cfa2da | 385 | |
c85cb3bc | 386 | have_close_range = false; |
5cfa0798 | 387 | break; |
c85cb3bc LP |
388 | } |
389 | } | |
5cfa0798 LP |
390 | |
391 | if (have_close_range) { | |
392 | /* The loop succeeded. Let's now close everything beyond the end */ | |
393 | ||
394 | if (sorted[n_sorted-1] >= INT_MAX) /* Dont let the addition below overflow */ | |
395 | return 0; | |
396 | ||
397 | if (close_range(sorted[n_sorted-1] + 1, -1, 0) >= 0) | |
398 | return 0; | |
399 | ||
400 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && !ERRNO_IS_PRIVILEGE(errno)) | |
401 | return -errno; | |
402 | ||
403 | have_close_range = false; | |
404 | } | |
b8cfa2da | 405 | } |
c85cb3bc LP |
406 | |
407 | /* Fallback on OOM or if close_range() is not supported */ | |
b8cfa2da LP |
408 | } |
409 | ||
e7e7c07c | 410 | d = opendir("/proc/self/fd"); |
11966552 | 411 | if (!d) |
5cfa0798 | 412 | return close_all_fds_frugal(except, n_except); /* ultimate fallback if /proc/ is not available */ |
3ffd4af2 | 413 | |
c85cb3bc LP |
414 | FOREACH_DIRENT(de, d, return -errno) { |
415 | int fd = -1, q; | |
3ffd4af2 | 416 | |
1f6639ea LP |
417 | if (!IN_SET(de->d_type, DT_LNK, DT_UNKNOWN)) |
418 | continue; | |
419 | ||
c85cb3bc LP |
420 | if (safe_atoi(de->d_name, &fd) < 0) |
421 | /* Let's better ignore this, just in case */ | |
422 | continue; | |
3ffd4af2 | 423 | |
c85cb3bc LP |
424 | if (fd < 3) |
425 | continue; | |
3ffd4af2 | 426 | |
c85cb3bc LP |
427 | if (fd == dirfd(d)) |
428 | continue; | |
3ffd4af2 LP |
429 | |
430 | if (fd_in_set(fd, except, n_except)) | |
431 | continue; | |
432 | ||
e43bc9f5 | 433 | q = close_nointr(fd); |
c85cb3bc | 434 | if (q < 0 && q != -EBADF && r >= 0) /* Valgrind has its own FD and doesn't want to have it closed */ |
e43bc9f5 | 435 | r = q; |
3ffd4af2 LP |
436 | } |
437 | ||
438 | return r; | |
439 | } | |
440 | ||
441 | int same_fd(int a, int b) { | |
442 | struct stat sta, stb; | |
443 | pid_t pid; | |
444 | int r, fa, fb; | |
445 | ||
446 | assert(a >= 0); | |
447 | assert(b >= 0); | |
448 | ||
675e7fc2 LP |
449 | /* Compares two file descriptors. Note that semantics are quite different depending on whether we |
450 | * have kcmp() or we don't. If we have kcmp() this will only return true for dup()ed file | |
451 | * descriptors, but not otherwise. If we don't have kcmp() this will also return true for two fds of | |
452 | * the same file, created by separate open() calls. Since we use this call mostly for filtering out | |
453 | * duplicates in the fd store this difference hopefully doesn't matter too much. */ | |
3ffd4af2 LP |
454 | |
455 | if (a == b) | |
456 | return true; | |
457 | ||
458 | /* Try to use kcmp() if we have it. */ | |
df0ff127 | 459 | pid = getpid_cached(); |
3ffd4af2 LP |
460 | r = kcmp(pid, pid, KCMP_FILE, a, b); |
461 | if (r == 0) | |
462 | return true; | |
463 | if (r > 0) | |
464 | return false; | |
675e7fc2 | 465 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && !ERRNO_IS_PRIVILEGE(errno)) |
3ffd4af2 LP |
466 | return -errno; |
467 | ||
468 | /* We don't have kcmp(), use fstat() instead. */ | |
469 | if (fstat(a, &sta) < 0) | |
470 | return -errno; | |
471 | ||
472 | if (fstat(b, &stb) < 0) | |
473 | return -errno; | |
474 | ||
a9dac7a6 | 475 | if (!stat_inode_same(&sta, &stb)) |
3ffd4af2 LP |
476 | return false; |
477 | ||
675e7fc2 LP |
478 | /* We consider all device fds different, since two device fds might refer to quite different device |
479 | * contexts even though they share the same inode and backing dev_t. */ | |
3ffd4af2 LP |
480 | |
481 | if (S_ISCHR(sta.st_mode) || S_ISBLK(sta.st_mode)) | |
482 | return false; | |
483 | ||
a9dac7a6 LP |
484 | /* The fds refer to the same inode on disk, let's also check if they have the same fd flags. This is |
485 | * useful to distinguish the read and write side of a pipe created with pipe(). */ | |
3ffd4af2 LP |
486 | fa = fcntl(a, F_GETFL); |
487 | if (fa < 0) | |
488 | return -errno; | |
489 | ||
490 | fb = fcntl(b, F_GETFL); | |
491 | if (fb < 0) | |
492 | return -errno; | |
493 | ||
494 | return fa == fb; | |
495 | } | |
496 | ||
497 | void cmsg_close_all(struct msghdr *mh) { | |
498 | struct cmsghdr *cmsg; | |
499 | ||
500 | assert(mh); | |
501 | ||
502 | CMSG_FOREACH(cmsg, mh) | |
503 | if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS) | |
504 | close_many((int*) CMSG_DATA(cmsg), (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int)); | |
505 | } | |
4fee3975 LP |
506 | |
507 | bool fdname_is_valid(const char *s) { | |
508 | const char *p; | |
509 | ||
510 | /* Validates a name for $LISTEN_FDNAMES. We basically allow | |
511 | * everything ASCII that's not a control character. Also, as | |
512 | * special exception the ":" character is not allowed, as we | |
513 | * use that as field separator in $LISTEN_FDNAMES. | |
514 | * | |
515 | * Note that the empty string is explicitly allowed | |
516 | * here. However, we limit the length of the names to 255 | |
517 | * characters. */ | |
518 | ||
519 | if (!s) | |
520 | return false; | |
521 | ||
522 | for (p = s; *p; p++) { | |
523 | if (*p < ' ') | |
524 | return false; | |
525 | if (*p >= 127) | |
526 | return false; | |
527 | if (*p == ':') | |
528 | return false; | |
529 | } | |
530 | ||
ae3f4bae | 531 | return p - s <= FDNAME_MAX; |
4fee3975 | 532 | } |
4aeb20f5 LP |
533 | |
534 | int fd_get_path(int fd, char **ret) { | |
a0fe2a2d | 535 | int r; |
4aeb20f5 | 536 | |
ddb6eeaf | 537 | r = readlink_malloc(FORMAT_PROC_FD_PATH(fd), ret); |
f267719c LP |
538 | if (r == -ENOENT) { |
539 | /* ENOENT can mean two things: that the fd does not exist or that /proc is not mounted. Let's make | |
5238e957 | 540 | * things debuggable and distinguish the two. */ |
4aeb20f5 | 541 | |
8fe8f3aa LP |
542 | if (proc_mounted() == 0) |
543 | return -ENOSYS; /* /proc is not available or not set up properly, we're most likely in some chroot | |
544 | * environment. */ | |
f267719c LP |
545 | return -EBADF; /* The directory exists, hence it's the fd that doesn't. */ |
546 | } | |
a0fe2a2d LP |
547 | |
548 | return r; | |
4aeb20f5 | 549 | } |
046a82c1 LP |
550 | |
551 | int move_fd(int from, int to, int cloexec) { | |
552 | int r; | |
553 | ||
554 | /* Move fd 'from' to 'to', make sure FD_CLOEXEC remains equal if requested, and release the old fd. If | |
555 | * 'cloexec' is passed as -1, the original FD_CLOEXEC is inherited for the new fd. If it is 0, it is turned | |
556 | * off, if it is > 0 it is turned on. */ | |
557 | ||
558 | if (from < 0) | |
559 | return -EBADF; | |
560 | if (to < 0) | |
561 | return -EBADF; | |
562 | ||
563 | if (from == to) { | |
564 | ||
565 | if (cloexec >= 0) { | |
566 | r = fd_cloexec(to, cloexec); | |
567 | if (r < 0) | |
568 | return r; | |
569 | } | |
570 | ||
571 | return to; | |
572 | } | |
573 | ||
574 | if (cloexec < 0) { | |
575 | int fl; | |
576 | ||
577 | fl = fcntl(from, F_GETFD, 0); | |
578 | if (fl < 0) | |
579 | return -errno; | |
580 | ||
581 | cloexec = !!(fl & FD_CLOEXEC); | |
582 | } | |
583 | ||
584 | r = dup3(from, to, cloexec ? O_CLOEXEC : 0); | |
585 | if (r < 0) | |
586 | return -errno; | |
587 | ||
588 | assert(r == to); | |
589 | ||
590 | safe_close(from); | |
591 | ||
592 | return to; | |
593 | } | |
a548e14d | 594 | |
7fe2903c LP |
595 | int fd_move_above_stdio(int fd) { |
596 | int flags, copy; | |
597 | PROTECT_ERRNO; | |
598 | ||
599 | /* Moves the specified file descriptor if possible out of the range [0…2], i.e. the range of | |
600 | * stdin/stdout/stderr. If it can't be moved outside of this range the original file descriptor is | |
601 | * returned. This call is supposed to be used for long-lasting file descriptors we allocate in our code that | |
602 | * might get loaded into foreign code, and where we want ensure our fds are unlikely used accidentally as | |
603 | * stdin/stdout/stderr of unrelated code. | |
604 | * | |
605 | * Note that this doesn't fix any real bugs, it just makes it less likely that our code will be affected by | |
606 | * buggy code from others that mindlessly invokes 'fprintf(stderr, …' or similar in places where stderr has | |
607 | * been closed before. | |
608 | * | |
609 | * This function is written in a "best-effort" and "least-impact" style. This means whenever we encounter an | |
610 | * error we simply return the original file descriptor, and we do not touch errno. */ | |
611 | ||
612 | if (fd < 0 || fd > 2) | |
613 | return fd; | |
614 | ||
615 | flags = fcntl(fd, F_GETFD, 0); | |
616 | if (flags < 0) | |
617 | return fd; | |
618 | ||
619 | if (flags & FD_CLOEXEC) | |
620 | copy = fcntl(fd, F_DUPFD_CLOEXEC, 3); | |
621 | else | |
622 | copy = fcntl(fd, F_DUPFD, 3); | |
623 | if (copy < 0) | |
624 | return fd; | |
625 | ||
626 | assert(copy > 2); | |
627 | ||
628 | (void) close(fd); | |
629 | return copy; | |
630 | } | |
aa11e28b LP |
631 | |
632 | int rearrange_stdio(int original_input_fd, int original_output_fd, int original_error_fd) { | |
633 | ||
634 | int fd[3] = { /* Put together an array of fds we work on */ | |
635 | original_input_fd, | |
636 | original_output_fd, | |
637 | original_error_fd | |
638 | }; | |
639 | ||
640 | int r, i, | |
641 | null_fd = -1, /* if we open /dev/null, we store the fd to it here */ | |
642 | copy_fd[3] = { -1, -1, -1 }; /* This contains all fds we duplicate here temporarily, and hence need to close at the end */ | |
643 | bool null_readable, null_writable; | |
644 | ||
645 | /* Sets up stdin, stdout, stderr with the three file descriptors passed in. If any of the descriptors is | |
646 | * specified as -1 it will be connected with /dev/null instead. If any of the file descriptors is passed as | |
647 | * itself (e.g. stdin as STDIN_FILENO) it is left unmodified, but the O_CLOEXEC bit is turned off should it be | |
648 | * on. | |
649 | * | |
650 | * Note that if any of the passed file descriptors are > 2 they will be closed — both on success and on | |
651 | * failure! Thus, callers should assume that when this function returns the input fds are invalidated. | |
652 | * | |
653 | * Note that when this function fails stdin/stdout/stderr might remain half set up! | |
654 | * | |
655 | * O_CLOEXEC is turned off for all three file descriptors (which is how it should be for | |
656 | * stdin/stdout/stderr). */ | |
657 | ||
658 | null_readable = original_input_fd < 0; | |
659 | null_writable = original_output_fd < 0 || original_error_fd < 0; | |
660 | ||
661 | /* First step, open /dev/null once, if we need it */ | |
662 | if (null_readable || null_writable) { | |
663 | ||
664 | /* Let's open this with O_CLOEXEC first, and convert it to non-O_CLOEXEC when we move the fd to the final position. */ | |
665 | null_fd = open("/dev/null", (null_readable && null_writable ? O_RDWR : | |
666 | null_readable ? O_RDONLY : O_WRONLY) | O_CLOEXEC); | |
667 | if (null_fd < 0) { | |
668 | r = -errno; | |
669 | goto finish; | |
670 | } | |
671 | ||
672 | /* If this fd is in the 0…2 range, let's move it out of it */ | |
673 | if (null_fd < 3) { | |
674 | int copy; | |
675 | ||
676 | copy = fcntl(null_fd, F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */ | |
677 | if (copy < 0) { | |
678 | r = -errno; | |
679 | goto finish; | |
680 | } | |
681 | ||
ee3455cf | 682 | close_and_replace(null_fd, copy); |
aa11e28b LP |
683 | } |
684 | } | |
685 | ||
686 | /* Let's assemble fd[] with the fds to install in place of stdin/stdout/stderr */ | |
687 | for (i = 0; i < 3; i++) { | |
688 | ||
689 | if (fd[i] < 0) | |
690 | fd[i] = null_fd; /* A negative parameter means: connect this one to /dev/null */ | |
691 | else if (fd[i] != i && fd[i] < 3) { | |
692 | /* 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. */ | |
693 | copy_fd[i] = fcntl(fd[i], F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */ | |
694 | if (copy_fd[i] < 0) { | |
695 | r = -errno; | |
696 | goto finish; | |
697 | } | |
698 | ||
699 | fd[i] = copy_fd[i]; | |
700 | } | |
701 | } | |
702 | ||
703 | /* At this point we now have the fds to use in fd[], and they are all above the stdio range, so that we | |
704 | * have freedom to move them around. If the fds already were at the right places then the specific fds are | |
705 | * -1. Let's now move them to the right places. This is the point of no return. */ | |
706 | for (i = 0; i < 3; i++) { | |
707 | ||
708 | if (fd[i] == i) { | |
709 | ||
710 | /* fd is already in place, but let's make sure O_CLOEXEC is off */ | |
711 | r = fd_cloexec(i, false); | |
712 | if (r < 0) | |
713 | goto finish; | |
714 | ||
715 | } else { | |
716 | assert(fd[i] > 2); | |
717 | ||
718 | if (dup2(fd[i], i) < 0) { /* Turns off O_CLOEXEC on the new fd. */ | |
719 | r = -errno; | |
720 | goto finish; | |
721 | } | |
722 | } | |
723 | } | |
724 | ||
725 | r = 0; | |
726 | ||
727 | finish: | |
728 | /* Close the original fds, but only if they were outside of the stdio range. Also, properly check for the same | |
729 | * fd passed in multiple times. */ | |
730 | safe_close_above_stdio(original_input_fd); | |
731 | if (original_output_fd != original_input_fd) | |
732 | safe_close_above_stdio(original_output_fd); | |
733 | if (original_error_fd != original_input_fd && original_error_fd != original_output_fd) | |
734 | safe_close_above_stdio(original_error_fd); | |
735 | ||
736 | /* Close the copies we moved > 2 */ | |
737 | for (i = 0; i < 3; i++) | |
738 | safe_close(copy_fd[i]); | |
739 | ||
740 | /* Close our null fd, if it's > 2 */ | |
741 | safe_close_above_stdio(null_fd); | |
742 | ||
743 | return r; | |
744 | } | |
f2324783 LP |
745 | |
746 | int fd_reopen(int fd, int flags) { | |
d6274e6b | 747 | int new_fd, r; |
f2324783 LP |
748 | |
749 | /* Reopens the specified fd with new flags. This is useful for convert an O_PATH fd into a regular one, or to | |
750 | * turn O_RDWR fds into O_RDONLY fds. | |
751 | * | |
752 | * This doesn't work on sockets (since they cannot be open()ed, ever). | |
753 | * | |
754 | * This implicitly resets the file read index to 0. */ | |
755 | ||
b4f73d1e LP |
756 | if (FLAGS_SET(flags, O_DIRECTORY)) { |
757 | /* If we shall reopen the fd as directory we can just go via "." and thus bypass the whole | |
758 | * magic /proc/ directory, and make ourselves independent of that being mounted. */ | |
759 | new_fd = openat(fd, ".", flags); | |
760 | if (new_fd < 0) | |
761 | return -errno; | |
762 | ||
763 | return new_fd; | |
764 | } | |
765 | ||
ddb6eeaf | 766 | new_fd = open(FORMAT_PROC_FD_PATH(fd), flags); |
f8606626 LP |
767 | if (new_fd < 0) { |
768 | if (errno != ENOENT) | |
769 | return -errno; | |
770 | ||
d6274e6b LP |
771 | r = proc_mounted(); |
772 | if (r == 0) | |
f8606626 LP |
773 | return -ENOSYS; /* if we have no /proc/, the concept is not implementable */ |
774 | ||
d6274e6b LP |
775 | return r > 0 ? -EBADF : -ENOENT; /* If /proc/ is definitely around then this means the fd is |
776 | * not valid, otherwise let's propagate the original | |
777 | * error */ | |
f8606626 | 778 | } |
f2324783 LP |
779 | |
780 | return new_fd; | |
781 | } | |
9264cc39 | 782 | |
5f5865f0 LP |
783 | int fd_reopen_condition( |
784 | int fd, | |
785 | int flags, | |
786 | int mask, | |
787 | int *ret_new_fd) { | |
788 | ||
789 | int r, new_fd; | |
790 | ||
791 | assert(fd >= 0); | |
792 | ||
793 | /* Invokes fd_reopen(fd, flags), but only if the existing F_GETFL flags don't match the specified | |
794 | * flags (masked by the specified mask). This is useful for converting O_PATH fds into real fds if | |
795 | * needed, but only then. */ | |
796 | ||
797 | r = fcntl(fd, F_GETFL); | |
798 | if (r < 0) | |
799 | return -errno; | |
800 | ||
801 | if ((r & mask) == (flags & mask)) { | |
802 | *ret_new_fd = -1; | |
803 | return fd; | |
804 | } | |
805 | ||
806 | new_fd = fd_reopen(fd, flags); | |
807 | if (new_fd < 0) | |
808 | return new_fd; | |
809 | ||
810 | *ret_new_fd = new_fd; | |
811 | return new_fd; | |
812 | } | |
813 | ||
9264cc39 LP |
814 | int read_nr_open(void) { |
815 | _cleanup_free_ char *nr_open = NULL; | |
816 | int r; | |
817 | ||
818 | /* Returns the kernel's current fd limit, either by reading it of /proc/sys if that works, or using the | |
819 | * hard-coded default compiled-in value of current kernels (1M) if not. This call will never fail. */ | |
820 | ||
821 | r = read_one_line_file("/proc/sys/fs/nr_open", &nr_open); | |
822 | if (r < 0) | |
823 | log_debug_errno(r, "Failed to read /proc/sys/fs/nr_open, ignoring: %m"); | |
824 | else { | |
825 | int v; | |
826 | ||
827 | r = safe_atoi(nr_open, &v); | |
828 | if (r < 0) | |
829 | log_debug_errno(r, "Failed to parse /proc/sys/fs/nr_open value '%s', ignoring: %m", nr_open); | |
830 | else | |
831 | return v; | |
832 | } | |
833 | ||
2aed63f4 | 834 | /* If we fail, fall back to the hard-coded kernel limit of 1024 * 1024. */ |
9264cc39 LP |
835 | return 1024 * 1024; |
836 | } | |
65ddc2c5 | 837 | |
7e93a658 YW |
838 | int fd_get_diskseq(int fd, uint64_t *ret) { |
839 | uint64_t diskseq; | |
840 | ||
841 | assert(fd >= 0); | |
842 | assert(ret); | |
843 | ||
844 | if (ioctl(fd, BLKGETDISKSEQ, &diskseq) < 0) { | |
845 | /* Note that the kernel is weird: non-existing ioctls currently return EINVAL | |
846 | * rather than ENOTTY on loopback block devices. They should fix that in the kernel, | |
847 | * but in the meantime we accept both here. */ | |
848 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && errno != EINVAL) | |
849 | return -errno; | |
850 | ||
851 | return -EOPNOTSUPP; | |
852 | } | |
853 | ||
854 | *ret = diskseq; | |
855 | ||
856 | return 0; | |
857 | } |