1 /* SPDX-License-Identifier: LGPL-2.1+ */
3 Copyright 2010 Lennart Poettering
8 #include <sys/resource.h>
9 #include <sys/socket.h>
13 //#include "alloc-util.h"
15 #include "dirent-util.h"
19 //#include "io-util.h"
21 #include "memfd-util.h"
23 #include "parse-util.h"
24 #include "path-util.h"
25 #include "process-util.h"
26 #include "socket-util.h"
27 #include "stdio-util.h"
30 int close_nointr(int fd) {
37 * Just ignore EINTR; a retry loop is the wrong thing to do on
40 * http://lkml.indiana.edu/hypermail/linux/kernel/0509.1/0877.html
41 * https://bugzilla.gnome.org/show_bug.cgi?id=682819
42 * http://utcc.utoronto.ca/~cks/space/blog/unix/CloseEINTR
43 * https://sites.google.com/site/michaelsafyan/software-engineering/checkforeintrwheninvokingclosethinkagain
51 int safe_close(int fd) {
54 * Like close_nointr() but cannot fail. Guarantees errno is
55 * unchanged. Is a NOP with negative fds passed, and returns
56 * -1, so that it can be used in this syntax:
58 * fd = safe_close(fd);
64 /* The kernel might return pretty much any error code
65 * via close(), but the fd will be closed anyway. The
66 * only condition we want to check for here is whether
67 * the fd was invalid at all... */
69 assert_se(close_nointr(fd) != -EBADF);
75 void safe_close_pair(int p[]) {
79 /* Special case pairs which use the same fd in both
81 p[0] = p[1] = safe_close(p[0]);
85 p[0] = safe_close(p[0]);
86 p[1] = safe_close(p[1]);
89 void close_many(const int fds[], size_t n_fd) {
92 assert(fds || n_fd <= 0);
94 for (i = 0; i < n_fd; i++)
98 int fclose_nointr(FILE *f) {
101 /* Same as close_nointr(), but for fclose() */
112 FILE* safe_fclose(FILE *f) {
114 /* Same as safe_close(), but for fclose() */
119 assert_se(fclose_nointr(f) != EBADF);
125 #if 0 /// UNNEEDED by elogind
126 DIR* safe_closedir(DIR *d) {
131 assert_se(closedir(d) >= 0 || errno != EBADF);
138 int fd_nonblock(int fd, bool nonblock) {
143 flags = fcntl(fd, F_GETFL, 0);
148 nflags = flags | O_NONBLOCK;
150 nflags = flags & ~O_NONBLOCK;
155 if (fcntl(fd, F_SETFL, nflags) < 0)
161 int fd_cloexec(int fd, bool cloexec) {
166 flags = fcntl(fd, F_GETFD, 0);
171 nflags = flags | FD_CLOEXEC;
173 nflags = flags & ~FD_CLOEXEC;
178 if (fcntl(fd, F_SETFD, nflags) < 0)
184 _pure_ static bool fd_in_set(int fd, const int fdset[], size_t n_fdset) {
187 assert(n_fdset == 0 || fdset);
189 for (i = 0; i < n_fdset; i++)
196 int close_all_fds(const int except[], size_t n_except) {
197 _cleanup_closedir_ DIR *d = NULL;
201 assert(n_except == 0 || except);
203 d = opendir("/proc/self/fd");
208 /* When /proc isn't available (for example in chroots) the fallback is brute forcing through the fd
211 assert_se(getrlimit(RLIMIT_NOFILE, &rl) >= 0);
213 if (rl.rlim_max == 0)
216 /* Let's take special care if the resource limit is set to unlimited, or actually larger than the range
217 * of 'int'. Let's avoid implicit overflows. */
218 max_fd = (rl.rlim_max == RLIM_INFINITY || rl.rlim_max > INT_MAX) ? INT_MAX : (int) (rl.rlim_max - 1);
220 for (fd = 3; fd >= 0; fd = fd < max_fd ? fd + 1 : -1) {
223 if (fd_in_set(fd, except, n_except))
226 q = close_nointr(fd);
227 if (q < 0 && q != -EBADF && r >= 0)
234 FOREACH_DIRENT(de, d, return -errno) {
237 if (safe_atoi(de->d_name, &fd) < 0)
238 /* Let's better ignore this, just in case */
247 if (fd_in_set(fd, except, n_except))
250 q = close_nointr(fd);
251 if (q < 0 && q != -EBADF && r >= 0) /* Valgrind has its own FD and doesn't want to have it closed */
258 #if 0 /// UNNEEDED by elogind
259 int same_fd(int a, int b) {
260 struct stat sta, stb;
267 /* Compares two file descriptors. Note that semantics are
268 * quite different depending on whether we have kcmp() or we
269 * don't. If we have kcmp() this will only return true for
270 * dup()ed file descriptors, but not otherwise. If we don't
271 * have kcmp() this will also return true for two fds of the same
272 * file, created by separate open() calls. Since we use this
273 * call mostly for filtering out duplicates in the fd store
274 * this difference hopefully doesn't matter too much. */
279 /* Try to use kcmp() if we have it. */
280 pid = getpid_cached();
281 r = kcmp(pid, pid, KCMP_FILE, a, b);
289 /* We don't have kcmp(), use fstat() instead. */
290 if (fstat(a, &sta) < 0)
293 if (fstat(b, &stb) < 0)
296 if ((sta.st_mode & S_IFMT) != (stb.st_mode & S_IFMT))
299 /* We consider all device fds different, since two device fds
300 * might refer to quite different device contexts even though
301 * they share the same inode and backing dev_t. */
303 if (S_ISCHR(sta.st_mode) || S_ISBLK(sta.st_mode))
306 if (sta.st_dev != stb.st_dev || sta.st_ino != stb.st_ino)
309 /* The fds refer to the same inode on disk, let's also check
310 * if they have the same fd flags. This is useful to
311 * distinguish the read and write side of a pipe created with
313 fa = fcntl(a, F_GETFL);
317 fb = fcntl(b, F_GETFL);
324 void cmsg_close_all(struct msghdr *mh) {
325 struct cmsghdr *cmsg;
329 CMSG_FOREACH(cmsg, mh)
330 if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS)
331 close_many((int*) CMSG_DATA(cmsg), (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int));
334 bool fdname_is_valid(const char *s) {
337 /* Validates a name for $LISTEN_FDNAMES. We basically allow
338 * everything ASCII that's not a control character. Also, as
339 * special exception the ":" character is not allowed, as we
340 * use that as field separator in $LISTEN_FDNAMES.
342 * Note that the empty string is explicitly allowed
343 * here. However, we limit the length of the names to 255
349 for (p = s; *p; p++) {
362 int fd_get_path(int fd, char **ret) {
363 _cleanup_close_ int dir = -1;
364 char fdname[DECIMAL_STR_MAX(int)];
367 dir = open("/proc/self/fd/", O_CLOEXEC | O_DIRECTORY | O_PATH);
369 /* /proc is not available or not set up properly, we're most likely
370 * in some chroot environment. */
371 return errno == ENOENT ? -EOPNOTSUPP : -errno;
373 xsprintf(fdname, "%i", fd);
375 r = readlinkat_malloc(dir, fdname, ret);
377 /* If the file doesn't exist the fd is invalid */
383 int move_fd(int from, int to, int cloexec) {
386 /* Move fd 'from' to 'to', make sure FD_CLOEXEC remains equal if requested, and release the old fd. If
387 * 'cloexec' is passed as -1, the original FD_CLOEXEC is inherited for the new fd. If it is 0, it is turned
388 * off, if it is > 0 it is turned on. */
398 r = fd_cloexec(to, cloexec);
409 fl = fcntl(from, F_GETFD, 0);
413 cloexec = !!(fl & FD_CLOEXEC);
416 r = dup3(from, to, cloexec ? O_CLOEXEC : 0);
427 int acquire_data_fd(const void *data, size_t size, unsigned flags) {
429 _cleanup_close_pair_ int pipefds[2] = { -1, -1 };
430 char pattern[] = "/dev/shm/data-fd-XXXXXX";
431 _cleanup_close_ int fd = -1;
436 assert(data || size == 0);
438 /* Acquire a read-only file descriptor that when read from returns the specified data. This is much more
439 * complex than I wish it was. But here's why:
441 * a) First we try to use memfds. They are the best option, as we can seal them nicely to make them
442 * read-only. Unfortunately they require kernel 3.17, and – at the time of writing – we still support 3.14.
444 * b) Then, we try classic pipes. They are the second best options, as we can close the writing side, retaining
445 * a nicely read-only fd in the reading side. However, they are by default quite small, and unprivileged
446 * clients can only bump their size to a system-wide limit, which might be quite low.
448 * c) Then, we try an O_TMPFILE file in /dev/shm (that dir is the only suitable one known to exist from
449 * earliest boot on). To make it read-only we open the fd a second time with O_RDONLY via
450 * /proc/self/<fd>. Unfortunately O_TMPFILE is not available on older kernels on tmpfs.
452 * d) Finally, we try creating a regular file in /dev/shm, which we then delete.
454 * It sucks a bit that depending on the situation we return very different objects here, but that's Linux I
457 if (size == 0 && ((flags & ACQUIRE_NO_DEV_NULL) == 0)) {
458 /* As a special case, return /dev/null if we have been called for an empty data block */
459 r = open("/dev/null", O_RDONLY|O_CLOEXEC|O_NOCTTY);
466 if ((flags & ACQUIRE_NO_MEMFD) == 0) {
467 fd = memfd_new("data-fd");
471 n = write(fd, data, size);
474 if ((size_t) n != size)
477 f = lseek(fd, 0, SEEK_SET);
481 r = memfd_set_sealed(fd);
489 if ((flags & ACQUIRE_NO_PIPE) == 0) {
490 if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0)
493 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
497 if ((size_t) isz < size) {
499 if (isz < 0 || (size_t) isz != size)
502 /* Try to bump the pipe size */
503 (void) fcntl(pipefds[1], F_SETPIPE_SZ, isz);
505 /* See if that worked */
506 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
510 if ((size_t) isz < size)
514 n = write(pipefds[1], data, size);
517 if ((size_t) n != size)
520 (void) fd_nonblock(pipefds[0], false);
522 return TAKE_FD(pipefds[0]);
526 if ((flags & ACQUIRE_NO_TMPFILE) == 0) {
527 fd = open("/dev/shm", O_RDWR|O_TMPFILE|O_CLOEXEC, 0500);
529 goto try_dev_shm_without_o_tmpfile;
531 n = write(fd, data, size);
534 if ((size_t) n != size)
537 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
538 return fd_reopen(fd, O_RDONLY|O_CLOEXEC);
541 try_dev_shm_without_o_tmpfile:
542 if ((flags & ACQUIRE_NO_REGULAR) == 0) {
543 fd = mkostemp_safe(pattern);
547 n = write(fd, data, size);
550 goto unlink_and_return;
552 if ((size_t) n != size) {
554 goto unlink_and_return;
557 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
558 r = open(pattern, O_RDONLY|O_CLOEXEC);
563 (void) unlink(pattern);
570 /* When the data is smaller or equal to 64K, try to place the copy in a memfd/pipe */
571 #define DATA_FD_MEMORY_LIMIT (64U*1024U)
573 /* 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. */
574 #define DATA_FD_TMP_LIMIT (1024U*1024U)
576 int fd_duplicate_data_fd(int fd) {
578 _cleanup_close_ int copy_fd = -1, tmp_fd = -1;
579 _cleanup_free_ void *remains = NULL;
580 size_t remains_size = 0;
585 /* Creates a 'data' fd from the specified source fd, containing all the same data in a read-only fashion, but
586 * independent of it (i.e. the source fd can be closed and unmounted after this call succeeded). Tries to be
587 * somewhat smart about where to place the data. In the best case uses a memfd(). If memfd() are not supported
588 * uses a pipe instead. For larger data will use an unlinked file in /tmp, and for even larger data one in
591 if (fstat(fd, &st) < 0)
594 /* For now, let's only accept regular files, sockets, pipes and char devices */
595 if (S_ISDIR(st.st_mode))
597 if (S_ISLNK(st.st_mode))
599 if (!S_ISREG(st.st_mode) && !S_ISSOCK(st.st_mode) && !S_ISFIFO(st.st_mode) && !S_ISCHR(st.st_mode))
602 /* If we have reason to believe the data is bounded in size, then let's use memfds or pipes as backing fd. Note
603 * that we use the reported regular file size only as a hint, given that there are plenty special files in
604 * /proc and /sys which report a zero file size but can be read from. */
606 if (!S_ISREG(st.st_mode) || st.st_size < DATA_FD_MEMORY_LIMIT) {
608 /* Try a memfd first */
609 copy_fd = memfd_new("data-fd");
613 r = copy_bytes(fd, copy_fd, DATA_FD_MEMORY_LIMIT, 0);
617 f = lseek(copy_fd, 0, SEEK_SET);
622 /* Did it fit into the limit? If so, we are done. */
623 r = memfd_set_sealed(copy_fd);
627 return TAKE_FD(copy_fd);
630 /* Hmm, pity, this didn't fit. Let's fall back to /tmp then, see below */
633 _cleanup_(close_pairp) int pipefds[2] = { -1, -1 };
636 /* If memfds aren't available, use a pipe. Set O_NONBLOCK so that we will get EAGAIN rather
637 * then block indefinitely when we hit the pipe size limit */
639 if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0)
642 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
646 /* Try to enlarge the pipe size if necessary */
647 if ((size_t) isz < DATA_FD_MEMORY_LIMIT) {
649 (void) fcntl(pipefds[1], F_SETPIPE_SZ, DATA_FD_MEMORY_LIMIT);
651 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
656 if ((size_t) isz >= DATA_FD_MEMORY_LIMIT) {
658 r = copy_bytes_full(fd, pipefds[1], DATA_FD_MEMORY_LIMIT, 0, &remains, &remains_size);
659 if (r < 0 && r != -EAGAIN)
660 return r; /* If we get EAGAIN it could be because of the source or because of
661 * the destination fd, we can't know, as sendfile() and friends won't
662 * tell us. Hence, treat this as reason to fall back, just to be
665 /* Everything fit in, yay! */
666 (void) fd_nonblock(pipefds[0], false);
668 return TAKE_FD(pipefds[0]);
671 /* Things didn't fit in. But we read data into the pipe, let's remember that, so that
672 * when writing the new file we incorporate this first. */
673 copy_fd = TAKE_FD(pipefds[0]);
678 /* If we have reason to believe this will fit fine in /tmp, then use that as first fallback. */
679 if ((!S_ISREG(st.st_mode) || st.st_size < DATA_FD_TMP_LIMIT) &&
680 (DATA_FD_MEMORY_LIMIT + remains_size) < DATA_FD_TMP_LIMIT) {
683 tmp_fd = open_tmpfile_unlinkable(NULL /* NULL as directory means /tmp */, O_RDWR|O_CLOEXEC);
688 /* If we tried a memfd/pipe first and it ended up being too large, then copy this into the
689 * temporary file first. */
691 r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, 0);
698 if (remains_size > 0) {
699 /* If there were remaining bytes (i.e. read into memory, but not written out yet) from the
700 * failed copy operation, let's flush them out next. */
702 r = loop_write(tmp_fd, remains, remains_size, false);
707 r = copy_bytes(fd, tmp_fd, DATA_FD_TMP_LIMIT - DATA_FD_MEMORY_LIMIT - remains_size, COPY_REFLINK);
711 goto finish; /* Yay, it fit in */
713 /* It didn't fit in. Let's not forget to use what we already used */
714 f = lseek(tmp_fd, 0, SEEK_SET);
719 copy_fd = TAKE_FD(tmp_fd);
721 remains = mfree(remains);
725 /* As last fallback use /var/tmp */
726 r = var_tmp_dir(&td);
730 tmp_fd = open_tmpfile_unlinkable(td, O_RDWR|O_CLOEXEC);
735 /* If we tried a memfd/pipe first, or a file in /tmp, and it ended up being too large, than copy this
736 * into the temporary file first. */
737 r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
744 if (remains_size > 0) {
745 /* Then, copy in any read but not yet written bytes. */
746 r = loop_write(tmp_fd, remains, remains_size, false);
751 /* Copy in the rest */
752 r = copy_bytes(fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
759 /* Now convert the O_RDWR file descriptor into an O_RDONLY one (and as side effect seek to the beginning of the
762 return fd_reopen(tmp_fd, O_RDONLY|O_CLOEXEC);
765 int fd_move_above_stdio(int fd) {
769 /* Moves the specified file descriptor if possible out of the range [0…2], i.e. the range of
770 * stdin/stdout/stderr. If it can't be moved outside of this range the original file descriptor is
771 * returned. This call is supposed to be used for long-lasting file descriptors we allocate in our code that
772 * might get loaded into foreign code, and where we want ensure our fds are unlikely used accidentally as
773 * stdin/stdout/stderr of unrelated code.
775 * Note that this doesn't fix any real bugs, it just makes it less likely that our code will be affected by
776 * buggy code from others that mindlessly invokes 'fprintf(stderr, …' or similar in places where stderr has
777 * been closed before.
779 * This function is written in a "best-effort" and "least-impact" style. This means whenever we encounter an
780 * error we simply return the original file descriptor, and we do not touch errno. */
782 if (fd < 0 || fd > 2)
785 flags = fcntl(fd, F_GETFD, 0);
789 if (flags & FD_CLOEXEC)
790 copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
792 copy = fcntl(fd, F_DUPFD, 3);
802 int rearrange_stdio(int original_input_fd, int original_output_fd, int original_error_fd) {
804 int fd[3] = { /* Put together an array of fds we work on */
811 null_fd = -1, /* if we open /dev/null, we store the fd to it here */
812 copy_fd[3] = { -1, -1, -1 }; /* This contains all fds we duplicate here temporarily, and hence need to close at the end */
813 bool null_readable, null_writable;
815 /* Sets up stdin, stdout, stderr with the three file descriptors passed in. If any of the descriptors is
816 * specified as -1 it will be connected with /dev/null instead. If any of the file descriptors is passed as
817 * itself (e.g. stdin as STDIN_FILENO) it is left unmodified, but the O_CLOEXEC bit is turned off should it be
820 * Note that if any of the passed file descriptors are > 2 they will be closed — both on success and on
821 * failure! Thus, callers should assume that when this function returns the input fds are invalidated.
823 * Note that when this function fails stdin/stdout/stderr might remain half set up!
825 * O_CLOEXEC is turned off for all three file descriptors (which is how it should be for
826 * stdin/stdout/stderr). */
828 null_readable = original_input_fd < 0;
829 null_writable = original_output_fd < 0 || original_error_fd < 0;
831 /* First step, open /dev/null once, if we need it */
832 if (null_readable || null_writable) {
834 /* Let's open this with O_CLOEXEC first, and convert it to non-O_CLOEXEC when we move the fd to the final position. */
835 null_fd = open("/dev/null", (null_readable && null_writable ? O_RDWR :
836 null_readable ? O_RDONLY : O_WRONLY) | O_CLOEXEC);
842 /* If this fd is in the 0…2 range, let's move it out of it */
846 copy = fcntl(null_fd, F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */
857 /* Let's assemble fd[] with the fds to install in place of stdin/stdout/stderr */
858 for (i = 0; i < 3; i++) {
861 fd[i] = null_fd; /* A negative parameter means: connect this one to /dev/null */
862 else if (fd[i] != i && fd[i] < 3) {
863 /* 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. */
864 copy_fd[i] = fcntl(fd[i], F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */
865 if (copy_fd[i] < 0) {
874 /* At this point we now have the fds to use in fd[], and they are all above the stdio range, so that we
875 * have freedom to move them around. If the fds already were at the right places then the specific fds are
876 * -1. Let's now move them to the right places. This is the point of no return. */
877 for (i = 0; i < 3; i++) {
881 /* fd is already in place, but let's make sure O_CLOEXEC is off */
882 r = fd_cloexec(i, false);
889 if (dup2(fd[i], i) < 0) { /* Turns off O_CLOEXEC on the new fd. */
899 /* Close the original fds, but only if they were outside of the stdio range. Also, properly check for the same
900 * fd passed in multiple times. */
901 safe_close_above_stdio(original_input_fd);
902 if (original_output_fd != original_input_fd)
903 safe_close_above_stdio(original_output_fd);
904 if (original_error_fd != original_input_fd && original_error_fd != original_output_fd)
905 safe_close_above_stdio(original_error_fd);
907 /* Close the copies we moved > 2 */
908 for (i = 0; i < 3; i++)
909 safe_close(copy_fd[i]);
911 /* Close our null fd, if it's > 2 */
912 safe_close_above_stdio(null_fd);
917 int fd_reopen(int fd, int flags) {
918 char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int)];
921 /* Reopens the specified fd with new flags. This is useful for convert an O_PATH fd into a regular one, or to
922 * turn O_RDWR fds into O_RDONLY fds.
924 * This doesn't work on sockets (since they cannot be open()ed, ever).
926 * This implicitly resets the file read index to 0. */
928 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
929 new_fd = open(procfs_path, flags);
936 int read_nr_open(void) {
937 _cleanup_free_ char *nr_open = NULL;
940 /* Returns the kernel's current fd limit, either by reading it of /proc/sys if that works, or using the
941 * hard-coded default compiled-in value of current kernels (1M) if not. This call will never fail. */
943 r = read_one_line_file("/proc/sys/fs/nr_open", &nr_open);
945 log_debug_errno(r, "Failed to read /proc/sys/fs/nr_open, ignoring: %m");
949 r = safe_atoi(nr_open, &v);
951 log_debug_errno(r, "Failed to parse /proc/sys/fs/nr_open value '%s', ignoring: %m", nr_open);
956 /* If we fail, fallback to the hard-coded kernel limit of 1024 * 1024. */