1 /* SPDX-License-Identifier: LGPL-2.1+ */
3 This file is part of systemd.
5 Copyright 2010 Lennart Poettering
10 #include <sys/resource.h>
11 #include <sys/socket.h>
15 //#include "alloc-util.h"
17 #include "dirent-util.h"
21 //#include "io-util.h"
23 #include "memfd-util.h"
25 #include "parse-util.h"
26 #include "path-util.h"
27 #include "process-util.h"
28 #include "socket-util.h"
29 #include "stdio-util.h"
32 int close_nointr(int fd) {
39 * Just ignore EINTR; a retry loop is the wrong thing to do on
42 * http://lkml.indiana.edu/hypermail/linux/kernel/0509.1/0877.html
43 * https://bugzilla.gnome.org/show_bug.cgi?id=682819
44 * http://utcc.utoronto.ca/~cks/space/blog/unix/CloseEINTR
45 * https://sites.google.com/site/michaelsafyan/software-engineering/checkforeintrwheninvokingclosethinkagain
53 int safe_close(int fd) {
56 * Like close_nointr() but cannot fail. Guarantees errno is
57 * unchanged. Is a NOP with negative fds passed, and returns
58 * -1, so that it can be used in this syntax:
60 * fd = safe_close(fd);
66 /* The kernel might return pretty much any error code
67 * via close(), but the fd will be closed anyway. The
68 * only condition we want to check for here is whether
69 * the fd was invalid at all... */
71 assert_se(close_nointr(fd) != -EBADF);
77 void safe_close_pair(int p[]) {
81 /* Special case pairs which use the same fd in both
83 p[0] = p[1] = safe_close(p[0]);
87 p[0] = safe_close(p[0]);
88 p[1] = safe_close(p[1]);
91 void close_many(const int fds[], size_t n_fd) {
94 assert(fds || n_fd <= 0);
96 for (i = 0; i < n_fd; i++)
100 int fclose_nointr(FILE *f) {
103 /* Same as close_nointr(), but for fclose() */
114 FILE* safe_fclose(FILE *f) {
116 /* Same as safe_close(), but for fclose() */
121 assert_se(fclose_nointr(f) != EBADF);
127 #if 0 /// UNNEEDED by elogind
128 DIR* safe_closedir(DIR *d) {
133 assert_se(closedir(d) >= 0 || errno != EBADF);
140 int fd_nonblock(int fd, bool nonblock) {
145 flags = fcntl(fd, F_GETFL, 0);
150 nflags = flags | O_NONBLOCK;
152 nflags = flags & ~O_NONBLOCK;
157 if (fcntl(fd, F_SETFL, nflags) < 0)
163 int fd_cloexec(int fd, bool cloexec) {
168 flags = fcntl(fd, F_GETFD, 0);
173 nflags = flags | FD_CLOEXEC;
175 nflags = flags & ~FD_CLOEXEC;
180 if (fcntl(fd, F_SETFD, nflags) < 0)
186 _pure_ static bool fd_in_set(int fd, const int fdset[], size_t n_fdset) {
189 assert(n_fdset == 0 || fdset);
191 for (i = 0; i < n_fdset; i++)
198 int close_all_fds(const int except[], size_t n_except) {
199 _cleanup_closedir_ DIR *d = NULL;
203 assert(n_except == 0 || except);
205 d = opendir("/proc/self/fd");
210 /* When /proc isn't available (for example in chroots) the fallback is brute forcing through the fd
213 assert_se(getrlimit(RLIMIT_NOFILE, &rl) >= 0);
215 if (rl.rlim_max == 0)
218 /* Let's take special care if the resource limit is set to unlimited, or actually larger than the range
219 * of 'int'. Let's avoid implicit overflows. */
220 max_fd = (rl.rlim_max == RLIM_INFINITY || rl.rlim_max > INT_MAX) ? INT_MAX : (int) (rl.rlim_max - 1);
222 for (fd = 3; fd >= 0; fd = fd < max_fd ? fd + 1 : -1) {
225 if (fd_in_set(fd, except, n_except))
228 q = close_nointr(fd);
229 if (q < 0 && q != -EBADF && r >= 0)
236 FOREACH_DIRENT(de, d, return -errno) {
239 if (safe_atoi(de->d_name, &fd) < 0)
240 /* Let's better ignore this, just in case */
249 if (fd_in_set(fd, except, n_except))
252 q = close_nointr(fd);
253 if (q < 0 && q != -EBADF && r >= 0) /* Valgrind has its own FD and doesn't want to have it closed */
260 #if 0 /// UNNEEDED by elogind
261 int same_fd(int a, int b) {
262 struct stat sta, stb;
269 /* Compares two file descriptors. Note that semantics are
270 * quite different depending on whether we have kcmp() or we
271 * don't. If we have kcmp() this will only return true for
272 * dup()ed file descriptors, but not otherwise. If we don't
273 * have kcmp() this will also return true for two fds of the same
274 * file, created by separate open() calls. Since we use this
275 * call mostly for filtering out duplicates in the fd store
276 * this difference hopefully doesn't matter too much. */
281 /* Try to use kcmp() if we have it. */
282 pid = getpid_cached();
283 r = kcmp(pid, pid, KCMP_FILE, a, b);
291 /* We don't have kcmp(), use fstat() instead. */
292 if (fstat(a, &sta) < 0)
295 if (fstat(b, &stb) < 0)
298 if ((sta.st_mode & S_IFMT) != (stb.st_mode & S_IFMT))
301 /* We consider all device fds different, since two device fds
302 * might refer to quite different device contexts even though
303 * they share the same inode and backing dev_t. */
305 if (S_ISCHR(sta.st_mode) || S_ISBLK(sta.st_mode))
308 if (sta.st_dev != stb.st_dev || sta.st_ino != stb.st_ino)
311 /* The fds refer to the same inode on disk, let's also check
312 * if they have the same fd flags. This is useful to
313 * distinguish the read and write side of a pipe created with
315 fa = fcntl(a, F_GETFL);
319 fb = fcntl(b, F_GETFL);
326 void cmsg_close_all(struct msghdr *mh) {
327 struct cmsghdr *cmsg;
331 CMSG_FOREACH(cmsg, mh)
332 if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS)
333 close_many((int*) CMSG_DATA(cmsg), (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int));
336 bool fdname_is_valid(const char *s) {
339 /* Validates a name for $LISTEN_FDNAMES. We basically allow
340 * everything ASCII that's not a control character. Also, as
341 * special exception the ":" character is not allowed, as we
342 * use that as field separator in $LISTEN_FDNAMES.
344 * Note that the empty string is explicitly allowed
345 * here. However, we limit the length of the names to 255
351 for (p = s; *p; p++) {
364 int fd_get_path(int fd, char **ret) {
365 _cleanup_close_ int dir = -1;
366 char fdname[DECIMAL_STR_MAX(int)];
369 dir = open("/proc/self/fd/", O_CLOEXEC | O_DIRECTORY | O_PATH);
371 /* /proc is not available or not set up properly, we're most likely
372 * in some chroot environment. */
373 return errno == ENOENT ? -EOPNOTSUPP : -errno;
375 xsprintf(fdname, "%i", fd);
377 r = readlinkat_malloc(dir, fdname, ret);
379 /* If the file doesn't exist the fd is invalid */
385 int move_fd(int from, int to, int cloexec) {
388 /* Move fd 'from' to 'to', make sure FD_CLOEXEC remains equal if requested, and release the old fd. If
389 * 'cloexec' is passed as -1, the original FD_CLOEXEC is inherited for the new fd. If it is 0, it is turned
390 * off, if it is > 0 it is turned on. */
400 r = fd_cloexec(to, cloexec);
411 fl = fcntl(from, F_GETFD, 0);
415 cloexec = !!(fl & FD_CLOEXEC);
418 r = dup3(from, to, cloexec ? O_CLOEXEC : 0);
429 int acquire_data_fd(const void *data, size_t size, unsigned flags) {
431 _cleanup_close_pair_ int pipefds[2] = { -1, -1 };
432 char pattern[] = "/dev/shm/data-fd-XXXXXX";
433 _cleanup_close_ int fd = -1;
438 assert(data || size == 0);
440 /* Acquire a read-only file descriptor that when read from returns the specified data. This is much more
441 * complex than I wish it was. But here's why:
443 * a) First we try to use memfds. They are the best option, as we can seal them nicely to make them
444 * read-only. Unfortunately they require kernel 3.17, and – at the time of writing – we still support 3.14.
446 * b) Then, we try classic pipes. They are the second best options, as we can close the writing side, retaining
447 * a nicely read-only fd in the reading side. However, they are by default quite small, and unprivileged
448 * clients can only bump their size to a system-wide limit, which might be quite low.
450 * c) Then, we try an O_TMPFILE file in /dev/shm (that dir is the only suitable one known to exist from
451 * earliest boot on). To make it read-only we open the fd a second time with O_RDONLY via
452 * /proc/self/<fd>. Unfortunately O_TMPFILE is not available on older kernels on tmpfs.
454 * d) Finally, we try creating a regular file in /dev/shm, which we then delete.
456 * It sucks a bit that depending on the situation we return very different objects here, but that's Linux I
459 if (size == 0 && ((flags & ACQUIRE_NO_DEV_NULL) == 0)) {
460 /* As a special case, return /dev/null if we have been called for an empty data block */
461 r = open("/dev/null", O_RDONLY|O_CLOEXEC|O_NOCTTY);
468 if ((flags & ACQUIRE_NO_MEMFD) == 0) {
469 fd = memfd_new("data-fd");
473 n = write(fd, data, size);
476 if ((size_t) n != size)
479 f = lseek(fd, 0, SEEK_SET);
483 r = memfd_set_sealed(fd);
491 if ((flags & ACQUIRE_NO_PIPE) == 0) {
492 if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0)
495 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
499 if ((size_t) isz < size) {
501 if (isz < 0 || (size_t) isz != size)
504 /* Try to bump the pipe size */
505 (void) fcntl(pipefds[1], F_SETPIPE_SZ, isz);
507 /* See if that worked */
508 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
512 if ((size_t) isz < size)
516 n = write(pipefds[1], data, size);
519 if ((size_t) n != size)
522 (void) fd_nonblock(pipefds[0], false);
524 return TAKE_FD(pipefds[0]);
528 if ((flags & ACQUIRE_NO_TMPFILE) == 0) {
529 fd = open("/dev/shm", O_RDWR|O_TMPFILE|O_CLOEXEC, 0500);
531 goto try_dev_shm_without_o_tmpfile;
533 n = write(fd, data, size);
536 if ((size_t) n != size)
539 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
540 return fd_reopen(fd, O_RDONLY|O_CLOEXEC);
543 try_dev_shm_without_o_tmpfile:
544 if ((flags & ACQUIRE_NO_REGULAR) == 0) {
545 fd = mkostemp_safe(pattern);
549 n = write(fd, data, size);
552 goto unlink_and_return;
554 if ((size_t) n != size) {
556 goto unlink_and_return;
559 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
560 r = open(pattern, O_RDONLY|O_CLOEXEC);
565 (void) unlink(pattern);
572 /* When the data is smaller or equal to 64K, try to place the copy in a memfd/pipe */
573 #define DATA_FD_MEMORY_LIMIT (64U*1024U)
575 /* 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. */
576 #define DATA_FD_TMP_LIMIT (1024U*1024U)
578 int fd_duplicate_data_fd(int fd) {
580 _cleanup_close_ int copy_fd = -1, tmp_fd = -1;
581 _cleanup_free_ void *remains = NULL;
582 size_t remains_size = 0;
587 /* Creates a 'data' fd from the specified source fd, containing all the same data in a read-only fashion, but
588 * independent of it (i.e. the source fd can be closed and unmounted after this call succeeded). Tries to be
589 * somewhat smart about where to place the data. In the best case uses a memfd(). If memfd() are not supported
590 * uses a pipe instead. For larger data will use an unlinked file in /tmp, and for even larger data one in
593 if (fstat(fd, &st) < 0)
596 /* For now, let's only accept regular files, sockets, pipes and char devices */
597 if (S_ISDIR(st.st_mode))
599 if (S_ISLNK(st.st_mode))
601 if (!S_ISREG(st.st_mode) && !S_ISSOCK(st.st_mode) && !S_ISFIFO(st.st_mode) && !S_ISCHR(st.st_mode))
604 /* If we have reason to believe the data is bounded in size, then let's use memfds or pipes as backing fd. Note
605 * that we use the reported regular file size only as a hint, given that there are plenty special files in
606 * /proc and /sys which report a zero file size but can be read from. */
608 if (!S_ISREG(st.st_mode) || st.st_size < DATA_FD_MEMORY_LIMIT) {
610 /* Try a memfd first */
611 copy_fd = memfd_new("data-fd");
615 r = copy_bytes(fd, copy_fd, DATA_FD_MEMORY_LIMIT, 0);
619 f = lseek(copy_fd, 0, SEEK_SET);
624 /* Did it fit into the limit? If so, we are done. */
625 r = memfd_set_sealed(copy_fd);
629 return TAKE_FD(copy_fd);
632 /* Hmm, pity, this didn't fit. Let's fall back to /tmp then, see below */
635 _cleanup_(close_pairp) int pipefds[2] = { -1, -1 };
638 /* If memfds aren't available, use a pipe. Set O_NONBLOCK so that we will get EAGAIN rather
639 * then block indefinitely when we hit the pipe size limit */
641 if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0)
644 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
648 /* Try to enlarge the pipe size if necessary */
649 if ((size_t) isz < DATA_FD_MEMORY_LIMIT) {
651 (void) fcntl(pipefds[1], F_SETPIPE_SZ, DATA_FD_MEMORY_LIMIT);
653 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
658 if ((size_t) isz >= DATA_FD_MEMORY_LIMIT) {
660 r = copy_bytes_full(fd, pipefds[1], DATA_FD_MEMORY_LIMIT, 0, &remains, &remains_size);
661 if (r < 0 && r != -EAGAIN)
662 return r; /* If we get EAGAIN it could be because of the source or because of
663 * the destination fd, we can't know, as sendfile() and friends won't
664 * tell us. Hence, treat this as reason to fall back, just to be
667 /* Everything fit in, yay! */
668 (void) fd_nonblock(pipefds[0], false);
670 return TAKE_FD(pipefds[0]);
673 /* Things didn't fit in. But we read data into the pipe, let's remember that, so that
674 * when writing the new file we incorporate this first. */
675 copy_fd = TAKE_FD(pipefds[0]);
680 /* If we have reason to believe this will fit fine in /tmp, then use that as first fallback. */
681 if ((!S_ISREG(st.st_mode) || st.st_size < DATA_FD_TMP_LIMIT) &&
682 (DATA_FD_MEMORY_LIMIT + remains_size) < DATA_FD_TMP_LIMIT) {
685 tmp_fd = open_tmpfile_unlinkable(NULL /* NULL as directory means /tmp */, O_RDWR|O_CLOEXEC);
690 /* If we tried a memfd/pipe first and it ended up being too large, then copy this into the
691 * temporary file first. */
693 r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, 0);
700 if (remains_size > 0) {
701 /* If there were remaining bytes (i.e. read into memory, but not written out yet) from the
702 * failed copy operation, let's flush them out next. */
704 r = loop_write(tmp_fd, remains, remains_size, false);
709 r = copy_bytes(fd, tmp_fd, DATA_FD_TMP_LIMIT - DATA_FD_MEMORY_LIMIT - remains_size, COPY_REFLINK);
713 goto finish; /* Yay, it fit in */
715 /* It didn't fit in. Let's not forget to use what we already used */
716 f = lseek(tmp_fd, 0, SEEK_SET);
721 copy_fd = TAKE_FD(tmp_fd);
723 remains = mfree(remains);
727 /* As last fallback use /var/tmp */
728 r = var_tmp_dir(&td);
732 tmp_fd = open_tmpfile_unlinkable(td, O_RDWR|O_CLOEXEC);
737 /* If we tried a memfd/pipe first, or a file in /tmp, and it ended up being too large, than copy this
738 * into the temporary file first. */
739 r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
746 if (remains_size > 0) {
747 /* Then, copy in any read but not yet written bytes. */
748 r = loop_write(tmp_fd, remains, remains_size, false);
753 /* Copy in the rest */
754 r = copy_bytes(fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
761 /* Now convert the O_RDWR file descriptor into an O_RDONLY one (and as side effect seek to the beginning of the
764 return fd_reopen(tmp_fd, O_RDONLY|O_CLOEXEC);
767 int fd_move_above_stdio(int fd) {
771 /* Moves the specified file descriptor if possible out of the range [0…2], i.e. the range of
772 * stdin/stdout/stderr. If it can't be moved outside of this range the original file descriptor is
773 * returned. This call is supposed to be used for long-lasting file descriptors we allocate in our code that
774 * might get loaded into foreign code, and where we want ensure our fds are unlikely used accidentally as
775 * stdin/stdout/stderr of unrelated code.
777 * Note that this doesn't fix any real bugs, it just makes it less likely that our code will be affected by
778 * buggy code from others that mindlessly invokes 'fprintf(stderr, …' or similar in places where stderr has
779 * been closed before.
781 * This function is written in a "best-effort" and "least-impact" style. This means whenever we encounter an
782 * error we simply return the original file descriptor, and we do not touch errno. */
784 if (fd < 0 || fd > 2)
787 flags = fcntl(fd, F_GETFD, 0);
791 if (flags & FD_CLOEXEC)
792 copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
794 copy = fcntl(fd, F_DUPFD, 3);
804 int rearrange_stdio(int original_input_fd, int original_output_fd, int original_error_fd) {
806 int fd[3] = { /* Put together an array of fds we work on */
813 null_fd = -1, /* if we open /dev/null, we store the fd to it here */
814 copy_fd[3] = { -1, -1, -1 }; /* This contains all fds we duplicate here temporarily, and hence need to close at the end */
815 bool null_readable, null_writable;
817 /* Sets up stdin, stdout, stderr with the three file descriptors passed in. If any of the descriptors is
818 * specified as -1 it will be connected with /dev/null instead. If any of the file descriptors is passed as
819 * itself (e.g. stdin as STDIN_FILENO) it is left unmodified, but the O_CLOEXEC bit is turned off should it be
822 * Note that if any of the passed file descriptors are > 2 they will be closed — both on success and on
823 * failure! Thus, callers should assume that when this function returns the input fds are invalidated.
825 * Note that when this function fails stdin/stdout/stderr might remain half set up!
827 * O_CLOEXEC is turned off for all three file descriptors (which is how it should be for
828 * stdin/stdout/stderr). */
830 null_readable = original_input_fd < 0;
831 null_writable = original_output_fd < 0 || original_error_fd < 0;
833 /* First step, open /dev/null once, if we need it */
834 if (null_readable || null_writable) {
836 /* Let's open this with O_CLOEXEC first, and convert it to non-O_CLOEXEC when we move the fd to the final position. */
837 null_fd = open("/dev/null", (null_readable && null_writable ? O_RDWR :
838 null_readable ? O_RDONLY : O_WRONLY) | O_CLOEXEC);
844 /* If this fd is in the 0…2 range, let's move it out of it */
848 copy = fcntl(null_fd, F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */
859 /* Let's assemble fd[] with the fds to install in place of stdin/stdout/stderr */
860 for (i = 0; i < 3; i++) {
863 fd[i] = null_fd; /* A negative parameter means: connect this one to /dev/null */
864 else if (fd[i] != i && fd[i] < 3) {
865 /* 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. */
866 copy_fd[i] = fcntl(fd[i], F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */
867 if (copy_fd[i] < 0) {
876 /* At this point we now have the fds to use in fd[], and they are all above the stdio range, so that we
877 * have freedom to move them around. If the fds already were at the right places then the specific fds are
878 * -1. Let's now move them to the right places. This is the point of no return. */
879 for (i = 0; i < 3; i++) {
883 /* fd is already in place, but let's make sure O_CLOEXEC is off */
884 r = fd_cloexec(i, false);
891 if (dup2(fd[i], i) < 0) { /* Turns off O_CLOEXEC on the new fd. */
901 /* Close the original fds, but only if they were outside of the stdio range. Also, properly check for the same
902 * fd passed in multiple times. */
903 safe_close_above_stdio(original_input_fd);
904 if (original_output_fd != original_input_fd)
905 safe_close_above_stdio(original_output_fd);
906 if (original_error_fd != original_input_fd && original_error_fd != original_output_fd)
907 safe_close_above_stdio(original_error_fd);
909 /* Close the copies we moved > 2 */
910 for (i = 0; i < 3; i++)
911 safe_close(copy_fd[i]);
913 /* Close our null fd, if it's > 2 */
914 safe_close_above_stdio(null_fd);
919 int fd_reopen(int fd, int flags) {
920 char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int)];
923 /* Reopens the specified fd with new flags. This is useful for convert an O_PATH fd into a regular one, or to
924 * turn O_RDWR fds into O_RDONLY fds.
926 * This doesn't work on sockets (since they cannot be open()ed, ever).
928 * This implicitly resets the file read index to 0. */
930 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
931 new_fd = open(procfs_path, flags);
938 int read_nr_open(void) {
939 _cleanup_free_ char *nr_open = NULL;
942 /* Returns the kernel's current fd limit, either by reading it of /proc/sys if that works, or using the
943 * hard-coded default compiled-in value of current kernels (1M) if not. This call will never fail. */
945 r = read_one_line_file("/proc/sys/fs/nr_open", &nr_open);
947 log_debug_errno(r, "Failed to read /proc/sys/fs/nr_open, ignoring: %m");
951 r = safe_atoi(nr_open, &v);
953 log_debug_errno(r, "Failed to parse /proc/sys/fs/nr_open value '%s', ignoring: %m", nr_open);
958 /* If we fail, fallback to the hard-coded kernel limit of 1024 * 1024. */