1 /* SPDX-License-Identifier: LGPL-2.1+ */
5 #include <sys/resource.h>
6 #include <sys/socket.h>
10 //#include "alloc-util.h"
12 #include "dirent-util.h"
16 //#include "io-util.h"
18 #include "memfd-util.h"
20 #include "parse-util.h"
21 #include "path-util.h"
22 #include "process-util.h"
23 #include "socket-util.h"
24 #include "stdio-util.h"
27 int close_nointr(int fd) {
34 * Just ignore EINTR; a retry loop is the wrong thing to do on
37 * http://lkml.indiana.edu/hypermail/linux/kernel/0509.1/0877.html
38 * https://bugzilla.gnome.org/show_bug.cgi?id=682819
39 * http://utcc.utoronto.ca/~cks/space/blog/unix/CloseEINTR
40 * https://sites.google.com/site/michaelsafyan/software-engineering/checkforeintrwheninvokingclosethinkagain
48 int safe_close(int fd) {
51 * Like close_nointr() but cannot fail. Guarantees errno is
52 * unchanged. Is a NOP with negative fds passed, and returns
53 * -1, so that it can be used in this syntax:
55 * fd = safe_close(fd);
61 /* The kernel might return pretty much any error code
62 * via close(), but the fd will be closed anyway. The
63 * only condition we want to check for here is whether
64 * the fd was invalid at all... */
66 assert_se(close_nointr(fd) != -EBADF);
72 void safe_close_pair(int p[]) {
76 /* Special case pairs which use the same fd in both
78 p[0] = p[1] = safe_close(p[0]);
82 p[0] = safe_close(p[0]);
83 p[1] = safe_close(p[1]);
86 void close_many(const int fds[], size_t n_fd) {
89 assert(fds || n_fd <= 0);
91 for (i = 0; i < n_fd; i++)
95 int fclose_nointr(FILE *f) {
98 /* Same as close_nointr(), but for fclose() */
109 FILE* safe_fclose(FILE *f) {
111 /* Same as safe_close(), but for fclose() */
116 assert_se(fclose_nointr(f) != EBADF);
122 #if 0 /// UNNEEDED by elogind
123 DIR* safe_closedir(DIR *d) {
128 assert_se(closedir(d) >= 0 || errno != EBADF);
135 int fd_nonblock(int fd, bool nonblock) {
140 flags = fcntl(fd, F_GETFL, 0);
145 nflags = flags | O_NONBLOCK;
147 nflags = flags & ~O_NONBLOCK;
152 if (fcntl(fd, F_SETFL, nflags) < 0)
158 int fd_cloexec(int fd, bool cloexec) {
163 flags = fcntl(fd, F_GETFD, 0);
168 nflags = flags | FD_CLOEXEC;
170 nflags = flags & ~FD_CLOEXEC;
175 if (fcntl(fd, F_SETFD, nflags) < 0)
181 _pure_ static bool fd_in_set(int fd, const int fdset[], size_t n_fdset) {
184 assert(n_fdset == 0 || fdset);
186 for (i = 0; i < n_fdset; i++)
193 int close_all_fds(const int except[], size_t n_except) {
194 _cleanup_closedir_ DIR *d = NULL;
198 assert(n_except == 0 || except);
200 d = opendir("/proc/self/fd");
205 /* When /proc isn't available (for example in chroots) the fallback is brute forcing through the fd
208 assert_se(getrlimit(RLIMIT_NOFILE, &rl) >= 0);
210 if (rl.rlim_max == 0)
213 /* Let's take special care if the resource limit is set to unlimited, or actually larger than the range
214 * of 'int'. Let's avoid implicit overflows. */
215 max_fd = (rl.rlim_max == RLIM_INFINITY || rl.rlim_max > INT_MAX) ? INT_MAX : (int) (rl.rlim_max - 1);
217 for (fd = 3; fd >= 0; fd = fd < max_fd ? fd + 1 : -1) {
220 if (fd_in_set(fd, except, n_except))
223 q = close_nointr(fd);
224 if (q < 0 && q != -EBADF && r >= 0)
231 FOREACH_DIRENT(de, d, return -errno) {
234 if (safe_atoi(de->d_name, &fd) < 0)
235 /* Let's better ignore this, just in case */
244 if (fd_in_set(fd, except, n_except))
247 q = close_nointr(fd);
248 if (q < 0 && q != -EBADF && r >= 0) /* Valgrind has its own FD and doesn't want to have it closed */
255 #if 0 /// UNNEEDED by elogind
256 int same_fd(int a, int b) {
257 struct stat sta, stb;
264 /* Compares two file descriptors. Note that semantics are
265 * quite different depending on whether we have kcmp() or we
266 * don't. If we have kcmp() this will only return true for
267 * dup()ed file descriptors, but not otherwise. If we don't
268 * have kcmp() this will also return true for two fds of the same
269 * file, created by separate open() calls. Since we use this
270 * call mostly for filtering out duplicates in the fd store
271 * this difference hopefully doesn't matter too much. */
276 /* Try to use kcmp() if we have it. */
277 pid = getpid_cached();
278 r = kcmp(pid, pid, KCMP_FILE, a, b);
286 /* We don't have kcmp(), use fstat() instead. */
287 if (fstat(a, &sta) < 0)
290 if (fstat(b, &stb) < 0)
293 if ((sta.st_mode & S_IFMT) != (stb.st_mode & S_IFMT))
296 /* We consider all device fds different, since two device fds
297 * might refer to quite different device contexts even though
298 * they share the same inode and backing dev_t. */
300 if (S_ISCHR(sta.st_mode) || S_ISBLK(sta.st_mode))
303 if (sta.st_dev != stb.st_dev || sta.st_ino != stb.st_ino)
306 /* The fds refer to the same inode on disk, let's also check
307 * if they have the same fd flags. This is useful to
308 * distinguish the read and write side of a pipe created with
310 fa = fcntl(a, F_GETFL);
314 fb = fcntl(b, F_GETFL);
321 void cmsg_close_all(struct msghdr *mh) {
322 struct cmsghdr *cmsg;
326 CMSG_FOREACH(cmsg, mh)
327 if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS)
328 close_many((int*) CMSG_DATA(cmsg), (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int));
331 bool fdname_is_valid(const char *s) {
334 /* Validates a name for $LISTEN_FDNAMES. We basically allow
335 * everything ASCII that's not a control character. Also, as
336 * special exception the ":" character is not allowed, as we
337 * use that as field separator in $LISTEN_FDNAMES.
339 * Note that the empty string is explicitly allowed
340 * here. However, we limit the length of the names to 255
346 for (p = s; *p; p++) {
359 int fd_get_path(int fd, char **ret) {
360 _cleanup_close_ int dir = -1;
361 char fdname[DECIMAL_STR_MAX(int)];
364 dir = open("/proc/self/fd/", O_CLOEXEC | O_DIRECTORY | O_PATH);
366 /* /proc is not available or not set up properly, we're most likely
367 * in some chroot environment. */
368 return errno == ENOENT ? -EOPNOTSUPP : -errno;
370 xsprintf(fdname, "%i", fd);
372 r = readlinkat_malloc(dir, fdname, ret);
374 /* If the file doesn't exist the fd is invalid */
380 int move_fd(int from, int to, int cloexec) {
383 /* Move fd 'from' to 'to', make sure FD_CLOEXEC remains equal if requested, and release the old fd. If
384 * 'cloexec' is passed as -1, the original FD_CLOEXEC is inherited for the new fd. If it is 0, it is turned
385 * off, if it is > 0 it is turned on. */
395 r = fd_cloexec(to, cloexec);
406 fl = fcntl(from, F_GETFD, 0);
410 cloexec = !!(fl & FD_CLOEXEC);
413 r = dup3(from, to, cloexec ? O_CLOEXEC : 0);
424 int acquire_data_fd(const void *data, size_t size, unsigned flags) {
426 _cleanup_close_pair_ int pipefds[2] = { -1, -1 };
427 char pattern[] = "/dev/shm/data-fd-XXXXXX";
428 _cleanup_close_ int fd = -1;
433 assert(data || size == 0);
435 /* Acquire a read-only file descriptor that when read from returns the specified data. This is much more
436 * complex than I wish it was. But here's why:
438 * a) First we try to use memfds. They are the best option, as we can seal them nicely to make them
439 * read-only. Unfortunately they require kernel 3.17, and – at the time of writing – we still support 3.14.
441 * b) Then, we try classic pipes. They are the second best options, as we can close the writing side, retaining
442 * a nicely read-only fd in the reading side. However, they are by default quite small, and unprivileged
443 * clients can only bump their size to a system-wide limit, which might be quite low.
445 * c) Then, we try an O_TMPFILE file in /dev/shm (that dir is the only suitable one known to exist from
446 * earliest boot on). To make it read-only we open the fd a second time with O_RDONLY via
447 * /proc/self/<fd>. Unfortunately O_TMPFILE is not available on older kernels on tmpfs.
449 * d) Finally, we try creating a regular file in /dev/shm, which we then delete.
451 * It sucks a bit that depending on the situation we return very different objects here, but that's Linux I
454 if (size == 0 && ((flags & ACQUIRE_NO_DEV_NULL) == 0)) {
455 /* As a special case, return /dev/null if we have been called for an empty data block */
456 r = open("/dev/null", O_RDONLY|O_CLOEXEC|O_NOCTTY);
463 if ((flags & ACQUIRE_NO_MEMFD) == 0) {
464 fd = memfd_new("data-fd");
468 n = write(fd, data, size);
471 if ((size_t) n != size)
474 f = lseek(fd, 0, SEEK_SET);
478 r = memfd_set_sealed(fd);
486 if ((flags & ACQUIRE_NO_PIPE) == 0) {
487 if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0)
490 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
494 if ((size_t) isz < size) {
496 if (isz < 0 || (size_t) isz != size)
499 /* Try to bump the pipe size */
500 (void) fcntl(pipefds[1], F_SETPIPE_SZ, isz);
502 /* See if that worked */
503 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
507 if ((size_t) isz < size)
511 n = write(pipefds[1], data, size);
514 if ((size_t) n != size)
517 (void) fd_nonblock(pipefds[0], false);
519 return TAKE_FD(pipefds[0]);
523 if ((flags & ACQUIRE_NO_TMPFILE) == 0) {
524 fd = open("/dev/shm", O_RDWR|O_TMPFILE|O_CLOEXEC, 0500);
526 goto try_dev_shm_without_o_tmpfile;
528 n = write(fd, data, size);
531 if ((size_t) n != size)
534 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
535 return fd_reopen(fd, O_RDONLY|O_CLOEXEC);
538 try_dev_shm_without_o_tmpfile:
539 if ((flags & ACQUIRE_NO_REGULAR) == 0) {
540 fd = mkostemp_safe(pattern);
544 n = write(fd, data, size);
547 goto unlink_and_return;
549 if ((size_t) n != size) {
551 goto unlink_and_return;
554 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
555 r = open(pattern, O_RDONLY|O_CLOEXEC);
560 (void) unlink(pattern);
567 /* When the data is smaller or equal to 64K, try to place the copy in a memfd/pipe */
568 #define DATA_FD_MEMORY_LIMIT (64U*1024U)
570 /* 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. */
571 #define DATA_FD_TMP_LIMIT (1024U*1024U)
573 int fd_duplicate_data_fd(int fd) {
575 _cleanup_close_ int copy_fd = -1, tmp_fd = -1;
576 _cleanup_free_ void *remains = NULL;
577 size_t remains_size = 0;
582 /* Creates a 'data' fd from the specified source fd, containing all the same data in a read-only fashion, but
583 * independent of it (i.e. the source fd can be closed and unmounted after this call succeeded). Tries to be
584 * somewhat smart about where to place the data. In the best case uses a memfd(). If memfd() are not supported
585 * uses a pipe instead. For larger data will use an unlinked file in /tmp, and for even larger data one in
588 if (fstat(fd, &st) < 0)
591 /* For now, let's only accept regular files, sockets, pipes and char devices */
592 if (S_ISDIR(st.st_mode))
594 if (S_ISLNK(st.st_mode))
596 if (!S_ISREG(st.st_mode) && !S_ISSOCK(st.st_mode) && !S_ISFIFO(st.st_mode) && !S_ISCHR(st.st_mode))
599 /* If we have reason to believe the data is bounded in size, then let's use memfds or pipes as backing fd. Note
600 * that we use the reported regular file size only as a hint, given that there are plenty special files in
601 * /proc and /sys which report a zero file size but can be read from. */
603 if (!S_ISREG(st.st_mode) || st.st_size < DATA_FD_MEMORY_LIMIT) {
605 /* Try a memfd first */
606 copy_fd = memfd_new("data-fd");
610 r = copy_bytes(fd, copy_fd, DATA_FD_MEMORY_LIMIT, 0);
614 f = lseek(copy_fd, 0, SEEK_SET);
619 /* Did it fit into the limit? If so, we are done. */
620 r = memfd_set_sealed(copy_fd);
624 return TAKE_FD(copy_fd);
627 /* Hmm, pity, this didn't fit. Let's fall back to /tmp then, see below */
630 _cleanup_(close_pairp) int pipefds[2] = { -1, -1 };
633 /* If memfds aren't available, use a pipe. Set O_NONBLOCK so that we will get EAGAIN rather
634 * then block indefinitely when we hit the pipe size limit */
636 if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0)
639 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
643 /* Try to enlarge the pipe size if necessary */
644 if ((size_t) isz < DATA_FD_MEMORY_LIMIT) {
646 (void) fcntl(pipefds[1], F_SETPIPE_SZ, DATA_FD_MEMORY_LIMIT);
648 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
653 if ((size_t) isz >= DATA_FD_MEMORY_LIMIT) {
655 r = copy_bytes_full(fd, pipefds[1], DATA_FD_MEMORY_LIMIT, 0, &remains, &remains_size);
656 if (r < 0 && r != -EAGAIN)
657 return r; /* If we get EAGAIN it could be because of the source or because of
658 * the destination fd, we can't know, as sendfile() and friends won't
659 * tell us. Hence, treat this as reason to fall back, just to be
662 /* Everything fit in, yay! */
663 (void) fd_nonblock(pipefds[0], false);
665 return TAKE_FD(pipefds[0]);
668 /* Things didn't fit in. But we read data into the pipe, let's remember that, so that
669 * when writing the new file we incorporate this first. */
670 copy_fd = TAKE_FD(pipefds[0]);
675 /* If we have reason to believe this will fit fine in /tmp, then use that as first fallback. */
676 if ((!S_ISREG(st.st_mode) || st.st_size < DATA_FD_TMP_LIMIT) &&
677 (DATA_FD_MEMORY_LIMIT + remains_size) < DATA_FD_TMP_LIMIT) {
680 tmp_fd = open_tmpfile_unlinkable(NULL /* NULL as directory means /tmp */, O_RDWR|O_CLOEXEC);
685 /* If we tried a memfd/pipe first and it ended up being too large, then copy this into the
686 * temporary file first. */
688 r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, 0);
695 if (remains_size > 0) {
696 /* If there were remaining bytes (i.e. read into memory, but not written out yet) from the
697 * failed copy operation, let's flush them out next. */
699 r = loop_write(tmp_fd, remains, remains_size, false);
704 r = copy_bytes(fd, tmp_fd, DATA_FD_TMP_LIMIT - DATA_FD_MEMORY_LIMIT - remains_size, COPY_REFLINK);
708 goto finish; /* Yay, it fit in */
710 /* It didn't fit in. Let's not forget to use what we already used */
711 f = lseek(tmp_fd, 0, SEEK_SET);
716 copy_fd = TAKE_FD(tmp_fd);
718 remains = mfree(remains);
722 /* As last fallback use /var/tmp */
723 r = var_tmp_dir(&td);
727 tmp_fd = open_tmpfile_unlinkable(td, O_RDWR|O_CLOEXEC);
732 /* If we tried a memfd/pipe first, or a file in /tmp, and it ended up being too large, than copy this
733 * into the temporary file first. */
734 r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
741 if (remains_size > 0) {
742 /* Then, copy in any read but not yet written bytes. */
743 r = loop_write(tmp_fd, remains, remains_size, false);
748 /* Copy in the rest */
749 r = copy_bytes(fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
756 /* Now convert the O_RDWR file descriptor into an O_RDONLY one (and as side effect seek to the beginning of the
759 return fd_reopen(tmp_fd, O_RDONLY|O_CLOEXEC);
762 int fd_move_above_stdio(int fd) {
766 /* Moves the specified file descriptor if possible out of the range [0…2], i.e. the range of
767 * stdin/stdout/stderr. If it can't be moved outside of this range the original file descriptor is
768 * returned. This call is supposed to be used for long-lasting file descriptors we allocate in our code that
769 * might get loaded into foreign code, and where we want ensure our fds are unlikely used accidentally as
770 * stdin/stdout/stderr of unrelated code.
772 * Note that this doesn't fix any real bugs, it just makes it less likely that our code will be affected by
773 * buggy code from others that mindlessly invokes 'fprintf(stderr, …' or similar in places where stderr has
774 * been closed before.
776 * This function is written in a "best-effort" and "least-impact" style. This means whenever we encounter an
777 * error we simply return the original file descriptor, and we do not touch errno. */
779 if (fd < 0 || fd > 2)
782 flags = fcntl(fd, F_GETFD, 0);
786 if (flags & FD_CLOEXEC)
787 copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
789 copy = fcntl(fd, F_DUPFD, 3);
799 int rearrange_stdio(int original_input_fd, int original_output_fd, int original_error_fd) {
801 int fd[3] = { /* Put together an array of fds we work on */
808 null_fd = -1, /* if we open /dev/null, we store the fd to it here */
809 copy_fd[3] = { -1, -1, -1 }; /* This contains all fds we duplicate here temporarily, and hence need to close at the end */
810 bool null_readable, null_writable;
812 /* Sets up stdin, stdout, stderr with the three file descriptors passed in. If any of the descriptors is
813 * specified as -1 it will be connected with /dev/null instead. If any of the file descriptors is passed as
814 * itself (e.g. stdin as STDIN_FILENO) it is left unmodified, but the O_CLOEXEC bit is turned off should it be
817 * Note that if any of the passed file descriptors are > 2 they will be closed — both on success and on
818 * failure! Thus, callers should assume that when this function returns the input fds are invalidated.
820 * Note that when this function fails stdin/stdout/stderr might remain half set up!
822 * O_CLOEXEC is turned off for all three file descriptors (which is how it should be for
823 * stdin/stdout/stderr). */
825 null_readable = original_input_fd < 0;
826 null_writable = original_output_fd < 0 || original_error_fd < 0;
828 /* First step, open /dev/null once, if we need it */
829 if (null_readable || null_writable) {
831 /* Let's open this with O_CLOEXEC first, and convert it to non-O_CLOEXEC when we move the fd to the final position. */
832 null_fd = open("/dev/null", (null_readable && null_writable ? O_RDWR :
833 null_readable ? O_RDONLY : O_WRONLY) | O_CLOEXEC);
839 /* If this fd is in the 0…2 range, let's move it out of it */
843 copy = fcntl(null_fd, F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */
854 /* Let's assemble fd[] with the fds to install in place of stdin/stdout/stderr */
855 for (i = 0; i < 3; i++) {
858 fd[i] = null_fd; /* A negative parameter means: connect this one to /dev/null */
859 else if (fd[i] != i && fd[i] < 3) {
860 /* 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. */
861 copy_fd[i] = fcntl(fd[i], F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */
862 if (copy_fd[i] < 0) {
871 /* At this point we now have the fds to use in fd[], and they are all above the stdio range, so that we
872 * have freedom to move them around. If the fds already were at the right places then the specific fds are
873 * -1. Let's now move them to the right places. This is the point of no return. */
874 for (i = 0; i < 3; i++) {
878 /* fd is already in place, but let's make sure O_CLOEXEC is off */
879 r = fd_cloexec(i, false);
886 if (dup2(fd[i], i) < 0) { /* Turns off O_CLOEXEC on the new fd. */
896 /* Close the original fds, but only if they were outside of the stdio range. Also, properly check for the same
897 * fd passed in multiple times. */
898 safe_close_above_stdio(original_input_fd);
899 if (original_output_fd != original_input_fd)
900 safe_close_above_stdio(original_output_fd);
901 if (original_error_fd != original_input_fd && original_error_fd != original_output_fd)
902 safe_close_above_stdio(original_error_fd);
904 /* Close the copies we moved > 2 */
905 for (i = 0; i < 3; i++)
906 safe_close(copy_fd[i]);
908 /* Close our null fd, if it's > 2 */
909 safe_close_above_stdio(null_fd);
914 int fd_reopen(int fd, int flags) {
915 char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int)];
918 /* Reopens the specified fd with new flags. This is useful for convert an O_PATH fd into a regular one, or to
919 * turn O_RDWR fds into O_RDONLY fds.
921 * This doesn't work on sockets (since they cannot be open()ed, ever).
923 * This implicitly resets the file read index to 0. */
925 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
926 new_fd = open(procfs_path, flags);
933 int read_nr_open(void) {
934 _cleanup_free_ char *nr_open = NULL;
937 /* Returns the kernel's current fd limit, either by reading it of /proc/sys if that works, or using the
938 * hard-coded default compiled-in value of current kernels (1M) if not. This call will never fail. */
940 r = read_one_line_file("/proc/sys/fs/nr_open", &nr_open);
942 log_debug_errno(r, "Failed to read /proc/sys/fs/nr_open, ignoring: %m");
946 r = safe_atoi(nr_open, &v);
948 log_debug_errno(r, "Failed to parse /proc/sys/fs/nr_open value '%s', ignoring: %m", nr_open);
953 /* If we fail, fallback to the hard-coded kernel limit of 1024 * 1024. */