1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
5 #include <sys/resource.h>
9 #include "alloc-util.h"
11 #include "dirent-util.h"
17 #include "memfd-util.h"
18 #include "missing_fcntl.h"
19 #include "missing_syscall.h"
20 #include "parse-util.h"
21 #include "path-util.h"
22 #include "process-util.h"
23 #include "socket-util.h"
24 #include "sort-util.h"
25 #include "stat-util.h"
26 #include "stdio-util.h"
27 #include "tmpfile-util.h"
30 /* The maximum number of iterations in the loop to close descriptors in the fallback case
31 * when /proc/self/fd/ is inaccessible. */
32 #define MAX_FD_LOOP_LIMIT (1024*1024)
34 int close_nointr(int fd
) {
41 * Just ignore EINTR; a retry loop is the wrong thing to do on
44 * http://lkml.indiana.edu/hypermail/linux/kernel/0509.1/0877.html
45 * https://bugzilla.gnome.org/show_bug.cgi?id=682819
46 * http://utcc.utoronto.ca/~cks/space/blog/unix/CloseEINTR
47 * https://sites.google.com/site/michaelsafyan/software-engineering/checkforeintrwheninvokingclosethinkagain
55 int safe_close(int fd
) {
58 * Like close_nointr() but cannot fail. Guarantees errno is
59 * unchanged. Is a NOP with negative fds passed, and returns
60 * -1, so that it can be used in this syntax:
62 * fd = safe_close(fd);
68 /* The kernel might return pretty much any error code
69 * via close(), but the fd will be closed anyway. The
70 * only condition we want to check for here is whether
71 * the fd was invalid at all... */
73 assert_se(close_nointr(fd
) != -EBADF
);
79 void safe_close_pair(int p
[static 2]) {
83 /* Special case pairs which use the same fd in both
85 p
[0] = p
[1] = safe_close(p
[0]);
89 p
[0] = safe_close(p
[0]);
90 p
[1] = safe_close(p
[1]);
93 void close_many(const int fds
[], size_t n_fd
) {
94 assert(fds
|| n_fd
<= 0);
96 for (size_t i
= 0; i
< n_fd
; i
++)
100 int fclose_nointr(FILE *f
) {
103 /* Same as close_nointr(), but for fclose() */
105 errno
= 0; /* Extra safety: if the FILE* object is not encapsulating an fd, it might not set errno
106 * correctly. Let's hence initialize it to zero first, so that we aren't confused by any
107 * prior errno here */
114 return errno_or_else(EIO
);
117 FILE* safe_fclose(FILE *f
) {
119 /* Same as safe_close(), but for fclose() */
124 assert_se(fclose_nointr(f
) != -EBADF
);
130 DIR* safe_closedir(DIR *d
) {
135 assert_se(closedir(d
) >= 0 || errno
!= EBADF
);
141 int fd_nonblock(int fd
, bool nonblock
) {
146 flags
= fcntl(fd
, F_GETFL
, 0);
150 nflags
= UPDATE_FLAG(flags
, O_NONBLOCK
, nonblock
);
154 if (fcntl(fd
, F_SETFL
, nflags
) < 0)
160 int fd_cloexec(int fd
, bool cloexec
) {
165 flags
= fcntl(fd
, F_GETFD
, 0);
169 nflags
= UPDATE_FLAG(flags
, FD_CLOEXEC
, cloexec
);
173 if (fcntl(fd
, F_SETFD
, nflags
) < 0)
179 _pure_
static bool fd_in_set(int fd
, const int fdset
[], size_t n_fdset
) {
180 assert(n_fdset
== 0 || fdset
);
182 for (size_t i
= 0; i
< n_fdset
; i
++)
189 static int get_max_fd(void) {
193 /* Return the highest possible fd, based RLIMIT_NOFILE, but enforcing FD_SETSIZE-1 as lower boundary
194 * and INT_MAX as upper boundary. */
196 if (getrlimit(RLIMIT_NOFILE
, &rl
) < 0)
199 m
= MAX(rl
.rlim_cur
, rl
.rlim_max
);
200 if (m
< FD_SETSIZE
) /* Let's always cover at least 1024 fds */
203 if (m
== RLIM_INFINITY
|| m
> INT_MAX
) /* Saturate on overflow. After all fds are "int", hence can
204 * never be above INT_MAX */
207 return (int) (m
- 1);
210 int close_all_fds(const int except
[], size_t n_except
) {
211 static bool have_close_range
= true; /* Assume we live in the future */
212 _cleanup_closedir_
DIR *d
= NULL
;
216 assert(n_except
== 0 || except
);
218 if (have_close_range
) {
219 /* In the best case we have close_range() to close all fds between a start and an end fd,
220 * which we can use on the "inverted" exception array, i.e. all intervals between all
221 * adjacent pairs from the sorted exception array. This changes loop complexity from O(n)
222 * where n is number of open fds to O(m⋅log(m)) where m is the number of fds to keep
223 * open. Given that we assume n ≫ m that's preferable to us. */
226 /* Close everything. Yay! */
228 if (close_range(3, -1, 0) >= 0)
231 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && !ERRNO_IS_PRIVILEGE(errno
))
234 have_close_range
= false;
236 _cleanup_free_
int *sorted_malloc
= NULL
;
240 assert(n_except
< SIZE_MAX
);
241 n_sorted
= n_except
+ 1;
243 if (n_sorted
> 64) /* Use heap for large numbers of fds, stack otherwise */
244 sorted
= sorted_malloc
= new(int, n_sorted
);
246 sorted
= newa(int, n_sorted
);
251 memcpy(sorted
, except
, n_except
* sizeof(int));
253 /* Let's add fd 2 to the list of fds, to simplify the loop below, as this
254 * allows us to cover the head of the array the same way as the body */
255 sorted
[n_sorted
-1] = 2;
257 typesafe_qsort(sorted
, n_sorted
, cmp_int
);
259 for (size_t i
= 0; i
< n_sorted
-1; i
++) {
262 start
= MAX(sorted
[i
], 2); /* The first three fds shall always remain open */
263 end
= MAX(sorted
[i
+1], 2);
265 assert(end
>= start
);
267 if (end
- start
<= 1)
270 /* Close everything between the start and end fds (both of which shall stay open) */
271 if (close_range(start
+ 1, end
- 1, 0) < 0) {
272 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && !ERRNO_IS_PRIVILEGE(errno
))
275 have_close_range
= false;
279 c
+= end
- start
- 1;
282 if (have_close_range
) {
283 /* The loop succeeded. Let's now close everything beyond the end */
285 if (sorted
[n_sorted
-1] >= INT_MAX
) /* Dont let the addition below overflow */
288 if (close_range(sorted
[n_sorted
-1] + 1, -1, 0) >= 0)
291 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && !ERRNO_IS_PRIVILEGE(errno
))
294 have_close_range
= false;
299 /* Fallback on OOM or if close_range() is not supported */
302 d
= opendir("/proc/self/fd");
306 /* When /proc isn't available (for example in chroots) the fallback is brute forcing through
309 max_fd
= get_max_fd();
313 /* Refuse to do the loop over more too many elements. It's better to fail immediately than to
314 * spin the CPU for a long time. */
315 if (max_fd
> MAX_FD_LOOP_LIMIT
)
316 return log_debug_errno(SYNTHETIC_ERRNO(EPERM
),
317 "/proc/self/fd is inaccessible. Refusing to loop over %d potential fds.",
320 for (fd
= 3; fd
>= 0; fd
= fd
< max_fd
? fd
+ 1 : -1) {
323 if (fd_in_set(fd
, except
, n_except
))
326 q
= close_nointr(fd
);
327 if (q
< 0 && q
!= -EBADF
&& r
>= 0)
334 FOREACH_DIRENT(de
, d
, return -errno
) {
337 if (safe_atoi(de
->d_name
, &fd
) < 0)
338 /* Let's better ignore this, just in case */
347 if (fd_in_set(fd
, except
, n_except
))
350 q
= close_nointr(fd
);
351 if (q
< 0 && q
!= -EBADF
&& r
>= 0) /* Valgrind has its own FD and doesn't want to have it closed */
358 int same_fd(int a
, int b
) {
359 struct stat sta
, stb
;
366 /* Compares two file descriptors. Note that semantics are
367 * quite different depending on whether we have kcmp() or we
368 * don't. If we have kcmp() this will only return true for
369 * dup()ed file descriptors, but not otherwise. If we don't
370 * have kcmp() this will also return true for two fds of the same
371 * file, created by separate open() calls. Since we use this
372 * call mostly for filtering out duplicates in the fd store
373 * this difference hopefully doesn't matter too much. */
378 /* Try to use kcmp() if we have it. */
379 pid
= getpid_cached();
380 r
= kcmp(pid
, pid
, KCMP_FILE
, a
, b
);
385 if (!IN_SET(errno
, ENOSYS
, EACCES
, EPERM
))
388 /* We don't have kcmp(), use fstat() instead. */
389 if (fstat(a
, &sta
) < 0)
392 if (fstat(b
, &stb
) < 0)
395 if ((sta
.st_mode
& S_IFMT
) != (stb
.st_mode
& S_IFMT
))
398 /* We consider all device fds different, since two device fds
399 * might refer to quite different device contexts even though
400 * they share the same inode and backing dev_t. */
402 if (S_ISCHR(sta
.st_mode
) || S_ISBLK(sta
.st_mode
))
405 if (sta
.st_dev
!= stb
.st_dev
|| sta
.st_ino
!= stb
.st_ino
)
408 /* The fds refer to the same inode on disk, let's also check
409 * if they have the same fd flags. This is useful to
410 * distinguish the read and write side of a pipe created with
412 fa
= fcntl(a
, F_GETFL
);
416 fb
= fcntl(b
, F_GETFL
);
423 void cmsg_close_all(struct msghdr
*mh
) {
424 struct cmsghdr
*cmsg
;
428 CMSG_FOREACH(cmsg
, mh
)
429 if (cmsg
->cmsg_level
== SOL_SOCKET
&& cmsg
->cmsg_type
== SCM_RIGHTS
)
430 close_many((int*) CMSG_DATA(cmsg
), (cmsg
->cmsg_len
- CMSG_LEN(0)) / sizeof(int));
433 bool fdname_is_valid(const char *s
) {
436 /* Validates a name for $LISTEN_FDNAMES. We basically allow
437 * everything ASCII that's not a control character. Also, as
438 * special exception the ":" character is not allowed, as we
439 * use that as field separator in $LISTEN_FDNAMES.
441 * Note that the empty string is explicitly allowed
442 * here. However, we limit the length of the names to 255
448 for (p
= s
; *p
; p
++) {
457 return p
- s
<= FDNAME_MAX
;
460 int fd_get_path(int fd
, char **ret
) {
461 char procfs_path
[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int)];
464 xsprintf(procfs_path
, "/proc/self/fd/%i", fd
);
465 r
= readlink_malloc(procfs_path
, ret
);
467 /* ENOENT can mean two things: that the fd does not exist or that /proc is not mounted. Let's make
468 * things debuggable and distinguish the two. */
470 if (proc_mounted() == 0)
471 return -ENOSYS
; /* /proc is not available or not set up properly, we're most likely in some chroot
473 return -EBADF
; /* The directory exists, hence it's the fd that doesn't. */
479 int move_fd(int from
, int to
, int cloexec
) {
482 /* Move fd 'from' to 'to', make sure FD_CLOEXEC remains equal if requested, and release the old fd. If
483 * 'cloexec' is passed as -1, the original FD_CLOEXEC is inherited for the new fd. If it is 0, it is turned
484 * off, if it is > 0 it is turned on. */
494 r
= fd_cloexec(to
, cloexec
);
505 fl
= fcntl(from
, F_GETFD
, 0);
509 cloexec
= !!(fl
& FD_CLOEXEC
);
512 r
= dup3(from
, to
, cloexec
? O_CLOEXEC
: 0);
523 int acquire_data_fd(const void *data
, size_t size
, unsigned flags
) {
525 _cleanup_close_pair_
int pipefds
[2] = { -1, -1 };
526 char pattern
[] = "/dev/shm/data-fd-XXXXXX";
527 _cleanup_close_
int fd
= -1;
532 assert(data
|| size
== 0);
534 /* Acquire a read-only file descriptor that when read from returns the specified data. This is much more
535 * complex than I wish it was. But here's why:
537 * a) First we try to use memfds. They are the best option, as we can seal them nicely to make them
538 * read-only. Unfortunately they require kernel 3.17, and – at the time of writing – we still support 3.14.
540 * b) Then, we try classic pipes. They are the second best options, as we can close the writing side, retaining
541 * a nicely read-only fd in the reading side. However, they are by default quite small, and unprivileged
542 * clients can only bump their size to a system-wide limit, which might be quite low.
544 * c) Then, we try an O_TMPFILE file in /dev/shm (that dir is the only suitable one known to exist from
545 * earliest boot on). To make it read-only we open the fd a second time with O_RDONLY via
546 * /proc/self/<fd>. Unfortunately O_TMPFILE is not available on older kernels on tmpfs.
548 * d) Finally, we try creating a regular file in /dev/shm, which we then delete.
550 * It sucks a bit that depending on the situation we return very different objects here, but that's Linux I
553 if (size
== 0 && ((flags
& ACQUIRE_NO_DEV_NULL
) == 0)) {
554 /* As a special case, return /dev/null if we have been called for an empty data block */
555 r
= open("/dev/null", O_RDONLY
|O_CLOEXEC
|O_NOCTTY
);
562 if ((flags
& ACQUIRE_NO_MEMFD
) == 0) {
563 fd
= memfd_new("data-fd");
567 n
= write(fd
, data
, size
);
570 if ((size_t) n
!= size
)
573 f
= lseek(fd
, 0, SEEK_SET
);
577 r
= memfd_set_sealed(fd
);
585 if ((flags
& ACQUIRE_NO_PIPE
) == 0) {
586 if (pipe2(pipefds
, O_CLOEXEC
|O_NONBLOCK
) < 0)
589 isz
= fcntl(pipefds
[1], F_GETPIPE_SZ
, 0);
593 if ((size_t) isz
< size
) {
595 if (isz
< 0 || (size_t) isz
!= size
)
598 /* Try to bump the pipe size */
599 (void) fcntl(pipefds
[1], F_SETPIPE_SZ
, isz
);
601 /* See if that worked */
602 isz
= fcntl(pipefds
[1], F_GETPIPE_SZ
, 0);
606 if ((size_t) isz
< size
)
610 n
= write(pipefds
[1], data
, size
);
613 if ((size_t) n
!= size
)
616 (void) fd_nonblock(pipefds
[0], false);
618 return TAKE_FD(pipefds
[0]);
622 if ((flags
& ACQUIRE_NO_TMPFILE
) == 0) {
623 fd
= open("/dev/shm", O_RDWR
|O_TMPFILE
|O_CLOEXEC
, 0500);
625 goto try_dev_shm_without_o_tmpfile
;
627 n
= write(fd
, data
, size
);
630 if ((size_t) n
!= size
)
633 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
634 return fd_reopen(fd
, O_RDONLY
|O_CLOEXEC
);
637 try_dev_shm_without_o_tmpfile
:
638 if ((flags
& ACQUIRE_NO_REGULAR
) == 0) {
639 fd
= mkostemp_safe(pattern
);
643 n
= write(fd
, data
, size
);
646 goto unlink_and_return
;
648 if ((size_t) n
!= size
) {
650 goto unlink_and_return
;
653 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
654 r
= open(pattern
, O_RDONLY
|O_CLOEXEC
);
659 (void) unlink(pattern
);
666 /* When the data is smaller or equal to 64K, try to place the copy in a memfd/pipe */
667 #define DATA_FD_MEMORY_LIMIT (64U*1024U)
669 /* 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. */
670 #define DATA_FD_TMP_LIMIT (1024U*1024U)
672 int fd_duplicate_data_fd(int fd
) {
674 _cleanup_close_
int copy_fd
= -1, tmp_fd
= -1;
675 _cleanup_free_
void *remains
= NULL
;
676 size_t remains_size
= 0;
681 /* Creates a 'data' fd from the specified source fd, containing all the same data in a read-only fashion, but
682 * independent of it (i.e. the source fd can be closed and unmounted after this call succeeded). Tries to be
683 * somewhat smart about where to place the data. In the best case uses a memfd(). If memfd() are not supported
684 * uses a pipe instead. For larger data will use an unlinked file in /tmp, and for even larger data one in
687 if (fstat(fd
, &st
) < 0)
690 /* For now, let's only accept regular files, sockets, pipes and char devices */
691 if (S_ISDIR(st
.st_mode
))
693 if (S_ISLNK(st
.st_mode
))
695 if (!S_ISREG(st
.st_mode
) && !S_ISSOCK(st
.st_mode
) && !S_ISFIFO(st
.st_mode
) && !S_ISCHR(st
.st_mode
))
698 /* If we have reason to believe the data is bounded in size, then let's use memfds or pipes as backing fd. Note
699 * that we use the reported regular file size only as a hint, given that there are plenty special files in
700 * /proc and /sys which report a zero file size but can be read from. */
702 if (!S_ISREG(st
.st_mode
) || st
.st_size
< DATA_FD_MEMORY_LIMIT
) {
704 /* Try a memfd first */
705 copy_fd
= memfd_new("data-fd");
709 r
= copy_bytes(fd
, copy_fd
, DATA_FD_MEMORY_LIMIT
, 0);
713 f
= lseek(copy_fd
, 0, SEEK_SET
);
718 /* Did it fit into the limit? If so, we are done. */
719 r
= memfd_set_sealed(copy_fd
);
723 return TAKE_FD(copy_fd
);
726 /* Hmm, pity, this didn't fit. Let's fall back to /tmp then, see below */
729 _cleanup_(close_pairp
) int pipefds
[2] = { -1, -1 };
732 /* If memfds aren't available, use a pipe. Set O_NONBLOCK so that we will get EAGAIN rather
733 * then block indefinitely when we hit the pipe size limit */
735 if (pipe2(pipefds
, O_CLOEXEC
|O_NONBLOCK
) < 0)
738 isz
= fcntl(pipefds
[1], F_GETPIPE_SZ
, 0);
742 /* Try to enlarge the pipe size if necessary */
743 if ((size_t) isz
< DATA_FD_MEMORY_LIMIT
) {
745 (void) fcntl(pipefds
[1], F_SETPIPE_SZ
, DATA_FD_MEMORY_LIMIT
);
747 isz
= fcntl(pipefds
[1], F_GETPIPE_SZ
, 0);
752 if ((size_t) isz
>= DATA_FD_MEMORY_LIMIT
) {
754 r
= copy_bytes_full(fd
, pipefds
[1], DATA_FD_MEMORY_LIMIT
, 0, &remains
, &remains_size
, NULL
, NULL
);
755 if (r
< 0 && r
!= -EAGAIN
)
756 return r
; /* If we get EAGAIN it could be because of the source or because of
757 * the destination fd, we can't know, as sendfile() and friends won't
758 * tell us. Hence, treat this as reason to fall back, just to be
761 /* Everything fit in, yay! */
762 (void) fd_nonblock(pipefds
[0], false);
764 return TAKE_FD(pipefds
[0]);
767 /* Things didn't fit in. But we read data into the pipe, let's remember that, so that
768 * when writing the new file we incorporate this first. */
769 copy_fd
= TAKE_FD(pipefds
[0]);
774 /* If we have reason to believe this will fit fine in /tmp, then use that as first fallback. */
775 if ((!S_ISREG(st
.st_mode
) || st
.st_size
< DATA_FD_TMP_LIMIT
) &&
776 (DATA_FD_MEMORY_LIMIT
+ remains_size
) < DATA_FD_TMP_LIMIT
) {
779 tmp_fd
= open_tmpfile_unlinkable(NULL
/* NULL as directory means /tmp */, O_RDWR
|O_CLOEXEC
);
784 /* If we tried a memfd/pipe first and it ended up being too large, then copy this into the
785 * temporary file first. */
787 r
= copy_bytes(copy_fd
, tmp_fd
, UINT64_MAX
, 0);
794 if (remains_size
> 0) {
795 /* If there were remaining bytes (i.e. read into memory, but not written out yet) from the
796 * failed copy operation, let's flush them out next. */
798 r
= loop_write(tmp_fd
, remains
, remains_size
, false);
803 r
= copy_bytes(fd
, tmp_fd
, DATA_FD_TMP_LIMIT
- DATA_FD_MEMORY_LIMIT
- remains_size
, COPY_REFLINK
);
807 goto finish
; /* Yay, it fit in */
809 /* It didn't fit in. Let's not forget to use what we already used */
810 f
= lseek(tmp_fd
, 0, SEEK_SET
);
814 CLOSE_AND_REPLACE(copy_fd
, tmp_fd
);
816 remains
= mfree(remains
);
820 /* As last fallback use /var/tmp */
821 r
= var_tmp_dir(&td
);
825 tmp_fd
= open_tmpfile_unlinkable(td
, O_RDWR
|O_CLOEXEC
);
830 /* If we tried a memfd/pipe first, or a file in /tmp, and it ended up being too large, than copy this
831 * into the temporary file first. */
832 r
= copy_bytes(copy_fd
, tmp_fd
, UINT64_MAX
, COPY_REFLINK
);
839 if (remains_size
> 0) {
840 /* Then, copy in any read but not yet written bytes. */
841 r
= loop_write(tmp_fd
, remains
, remains_size
, false);
846 /* Copy in the rest */
847 r
= copy_bytes(fd
, tmp_fd
, UINT64_MAX
, COPY_REFLINK
);
854 /* Now convert the O_RDWR file descriptor into an O_RDONLY one (and as side effect seek to the beginning of the
857 return fd_reopen(tmp_fd
, O_RDONLY
|O_CLOEXEC
);
860 int fd_move_above_stdio(int fd
) {
864 /* Moves the specified file descriptor if possible out of the range [0…2], i.e. the range of
865 * stdin/stdout/stderr. If it can't be moved outside of this range the original file descriptor is
866 * returned. This call is supposed to be used for long-lasting file descriptors we allocate in our code that
867 * might get loaded into foreign code, and where we want ensure our fds are unlikely used accidentally as
868 * stdin/stdout/stderr of unrelated code.
870 * Note that this doesn't fix any real bugs, it just makes it less likely that our code will be affected by
871 * buggy code from others that mindlessly invokes 'fprintf(stderr, …' or similar in places where stderr has
872 * been closed before.
874 * This function is written in a "best-effort" and "least-impact" style. This means whenever we encounter an
875 * error we simply return the original file descriptor, and we do not touch errno. */
877 if (fd
< 0 || fd
> 2)
880 flags
= fcntl(fd
, F_GETFD
, 0);
884 if (flags
& FD_CLOEXEC
)
885 copy
= fcntl(fd
, F_DUPFD_CLOEXEC
, 3);
887 copy
= fcntl(fd
, F_DUPFD
, 3);
897 int rearrange_stdio(int original_input_fd
, int original_output_fd
, int original_error_fd
) {
899 int fd
[3] = { /* Put together an array of fds we work on */
906 null_fd
= -1, /* if we open /dev/null, we store the fd to it here */
907 copy_fd
[3] = { -1, -1, -1 }; /* This contains all fds we duplicate here temporarily, and hence need to close at the end */
908 bool null_readable
, null_writable
;
910 /* Sets up stdin, stdout, stderr with the three file descriptors passed in. If any of the descriptors is
911 * specified as -1 it will be connected with /dev/null instead. If any of the file descriptors is passed as
912 * itself (e.g. stdin as STDIN_FILENO) it is left unmodified, but the O_CLOEXEC bit is turned off should it be
915 * Note that if any of the passed file descriptors are > 2 they will be closed — both on success and on
916 * failure! Thus, callers should assume that when this function returns the input fds are invalidated.
918 * Note that when this function fails stdin/stdout/stderr might remain half set up!
920 * O_CLOEXEC is turned off for all three file descriptors (which is how it should be for
921 * stdin/stdout/stderr). */
923 null_readable
= original_input_fd
< 0;
924 null_writable
= original_output_fd
< 0 || original_error_fd
< 0;
926 /* First step, open /dev/null once, if we need it */
927 if (null_readable
|| null_writable
) {
929 /* Let's open this with O_CLOEXEC first, and convert it to non-O_CLOEXEC when we move the fd to the final position. */
930 null_fd
= open("/dev/null", (null_readable
&& null_writable
? O_RDWR
:
931 null_readable
? O_RDONLY
: O_WRONLY
) | O_CLOEXEC
);
937 /* If this fd is in the 0…2 range, let's move it out of it */
941 copy
= fcntl(null_fd
, F_DUPFD_CLOEXEC
, 3); /* Duplicate this with O_CLOEXEC set */
947 CLOSE_AND_REPLACE(null_fd
, copy
);
951 /* Let's assemble fd[] with the fds to install in place of stdin/stdout/stderr */
952 for (i
= 0; i
< 3; i
++) {
955 fd
[i
] = null_fd
; /* A negative parameter means: connect this one to /dev/null */
956 else if (fd
[i
] != i
&& fd
[i
] < 3) {
957 /* 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. */
958 copy_fd
[i
] = fcntl(fd
[i
], F_DUPFD_CLOEXEC
, 3); /* Duplicate this with O_CLOEXEC set */
959 if (copy_fd
[i
] < 0) {
968 /* At this point we now have the fds to use in fd[], and they are all above the stdio range, so that we
969 * have freedom to move them around. If the fds already were at the right places then the specific fds are
970 * -1. Let's now move them to the right places. This is the point of no return. */
971 for (i
= 0; i
< 3; i
++) {
975 /* fd is already in place, but let's make sure O_CLOEXEC is off */
976 r
= fd_cloexec(i
, false);
983 if (dup2(fd
[i
], i
) < 0) { /* Turns off O_CLOEXEC on the new fd. */
993 /* Close the original fds, but only if they were outside of the stdio range. Also, properly check for the same
994 * fd passed in multiple times. */
995 safe_close_above_stdio(original_input_fd
);
996 if (original_output_fd
!= original_input_fd
)
997 safe_close_above_stdio(original_output_fd
);
998 if (original_error_fd
!= original_input_fd
&& original_error_fd
!= original_output_fd
)
999 safe_close_above_stdio(original_error_fd
);
1001 /* Close the copies we moved > 2 */
1002 for (i
= 0; i
< 3; i
++)
1003 safe_close(copy_fd
[i
]);
1005 /* Close our null fd, if it's > 2 */
1006 safe_close_above_stdio(null_fd
);
1011 int fd_reopen(int fd
, int flags
) {
1012 char procfs_path
[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int)];
1015 /* Reopens the specified fd with new flags. This is useful for convert an O_PATH fd into a regular one, or to
1016 * turn O_RDWR fds into O_RDONLY fds.
1018 * This doesn't work on sockets (since they cannot be open()ed, ever).
1020 * This implicitly resets the file read index to 0. */
1022 xsprintf(procfs_path
, "/proc/self/fd/%i", fd
);
1023 new_fd
= open(procfs_path
, flags
);
1025 if (errno
!= ENOENT
)
1028 if (proc_mounted() == 0)
1029 return -ENOSYS
; /* if we have no /proc/, the concept is not implementable */
1037 int read_nr_open(void) {
1038 _cleanup_free_
char *nr_open
= NULL
;
1041 /* Returns the kernel's current fd limit, either by reading it of /proc/sys if that works, or using the
1042 * hard-coded default compiled-in value of current kernels (1M) if not. This call will never fail. */
1044 r
= read_one_line_file("/proc/sys/fs/nr_open", &nr_open
);
1046 log_debug_errno(r
, "Failed to read /proc/sys/fs/nr_open, ignoring: %m");
1050 r
= safe_atoi(nr_open
, &v
);
1052 log_debug_errno(r
, "Failed to parse /proc/sys/fs/nr_open value '%s', ignoring: %m", nr_open
);
1057 /* If we fail, fall back to the hard-coded kernel limit of 1024 * 1024. */