/* SPDX-License-Identifier: LGPL-2.1+ */
-/***
- This file is part of systemd.
-
- Copyright 2010 Lennart Poettering
-***/
#include <errno.h>
#include <fcntl.h>
#include <sys/stat.h>
#include <unistd.h>
+#include "alloc-util.h"
+#include "copy.h"
#include "dirent-util.h"
#include "fd-util.h"
#include "fileio.h"
#include "fs-util.h"
+//#include "io-util.h"
#include "macro.h"
#include "memfd-util.h"
#include "missing.h"
p[1] = safe_close(p[1]);
}
-void close_many(const int fds[], unsigned n_fd) {
- unsigned i;
+void close_many(const int fds[], size_t n_fd) {
+ size_t i;
assert(fds || n_fd <= 0);
return 0;
}
-_pure_ static bool fd_in_set(int fd, const int fdset[], unsigned n_fdset) {
- unsigned i;
+_pure_ static bool fd_in_set(int fd, const int fdset[], size_t n_fdset) {
+ size_t i;
assert(n_fdset == 0 || fdset);
return false;
}
-int close_all_fds(const int except[], unsigned n_except) {
+int close_all_fds(const int except[], size_t n_except) {
_cleanup_closedir_ DIR *d = NULL;
struct dirent *de;
int r = 0;
d = opendir("/proc/self/fd");
if (!d) {
- int fd;
struct rlimit rl;
+ int fd, max_fd;
- /* When /proc isn't available (for example in chroots)
- * the fallback is brute forcing through the fd
+ /* When /proc isn't available (for example in chroots) the fallback is brute forcing through the fd
* table */
assert_se(getrlimit(RLIMIT_NOFILE, &rl) >= 0);
- for (fd = 3; fd < (int) rl.rlim_max; fd ++) {
+
+ if (rl.rlim_max == 0)
+ return -EINVAL;
+
+ /* Let's take special care if the resource limit is set to unlimited, or actually larger than the range
+ * of 'int'. Let's avoid implicit overflows. */
+ max_fd = (rl.rlim_max == RLIM_INFINITY || rl.rlim_max > INT_MAX) ? INT_MAX : (int) (rl.rlim_max - 1);
+
+ for (fd = 3; fd >= 0; fd = fd < max_fd ? fd + 1 : -1) {
int q;
if (fd_in_set(fd, except, n_except))
return -EOPNOTSUPP;
}
+/* When the data is smaller or equal to 64K, try to place the copy in a memfd/pipe */
+#define DATA_FD_MEMORY_LIMIT (64U*1024U)
+
+/* 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. */
+#define DATA_FD_TMP_LIMIT (1024U*1024U)
+
+int fd_duplicate_data_fd(int fd) {
+
+ _cleanup_close_ int copy_fd = -1, tmp_fd = -1;
+ _cleanup_free_ void *remains = NULL;
+ size_t remains_size = 0;
+ const char *td;
+ struct stat st;
+ int r;
+
+ /* Creates a 'data' fd from the specified source fd, containing all the same data in a read-only fashion, but
+ * independent of it (i.e. the source fd can be closed and unmounted after this call succeeded). Tries to be
+ * somewhat smart about where to place the data. In the best case uses a memfd(). If memfd() are not supported
+ * uses a pipe instead. For larger data will use an unlinked file in /tmp, and for even larger data one in
+ * /var/tmp. */
+
+ if (fstat(fd, &st) < 0)
+ return -errno;
+
+ /* For now, let's only accept regular files, sockets, pipes and char devices */
+ if (S_ISDIR(st.st_mode))
+ return -EISDIR;
+ if (S_ISLNK(st.st_mode))
+ return -ELOOP;
+ if (!S_ISREG(st.st_mode) && !S_ISSOCK(st.st_mode) && !S_ISFIFO(st.st_mode) && !S_ISCHR(st.st_mode))
+ return -EBADFD;
+
+ /* If we have reason to believe the data is bounded in size, then let's use memfds or pipes as backing fd. Note
+ * that we use the reported regular file size only as a hint, given that there are plenty special files in
+ * /proc and /sys which report a zero file size but can be read from. */
+
+ if (!S_ISREG(st.st_mode) || st.st_size < DATA_FD_MEMORY_LIMIT) {
+
+ /* Try a memfd first */
+ copy_fd = memfd_new("data-fd");
+ if (copy_fd >= 0) {
+ off_t f;
+
+ r = copy_bytes(fd, copy_fd, DATA_FD_MEMORY_LIMIT, 0);
+ if (r < 0)
+ return r;
+
+ f = lseek(copy_fd, 0, SEEK_SET);
+ if (f != 0)
+ return -errno;
+
+ if (r == 0) {
+ /* Did it fit into the limit? If so, we are done. */
+ r = memfd_set_sealed(copy_fd);
+ if (r < 0)
+ return r;
+
+ return TAKE_FD(copy_fd);
+ }
+
+ /* Hmm, pity, this didn't fit. Let's fall back to /tmp then, see below */
+
+ } else {
+ _cleanup_(close_pairp) int pipefds[2] = { -1, -1 };
+ int isz;
+
+ /* If memfds aren't available, use a pipe. Set O_NONBLOCK so that we will get EAGAIN rather
+ * then block indefinitely when we hit the pipe size limit */
+
+ if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0)
+ return -errno;
+
+ isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
+ if (isz < 0)
+ return -errno;
+
+ /* Try to enlarge the pipe size if necessary */
+ if ((size_t) isz < DATA_FD_MEMORY_LIMIT) {
+
+ (void) fcntl(pipefds[1], F_SETPIPE_SZ, DATA_FD_MEMORY_LIMIT);
+
+ isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
+ if (isz < 0)
+ return -errno;
+ }
+
+ if ((size_t) isz >= DATA_FD_MEMORY_LIMIT) {
+
+ r = copy_bytes_full(fd, pipefds[1], DATA_FD_MEMORY_LIMIT, 0, &remains, &remains_size);
+ if (r < 0 && r != -EAGAIN)
+ return r; /* If we get EAGAIN it could be because of the source or because of
+ * the destination fd, we can't know, as sendfile() and friends won't
+ * tell us. Hence, treat this as reason to fall back, just to be
+ * sure. */
+ if (r == 0) {
+ /* Everything fit in, yay! */
+ (void) fd_nonblock(pipefds[0], false);
+
+ return TAKE_FD(pipefds[0]);
+ }
+
+ /* Things didn't fit in. But we read data into the pipe, let's remember that, so that
+ * when writing the new file we incorporate this first. */
+ copy_fd = TAKE_FD(pipefds[0]);
+ }
+ }
+ }
+
+ /* If we have reason to believe this will fit fine in /tmp, then use that as first fallback. */
+ if ((!S_ISREG(st.st_mode) || st.st_size < DATA_FD_TMP_LIMIT) &&
+ (DATA_FD_MEMORY_LIMIT + remains_size) < DATA_FD_TMP_LIMIT) {
+ off_t f;
+
+ tmp_fd = open_tmpfile_unlinkable(NULL /* NULL as directory means /tmp */, O_RDWR|O_CLOEXEC);
+ if (tmp_fd < 0)
+ return tmp_fd;
+
+ if (copy_fd >= 0) {
+ /* If we tried a memfd/pipe first and it ended up being too large, then copy this into the
+ * temporary file first. */
+
+ r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, 0);
+ if (r < 0)
+ return r;
+
+ assert(r == 0);
+ }
+
+ if (remains_size > 0) {
+ /* If there were remaining bytes (i.e. read into memory, but not written out yet) from the
+ * failed copy operation, let's flush them out next. */
+
+ r = loop_write(tmp_fd, remains, remains_size, false);
+ if (r < 0)
+ return r;
+ }
+
+ r = copy_bytes(fd, tmp_fd, DATA_FD_TMP_LIMIT - DATA_FD_MEMORY_LIMIT - remains_size, COPY_REFLINK);
+ if (r < 0)
+ return r;
+ if (r == 0)
+ goto finish; /* Yay, it fit in */
+
+ /* It didn't fit in. Let's not forget to use what we already used */
+ f = lseek(tmp_fd, 0, SEEK_SET);
+ if (f != 0)
+ return -errno;
+
+ safe_close(copy_fd);
+ copy_fd = TAKE_FD(tmp_fd);
+
+ remains = mfree(remains);
+ remains_size = 0;
+ }
+
+ /* As last fallback use /var/tmp */
+ r = var_tmp_dir(&td);
+ if (r < 0)
+ return r;
+
+ tmp_fd = open_tmpfile_unlinkable(td, O_RDWR|O_CLOEXEC);
+ if (tmp_fd < 0)
+ return tmp_fd;
+
+ if (copy_fd >= 0) {
+ /* If we tried a memfd/pipe first, or a file in /tmp, and it ended up being too large, than copy this
+ * into the temporary file first. */
+ r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
+ if (r < 0)
+ return r;
+
+ assert(r == 0);
+ }
+
+ if (remains_size > 0) {
+ /* Then, copy in any read but not yet written bytes. */
+ r = loop_write(tmp_fd, remains, remains_size, false);
+ if (r < 0)
+ return r;
+ }
+
+ /* Copy in the rest */
+ r = copy_bytes(fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
+ if (r < 0)
+ return r;
+
+ assert(r == 0);
+
+finish:
+ /* Now convert the O_RDWR file descriptor into an O_RDONLY one (and as side effect seek to the beginning of the
+ * file again */
+
+ return fd_reopen(tmp_fd, O_RDONLY|O_CLOEXEC);
+}
+
int fd_move_above_stdio(int fd) {
int flags, copy;
PROTECT_ERRNO;
return new_fd;
}
+
+int read_nr_open(void) {
+ _cleanup_free_ char *nr_open = NULL;
+ int r;
+
+ /* Returns the kernel's current fd limit, either by reading it of /proc/sys if that works, or using the
+ * hard-coded default compiled-in value of current kernels (1M) if not. This call will never fail. */
+
+ r = read_one_line_file("/proc/sys/fs/nr_open", &nr_open);
+ if (r < 0)
+ log_debug_errno(r, "Failed to read /proc/sys/fs/nr_open, ignoring: %m");
+ else {
+ int v;
+
+ r = safe_atoi(nr_open, &v);
+ if (r < 0)
+ log_debug_errno(r, "Failed to parse /proc/sys/fs/nr_open value '%s', ignoring: %m", nr_open);
+ else
+ return v;
+ }
+
+ /* If we fail, fallback to the hard-coded kernel limit of 1024 * 1024. */
+ return 1024 * 1024;
+}