HALF ECHO LEARN / REFERENCE COLLECTION

v262.2

src/shared/mount-util.c

97,047 bytesGit mode 100644Download original fileLGPL-2.1-or-later

Git blob: c8f0fd53d4dbb8a72edef75f564df7d9d77bbaae

Lines 1–1000 of 2,291

/* SPDX-License-Identifier: LGPL-2.1-or-later */
 
#include <stdlib.h>
#include <sys/mount.h>
#include <sys/socket.h>
#include <sys/stat.h>
#include <unistd.h>
 
#include "alloc-util.h"
#include "chase.h"
#include "creds-util.h"
#include "dissect-image.h"
#include "errno-util.h"
#include "extract-word.h"
#include "fd-util.h"
#include "fileio.h"
#include "format-util.h"
#include "fs-util.h"
#include "fstab-util.h"
#include "glyph-util.h"
#include "hashmap.h"
#include "libmount-util.h"
#include "log.h"
#include "mkdir.h"
#include "mount-util.h"
#include "mountpoint-util.h"
#include "namespace-util.h"
#include "os-util.h"
#include "path-util.h"
#include "pidref.h"
#include "process-util.h"
#include "runtime-scope.h"
#include "set.h"
#include "socket-util.h"
#include "sort-util.h"
#include "stat-util.h"
#include "string-util.h"
#include "strv.h"
#include "tmpfile-util.h"
#include "user-util.h"
 
int umount_recursive_full(const char *prefix, int flags, char **keep) {
#if HAVE_LIBMOUNT
        _cleanup_fclose_ FILE *f = NULL;
        int n = 0, r;
 
        /* Try to umount everything recursively below a directory. Also, take care of stacked mounts, and
         * keep unmounting them until they are gone.
         *
         * Each pass unmounts everything it can from a single snapshot of the mount table, and we only
         * re-parse the table for another pass if the previous one made progress. Mounts that could not be
         * unmounted yet (e.g. because a child of theirs was still around) are retried on the next pass.
         *
         * Note that an unmount can invalidate later entries of the snapshot: with MNT_DETACH a parent takes
         * its children along, and their old paths then resolve through whatever the parent was covering,
         * possibly via symlinks and hence to some other mount outside of the prefix or on the keep list.
         * Hence verify that a path still refers to the mount we read from the table before unmounting it,
         * and if anything looks off after we unmounted something, start over with a fresh snapshot. A
         * snapshot without unmounts of ours can be trusted, hence there we move on to the next entry instead
         * (so that one stuck mount doesn't keep us from unmounting the others), and if the check itself fails
         * we try to unmount anyway, since that might work even if we cannot look at the mount. */
 
        f = fopen("/proc/self/mountinfo", "re"); /* Pin the file, in case we unmount /proc/ as part of the logic here */
        if (!f)
                return log_debug_errno(errno, "Failed to open %s: %m", "/proc/self/mountinfo");
 
        for (;;) {
                _cleanup_(mnt_free_tablep) struct libmnt_table *table = NULL;
                _cleanup_(mnt_free_iterp) struct libmnt_iter *iter = NULL;
                bool again = false;
 
                r = libmount_parse_full("/proc/self/mountinfo", f, MNT_ITER_BACKWARD, &table, &iter);
                if (r < 0)
                        return log_debug_errno(r, "Failed to parse /proc/self/mountinfo: %m");
 
                for (;;) {
                        bool shall_keep = false;
                        struct libmnt_fs *fs;
                        const char *path;
 
                        r = sym_mnt_table_next_fs(table, iter, &fs);
                        if (r == 1)
                                break;
                        if (r < 0)
                                return log_debug_errno(r, "Failed to get next entry from /proc/self/mountinfo: %m");
 
                        path = sym_mnt_fs_get_target(fs);
                        if (!path)
                                continue;
 
                        if (prefix && !path_startswith(path, prefix)) {
                                // FIXME: This is extremely noisy, we're probably doing something very wrong
                                // to trigger this so often, needs more investigation.
                                // log_trace("Not unmounting %s, outside of prefix: %s", path, prefix);
                                continue;
                        }
 
                        STRV_FOREACH(k, keep)
                                /* Match against anything in the path to the dirs to keep, or below the dirs to keep */
                                if (path_startswith(path, *k) || path_startswith(*k, path)) {
                                        shall_keep = true;
                                        break;
                                }
                        if (shall_keep) {
                                log_debug("Not unmounting %s, referenced by keep list.", path);
                                continue;
                        }
 
                        r = libmount_fs_id_matches_path(fs, path);
                        if (r <= 0 && again)
                                break;
                        if (r == 0)
                                continue;
 
                        if (umount2(path, flags | UMOUNT_NOFOLLOW) < 0) {
                                log_debug_errno(errno, "Failed to umount %s, ignoring: %m", path);
                                if (again)
                                        break;
                                continue;
                        }
 
                        log_trace("Successfully unmounted %s", path);
 
                        again = true;
                        n++;
                }
 
                if (!again)
                        break;
 
                rewind(f);
        }
 
        return n;
#else
        return log_debug_errno(SYNTHETIC_ERRNO(EOPNOTSUPP), "libmount support not compiled in");
#endif
}
 
#define MS_CONVERTIBLE_FLAGS (MS_RDONLY|MS_NOSUID|MS_NODEV|MS_NOEXEC|MS_NOSYMFOLLOW|MS_RELATIME|MS_NOATIME|MS_STRICTATIME|MS_NODIRATIME)
 
static uint64_t ms_flags_to_mount_attr(unsigned long a) {
        uint64_t f = 0;
 
        if (FLAGS_SET(a, MS_RDONLY))
                f |= MOUNT_ATTR_RDONLY;
 
        if (FLAGS_SET(a, MS_NOSUID))
                f |= MOUNT_ATTR_NOSUID;
 
        if (FLAGS_SET(a, MS_NODEV))
                f |= MOUNT_ATTR_NODEV;
 
        if (FLAGS_SET(a, MS_NOEXEC))
                f |= MOUNT_ATTR_NOEXEC;
 
        if (FLAGS_SET(a, MS_NOSYMFOLLOW))
                f |= MOUNT_ATTR_NOSYMFOLLOW;
 
        if (FLAGS_SET(a, MS_RELATIME))
                f |= MOUNT_ATTR_RELATIME;
 
        if (FLAGS_SET(a, MS_NOATIME))
                f |= MOUNT_ATTR_NOATIME;
 
        if (FLAGS_SET(a, MS_STRICTATIME))
                f |= MOUNT_ATTR_STRICTATIME;
 
        if (FLAGS_SET(a, MS_NODIRATIME))
                f |= MOUNT_ATTR_NODIRATIME;
 
        return f;
}
 
static uint64_t ms_flags_to_mount_attr_clr(unsigned long a) {
        uint64_t f = 0;
 
        /* As per documentation, if relatime/noatime/strictatime are set, we need to clear the atime flag
         * too, otherwise -EINVAL will be returned by the kernel. */
        if (FLAGS_SET(a, MS_RELATIME))
                f |= MOUNT_ATTR__ATIME;
 
        if (FLAGS_SET(a, MS_NOATIME))
                f |= MOUNT_ATTR__ATIME;
 
        if (FLAGS_SET(a, MS_STRICTATIME))
                f |= MOUNT_ATTR__ATIME;
 
        return f;
}
 
static bool skip_mount_set_attr = false;
 
/* Use this function only if you do not have direct access to /proc/self/mountinfo but the caller can open it
 * for you. This is the case when /proc is masked or not mounted. Otherwise, use bind_remount_recursive. */
int bind_remount_recursive_with_mountinfo(
                const char *prefix,
                unsigned long new_flags,
                unsigned long flags_mask,
                char **deny_list,
                FILE *proc_self_mountinfo) {
 
        assert(prefix);
 
        if ((flags_mask & ~MS_CONVERTIBLE_FLAGS) == 0 && strv_isempty(deny_list) && !skip_mount_set_attr) {
                /* Let's take a shortcut for all the flags we know how to convert into mount_setattr() flags */
 
                if (mount_setattr(AT_FDCWD, prefix, AT_SYMLINK_NOFOLLOW|AT_RECURSIVE,
                                  &(struct mount_attr) {
                                          .attr_set = ms_flags_to_mount_attr(new_flags & flags_mask),
                                          .attr_clr = ms_flags_to_mount_attr(~new_flags & flags_mask),
                                  }, MOUNT_ATTR_SIZE_VER0) < 0) {
 
                        log_debug_errno(errno, "mount_setattr() failed, falling back to classic remounting: %m");
 
                        /* We fall through to classic behaviour if not supported (i.e. kernel < 5.12). We
                         * also do this for all other kinds of errors since they are so many different, and
                         * mount_setattr() has no graceful mode where it continues despite seeing errors one
                         * some mounts, but we want that. Moreover mount_setattr() only works on the mount
                         * point inode itself, not a non-mount point inode, and we want to support arbitrary
                         * prefixes here. */
 
                        if (ERRNO_IS_NOT_SUPPORTED(errno)) /* if not supported, then don't bother at all anymore */
                                skip_mount_set_attr = true;
                } else
                        return 0; /* Nice, this worked! */
        }
 
#if HAVE_LIBMOUNT
        _cleanup_fclose_ FILE *proc_self_mountinfo_opened = NULL;
        _cleanup_set_free_ Set *done = NULL;
        unsigned n_tries = 0;
        int r;
 
        if (!proc_self_mountinfo) {
                r = fopen_unlocked("/proc/self/mountinfo", "re", &proc_self_mountinfo_opened);
                if (r < 0)
                        return r;
 
                proc_self_mountinfo = proc_self_mountinfo_opened;
        }
 
        /* Recursively remount a directory (and all its submounts) with desired flags (MS_READONLY,
         * MS_NOSUID, MS_NOEXEC). If the directory is already mounted, we reuse the mount and simply mark it
         * MS_BIND|MS_RDONLY (or remove the MS_RDONLY for read-write operation), ditto for other flags. If it
         * isn't we first make it one. Afterwards we apply (or remove) the flags to all submounts we can
         * access, too. When mounts are stacked on the same mount point we only care for each individual
         * "top-level" mount on each point, as we cannot influence/access the underlying mounts anyway. We do
         * not have any effect on future submounts that might get propagated, they might be writable
         * etc. This includes future submounts that have been triggered via autofs. Also note that we can't
         * operate atomically here. Mounts established while we process the tree might or might not get
         * noticed and thus might or might not be covered.
         *
         * If the "deny_list" parameter is specified it may contain a list of subtrees to exclude from the
         * remount operation. Note that we'll ignore the deny list for the top-level path. */
 
        for (;;) {
                _cleanup_(mnt_free_tablep) struct libmnt_table *table = NULL;
                _cleanup_(mnt_free_iterp) struct libmnt_iter *iter = NULL;
                _cleanup_hashmap_free_ Hashmap *todo = NULL;
                bool top_autofs = false;
 
                if (n_tries++ >= 32) /* Let's not retry this loop forever */
                        return -EBUSY;
 
                rewind(proc_self_mountinfo);
 
                r = libmount_parse_mountinfo(proc_self_mountinfo, &table, &iter);
                if (r < 0)
                        return log_debug_errno(r, "Failed to parse /proc/self/mountinfo: %m");
 
                for (;;) {
                        _cleanup_free_ char *d = NULL;
                        const char *path, *type, *opts;
                        unsigned long flags = 0;
                        struct libmnt_fs *fs;
 
                        r = sym_mnt_table_next_fs(table, iter, &fs);
                        if (r == 1) /* EOF */
                                break;
                        if (r < 0)
                                return log_debug_errno(r, "Failed to get next entry from /proc/self/mountinfo: %m");
 
                        path = sym_mnt_fs_get_target(fs);
                        if (!path)
                                continue;
 
                        if (!path_startswith(path, prefix))
                                continue;
 
                        type = sym_mnt_fs_get_fstype(fs);
                        if (!type)
                                continue;
 
                        /* Let's ignore autofs mounts. If they aren't triggered yet, we want to avoid
                         * triggering them, as we don't make any guarantees for future submounts anyway. If
                         * they are already triggered, then we will find another entry for this. */
                        if (streq(type, "autofs")) {
                                top_autofs = top_autofs || path_equal(path, prefix);
                                continue;
                        }
 
                        if (set_contains(done, path))
                                continue;
 
                        /* Ignore this mount if it is deny-listed, but only if it isn't the top-level mount
                         * we shall operate on. */
                        if (!path_equal(path, prefix)) {
                                bool deny_listed = false;
 
                                STRV_FOREACH(i, deny_list) {
                                        if (path_equal(*i, prefix))
                                                continue;
 
                                        if (!path_startswith(*i, prefix))
                                                continue;
 
                                        if (path_startswith(path, *i)) {
                                                deny_listed = true;
                                                log_trace("Not remounting %s deny-listed by %s, called for %s", path, *i, prefix);
                                                break;
                                        }
                                }
 
                                if (deny_listed)
                                        continue;
                        }
 
                        opts = sym_mnt_fs_get_vfs_options(fs);
                        if (opts) {
                                r = sym_mnt_optstr_get_flags(opts, &flags, sym_mnt_get_builtin_optmap(MNT_LINUX_MAP));
                                if (r < 0)
                                        log_debug_errno(r, "Could not get flags for '%s', ignoring: %m", path);
                        }
 
                        d = strdup(path);
                        if (!d)
                                return -ENOMEM;
 
                        r = hashmap_ensure_put(&todo, &path_hash_ops_free, d, ULONG_TO_PTR(flags));
                        if (r == -EEXIST)
                                /* If the same path was recorded, but with different mount flags, update it:
                                 * it means a mount point is overmounted, and libmount returns the "bottom" (or
                                 * older one) first, but we want to reapply the flags from the "top" (or newer
                                 * one). See: https://github.com/systemd/systemd/issues/20032
                                 * Note that this shouldn't really fail, as we were just told that the key
                                 * exists, and it's an update so we want 'd' to be freed immediately. */
                                r = hashmap_update(todo, d, ULONG_TO_PTR(flags));
                        if (r < 0)
                                return r;
                        if (r > 0)
                                TAKE_PTR(d);
                }
 
                /* Check if the top-level directory was among what we have seen so far. For that check both
                 * 'done' and 'todo'. Also check 'top_autofs' because if the top-level dir is an autofs we'll
                 * not include it in either set but will set this bool. */
                if (!set_contains(done, prefix) &&
                    !(top_autofs || hashmap_contains(todo, prefix))) {
 
                        /* The prefix directory itself is not yet a mount, make it one. */
                        r = mount_nofollow(prefix, prefix, NULL, MS_BIND|MS_REC, NULL);
                        if (r < 0)
                                return r;
 
                        /* Immediately rescan, so that we pick up the new mount's flags */
                        continue;
                }
 
                /* If we have no submounts to process anymore, we are done */
                if (hashmap_isempty(todo))
                        return 0;
 
                for (;;) {
                        unsigned long flags;
                        char *x = NULL;
 
                        /* Take the first mount from our list of mounts to still process */
                        flags = PTR_TO_ULONG(hashmap_steal_first_key_and_value(todo, (void**) &x));
                        if (!x)
                                break;
 
                        r = set_ensure_consume(&done, &path_hash_ops_free, x);
                        if (IN_SET(r, 0, -EEXIST))
                                continue; /* Already done */
                        if (r < 0)
                                return r;
 
                        /* Now, remount this with the new flags set, but exclude MS_RELATIME from it. (It's
                         * the default anyway, thus redundant, and in userns we'll get an error if we try to
                         * explicitly enable it) */
                        r = mount_nofollow(NULL, x, NULL, ((flags & ~flags_mask)|MS_BIND|MS_REMOUNT|new_flags) & ~MS_RELATIME, NULL);
                        if (r < 0) {
                                int q;
 
                                /* OK, so the remount of this entry failed. We'll ultimately ignore this in
                                 * almost all cases (there are simply so many reasons why this can fail,
                                 * think autofs, NFS, FUSE, …), but let's generate useful debug messages at
                                 * the very least. */
 
                                q = path_is_mount_point(x);
                                if (IN_SET(q, 0, -ENOENT)) {
                                        /* Hmm, whaaaa? The mount point is not actually a mount point? Then
                                         * it is either obstructed by a later mount or somebody has been
                                         * racing against us and removed it. Either way the mount point
                                         * doesn't matter to us, let's ignore it hence. */
                                        log_debug_errno(r, "Mount point '%s' to remount is not a mount point anymore, ignoring remount failure: %m", x);
                                        continue;
                                }
                                if (q < 0) /* Any other error on this? Just log and continue */
                                        log_debug_errno(q, "Failed to determine whether '%s' is a mount point or not, ignoring: %m", x);
 
                                if (((flags ^ new_flags) & flags_mask & ~MS_RELATIME) == 0) { /* ignore MS_RELATIME while comparing */
                                        log_debug_errno(r, "Couldn't remount '%s', but the flags already match what we want, hence ignoring: %m", x);
                                        continue;
                                }
 
                                /* Make this fatal if this is the top-level mount */
                                if (path_equal(x, prefix))
                                        return r;
 
                                /* If this is not the top-level mount, then handle this gracefully: log but
                                 * otherwise ignore. With NFS, FUSE, autofs there are just too many reasons
                                 * this might fail without a chance for us to do anything about it, let's
                                 * hence be strict on the top-level mount and lenient on the inner ones. */
                                log_debug_errno(r, "Couldn't remount submount '%s' for unexpected reason, ignoring: %m", x);
                                continue;
                        }
 
                        log_trace("Remounted %s.", x);
                }
        }
#else
        return log_debug_errno(SYNTHETIC_ERRNO(EOPNOTSUPP), "libmount support not compiled in");
#endif
}
 
int bind_remount_one_with_mountinfo(
                const char *path,
                unsigned long new_flags,
                unsigned long flags_mask,
                FILE *proc_self_mountinfo) {
 
        assert(path);
        assert(proc_self_mountinfo);
 
        if ((flags_mask & ~MS_CONVERTIBLE_FLAGS) == 0 && !skip_mount_set_attr) {
                /* Let's take a shortcut for all the flags we know how to convert into mount_setattr() flags */
 
                if (mount_setattr(AT_FDCWD, path, AT_SYMLINK_NOFOLLOW,
                                  &(struct mount_attr) {
                                          .attr_set = ms_flags_to_mount_attr(new_flags & flags_mask),
                                          .attr_clr = ms_flags_to_mount_attr(~new_flags & flags_mask),
                                  }, MOUNT_ATTR_SIZE_VER0) < 0) {
 
                        log_debug_errno(errno, "mount_setattr() didn't work, falling back to classic remounting: %m");
 
                        if (ERRNO_IS_NOT_SUPPORTED(errno)) /* if not supported, then don't bother at all anymore */
                                skip_mount_set_attr = true;
                } else
                        return 0; /* Nice, this worked! */
        }
 
#if HAVE_LIBMOUNT
        _cleanup_(mnt_free_tablep) struct libmnt_table *table = NULL;
        unsigned long flags = 0;
        struct libmnt_fs *fs;
        const char *opts;
        int r;
 
        rewind(proc_self_mountinfo);
 
        r = dlopen_libmount(LOG_DEBUG);
        if (r < 0)
                return r;
 
        table = sym_mnt_new_table();
        if (!table)
                return -ENOMEM;
 
        r = sym_mnt_table_parse_stream(table, proc_self_mountinfo, "/proc/self/mountinfo");
        if (r < 0)
                return r;
 
        fs = sym_mnt_table_find_target(table, path, MNT_ITER_FORWARD);
        if (!fs) {
                r = access_nofollow(path, F_OK); /* Hmm, it's not in the mount table, but does it exist at all? */
                if (r < 0)
                        return r;
 
                return -EINVAL; /* Not a mount point we recognize */
        }
 
        opts = sym_mnt_fs_get_vfs_options(fs);
        if (opts) {
                r = sym_mnt_optstr_get_flags(opts, &flags, sym_mnt_get_builtin_optmap(MNT_LINUX_MAP));
                if (r < 0)
                        log_debug_errno(r, "Could not get flags for '%s', ignoring: %m", path);
        }
 
        r = mount_nofollow(NULL, path, NULL, ((flags & ~flags_mask)|MS_BIND|MS_REMOUNT|new_flags) & ~MS_RELATIME, NULL);
        if (r < 0) {
                if (((flags ^ new_flags) & flags_mask & ~MS_RELATIME) != 0) /* Ignore MS_RELATIME again,
                                                                             * since kernel adds it in
                                                                             * everywhere, because it's the
                                                                             * default. */
                        return r;
 
                /* Let's handle redundant remounts gracefully */
                log_debug_errno(r, "Failed to remount '%s' but flags already match what we want, ignoring: %m", path);
        }
 
        return 0;
#else
        return log_debug_errno(SYNTHETIC_ERRNO(EOPNOTSUPP), "libmount support not compiled in");
#endif
}
 
int bind_remount_one(const char *path, unsigned long new_flags, unsigned long flags_mask) {
        _cleanup_fclose_ FILE *proc_self_mountinfo = NULL;
 
        proc_self_mountinfo = fopen("/proc/self/mountinfo", "re");
        if (!proc_self_mountinfo)
                return log_debug_errno(errno, "Failed to open %s: %m", "/proc/self/mountinfo");
 
        return bind_remount_one_with_mountinfo(path, new_flags, flags_mask, proc_self_mountinfo);
}
 
static int mount_switch_root_pivot(int fd_newroot, const char *path) {
        assert(fd_newroot >= 0);
        assert(path);
 
        /* Let the kernel tuck the new root under the old one. */
        if (pivot_root(".", ".") < 0)
                return log_debug_errno(errno, "Failed to pivot root to new rootfs '%s': %m", path);
 
        /* Get rid of the old root and reveal our brand new root. (This will always operate on the top-most
         * mount on our cwd, regardless what our current directory actually points to.) */
        if (umount2(".", MNT_DETACH) < 0)
                return log_debug_errno(errno, "Failed to unmount old rootfs: %m");
 
        return 0;
}
 
static int mount_switch_root_move(int fd_newroot, const char *path) {
        assert(fd_newroot >= 0);
        assert(path);
 
        /* Move the new root fs */
        if (mount(".", "/", NULL, MS_MOVE, NULL) < 0)
                return log_debug_errno(errno, "Failed to move new rootfs '%s': %m", path);
 
        /* Also change root dir */
        if (chroot(".") < 0)
                return log_debug_errno(errno, "Failed to chroot to new rootfs '%s': %m", path);
 
        return 0;
}
 
int mount_switch_root_full(const char *path, unsigned long mount_propagation_flag, bool force_ms_move) {
        _cleanup_close_ int fd_newroot = -EBADF;
        int r, is_current_root;
 
        assert(path);
        assert(mount_propagation_flag_is_valid(mount_propagation_flag));
 
        fd_newroot = open(path, O_PATH|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW);
        if (fd_newroot < 0)
                return log_debug_errno(errno, "Failed to open new rootfs '%s': %m", path);
 
        is_current_root = path_is_root_at(fd_newroot, NULL);
        if (is_current_root < 0)
                return log_debug_errno(is_current_root, "Failed to determine if target dir is our root already: %m");
 
        /* Change into the new rootfs. */
        if (fchdir(fd_newroot) < 0)
                return log_debug_errno(errno, "Failed to chdir into new rootfs '%s': %m", path);
 
        /* Make this a NOP if we are supposed to switch to our current root fs. After all, both pivot_root()
         * and MS_MOVE don't like that. */
        if (!is_current_root) {
                if (!force_ms_move) {
                        r = mount_switch_root_pivot(fd_newroot, path);
                        if (r < 0) {
                                log_debug_errno(r, "Failed to pivot into new rootfs '%s', will try to use MS_MOVE instead: %m", path);
                                force_ms_move = true;
                        }
                }
                if (force_ms_move) {
                        /* Failed to pivot_root() fallback to MS_MOVE. For example, this may happen if the rootfs is
                         * an initramfs in which case pivot_root() isn't supported. */
                        r = mount_switch_root_move(fd_newroot, path);
                        if (r < 0)
                                return log_debug_errno(r, "Failed to switch to new rootfs '%s' with MS_MOVE: %m", path);
                }
        }
 
        log_debug("Successfully switched root to '%s'.", path);
 
        /* Finally, let's establish the requested propagation flags. */
        if (mount_propagation_flag == 0)
                return 0;
 
        if (mount(NULL, ".", NULL, mount_propagation_flag | MS_REC, NULL) < 0)
                return log_debug_errno(errno, "Failed to turn new rootfs '%s' into %s mount: %m",
                                       mount_propagation_flag_to_string(mount_propagation_flag), path);
 
        return 0;
}
 
int repeat_unmount(const char *path, int flags) {
        bool done = false;
 
        assert(path);
 
        /* If there are multiple mounts on a mount point, this
         * removes them all */
 
        for (;;) {
                if (umount2(path, flags) < 0) {
 
                        if (errno == EINVAL)
                                return done;
 
                        return -errno;
                }
 
                done = true;
        }
}
 
int mode_to_inaccessible_node(
                const char *runtime_dir,
                mode_t mode,
                char **ret) {
 
        /* This function maps a node type to a corresponding inaccessible file node. These nodes are created
         * during early boot by PID 1. In some cases we lacked the privs to create the character and block
         * devices (maybe because we run in an userns environment, or miss CAP_SYS_MKNOD, or run with a
         * devices policy that excludes device nodes with major and minor of 0), but that's fine, in that
         * case we use an AF_UNIX file node instead, which is not the same, but close enough for most
         * uses. And most importantly, the kernel allows bind mounts from socket nodes to any non-directory
         * file nodes, and that's the most important thing that matters.
         *
         * Note that the runtime directory argument shall be the top-level runtime directory, i.e. /run/ if
         * we operate in system context and $XDG_RUNTIME_DIR if we operate in user context. */
 
        _cleanup_free_ char *d = NULL;
        const char *node;
 
        assert(ret);
 
        if (!runtime_dir)
                runtime_dir = "/run";
 
        if (S_ISLNK(mode))
                return -EINVAL;
 
        node = inode_type_to_string(mode);
        if (!node)
                return -EINVAL;
 
        d = path_join(runtime_dir, "systemd/inaccessible", node);
        if (!d)
                return -ENOMEM;
 
        /* On new kernels unprivileged users are permitted to create 0:0 char device nodes (because they also
         * act as whiteout inode for overlayfs), but no other char or block device nodes. On old kernels no
         * device node whatsoever may be created by unprivileged processes. Hence, if the caller asks for the
         * inaccessible block device node let's see if the block device node actually exists, and if not,
         * fall back to the character device node. From there fall back to the socket device node. This means
         * in the best case we'll get the right device node type — but if not we'll hopefully at least get a
         * device node at all. */
 
        if (S_ISBLK(mode) &&
            access(d, F_OK) < 0 && errno == ENOENT) {
                free(d);
                d = path_join(runtime_dir, "/systemd/inaccessible/chr");
                if (!d)
                        return -ENOMEM;
        }
 
        if (IN_SET(mode & S_IFMT, S_IFBLK, S_IFCHR) &&
            access(d, F_OK) < 0 && errno == ENOENT) {
                free(d);
                d = path_join(runtime_dir, "/systemd/inaccessible/sock");
                if (!d)
                        return -ENOMEM;
        }
 
        *ret = TAKE_PTR(d);
        return 0;
}
 
int mount_flags_to_string(unsigned long flags, char **ret) {
        static const struct {
                unsigned long flag;
                const char *name;
        } map[] = {
                { .flag = MS_RDONLY,      .name = "MS_RDONLY",      },
                { .flag = MS_NOSUID,      .name = "MS_NOSUID",      },
                { .flag = MS_NODEV,       .name = "MS_NODEV",       },
                { .flag = MS_NOEXEC,      .name = "MS_NOEXEC",      },
                { .flag = MS_SYNCHRONOUS, .name = "MS_SYNCHRONOUS", },
                { .flag = MS_REMOUNT,     .name = "MS_REMOUNT",     },
                { .flag = MS_MANDLOCK,    .name = "MS_MANDLOCK",    },
                { .flag = MS_DIRSYNC,     .name = "MS_DIRSYNC",     },
                { .flag = MS_NOSYMFOLLOW, .name = "MS_NOSYMFOLLOW", },
                { .flag = MS_NOATIME,     .name = "MS_NOATIME",     },
                { .flag = MS_NODIRATIME,  .name = "MS_NODIRATIME",  },
                { .flag = MS_BIND,        .name = "MS_BIND",        },
                { .flag = MS_MOVE,        .name = "MS_MOVE",        },
                { .flag = MS_REC,         .name = "MS_REC",         },
                { .flag = MS_SILENT,      .name = "MS_SILENT",      },
                { .flag = MS_POSIXACL,    .name = "MS_POSIXACL",    },
                { .flag = MS_UNBINDABLE,  .name = "MS_UNBINDABLE",  },
                { .flag = MS_PRIVATE,     .name = "MS_PRIVATE",     },
                { .flag = MS_SLAVE,       .name = "MS_SLAVE",       },
                { .flag = MS_SHARED,      .name = "MS_SHARED",      },
                { .flag = MS_RELATIME,    .name = "MS_RELATIME",    },
                { .flag = MS_KERNMOUNT,   .name = "MS_KERNMOUNT",   },
                { .flag = MS_I_VERSION,   .name = "MS_I_VERSION",   },
                { .flag = MS_STRICTATIME, .name = "MS_STRICTATIME", },
                { .flag = MS_LAZYTIME,    .name = "MS_LAZYTIME",    },
        };
        _cleanup_free_ char *str = NULL;
 
        assert(ret);
 
        FOREACH_ELEMENT(entry, map)
                if (flags & entry->flag) {
                        if (!strextend_with_separator(&str, "|", entry->name))
                                return -ENOMEM;
                        flags &= ~entry->flag;
                }
 
        if (!str || flags != 0)
                if (strextendf_with_separator(&str, "|", "%lx", flags) < 0)
                        return -ENOMEM;
 
        *ret = TAKE_PTR(str);
        return 0;
}
 
int mount_verbose_full(
                int error_log_level,
                const char *what,
                const char *where,
                const char *type,
                unsigned long flags,
                const char *options,
                bool follow_symlink) {
 
        _cleanup_free_ char *fl = NULL, *o = NULL;
        unsigned long f;
        int r;
 
        r = mount_option_mangle(options, flags, &f, &o);
        if (r < 0)
                return log_full_errno(error_log_level, r,
                                      "Failed to mangle mount options %s: %m",
                                      strempty(options));
 
        (void) mount_flags_to_string(f, &fl);
 
        if (FLAGS_SET(f, MS_REMOUNT|MS_BIND))
                log_debug("Changing mount flags %s (%s \"%s\")...",
                          where, strnull(fl), strempty(o));
        else if (f & MS_REMOUNT)
                log_debug("Remounting superblock %s (%s \"%s\")...",
                          where, strnull(fl), strempty(o));
        else if (f & (MS_SHARED|MS_PRIVATE|MS_SLAVE|MS_UNBINDABLE))
                log_debug("Changing mount propagation %s (%s \"%s\")",
                          where, strnull(fl), strempty(o));
        else if (f & MS_BIND)
                log_debug("Bind-mounting %s on %s (%s \"%s\")...",
                          what, where, strnull(fl), strempty(o));
        else if (f & MS_MOVE)
                log_debug("Moving mount %s %s %s (%s \"%s\")...",
                          what, glyph(GLYPH_ARROW_RIGHT), where, strnull(fl), strempty(o));
        else
                log_debug("Mounting %s (%s) on %s (%s \"%s\")...",
                          strna(what), strna(type), where, strnull(fl), strempty(o));
 
        if (follow_symlink)
                r = RET_NERRNO(mount(what, where, type, f, o));
        else
                r = mount_nofollow(what, where, type, f, o);
        if (r < 0)
                return log_full_errno(error_log_level, r,
                                      "Failed to mount %s (type %s) on %s (%s \"%s\"): %m",
                                      strna(what), strna(type), where, strnull(fl), strempty(o));
        return 0;
}
 
int umount_verbose(
                int error_log_level,
                const char *where,
                int flags) {
 
        assert(where);
 
        log_debug("Unmounting '%s'...", where);
 
        if (umount2(where, flags) < 0)
                return log_full_errno(error_log_level, errno, "Failed to unmount '%s': %m", where);
 
        return 0;
}
 
int umountat_detach_verbose(
                int error_log_level,
                int fd,
                const char *where) {
 
        /* Similar to umountat_verbose(), but goes by fd + path. This implies MNT_DETACH, since to do this we
         * must pin the inode in question via an fd. */
 
        assert(fd >= 0 || fd == AT_FDCWD);
 
        /* If neither fd nor path are specified take this as reference to the cwd */
        if (fd == AT_FDCWD && isempty(where))
                return umount_verbose(error_log_level, ".", MNT_DETACH|UMOUNT_NOFOLLOW);
 
        /* If we don't actually take the fd into consideration for this operation shortcut things, so that we
         * don't have to open the inode */
        if (fd == AT_FDCWD || path_is_absolute(where))
                return umount_verbose(error_log_level, where, MNT_DETACH|UMOUNT_NOFOLLOW);
 
        _cleanup_free_ char *prefix = NULL;
        const char *p;
        if (fd_get_path(fd, &prefix) < 0)
                p = "<fd>"; /* if we can't get the path, return something vaguely useful */
        else
                p = prefix;
        _cleanup_free_ char *joined = isempty(where) ? strdup(p) : path_join(p, where);
 
        log_debug("Unmounting '%s'...", strna(joined));
 
        _cleanup_close_ int inode_fd = -EBADF;
        int mnt_fd;
        if (isempty(where))
                mnt_fd = fd;
        else {
                inode_fd = openat(fd, where, O_PATH|O_CLOEXEC|O_NOFOLLOW);
                if (inode_fd < 0)
                        return log_full_errno(error_log_level, errno, "Failed to pin '%s': %m", strna(joined));
 
                mnt_fd = inode_fd;
        }
 
        if (umount2(FORMAT_PROC_FD_PATH(mnt_fd), MNT_DETACH) < 0)
                return log_full_errno(error_log_level, errno, "Failed to unmount '%s': %m", strna(joined));
 
        return 0;
}
 
int mount_exchange_graceful(int fsmount_fd, const char *dest, bool mount_beneath) {
        int r;
 
        assert(fsmount_fd >= 0);
        assert(dest);
 
        /* First, try to mount beneath an existing mount point, and if that works, umount the old mount,
         * which is now at the top. This will ensure we can atomically replace a mount. Note that this works
         * also in the case where there are submounts down the tree. Mount propagation is allowed but
         * restricted to layouts that don't end up propagation the new mount on top of the mount stack.  If
         * this is not supported (minimum kernel v6.5), or if there is no mount on the mountpoint, we get
         * -EINVAL and then we fallback to normal mounting. */
 
        r = RET_NERRNO(move_mount(fsmount_fd, /* from_path= */ "",
                                  /* to_fd= */ -EBADF, dest,
                                  MOVE_MOUNT_F_EMPTY_PATH | (mount_beneath ? MOVE_MOUNT_BENEATH : 0)));
        if (mount_beneath) {
                if (r >= 0) /* Mounting beneath worked! Now unmount the upper mount. */
                        return umount_verbose(LOG_DEBUG, dest, UMOUNT_NOFOLLOW|MNT_DETACH);
 
                if (r == -EINVAL) { /* Fallback if mount_beneath is not supported */
                        log_debug_errno(r,
                                        "Cannot mount beneath '%s', falling back to overmount: %m",
                                        dest);
                        return mount_exchange_graceful(fsmount_fd, dest, /* mount_beneath= */ false);
                }
        }
 
        return r;
}
 
int mount_option_mangle(
                const char *options,
                unsigned long mount_flags,
                unsigned long *ret_mount_flags,
                char **ret_remaining_options) {
 
        assert(ret_mount_flags);
        assert(ret_remaining_options);
 
        if (!options) {
                *ret_mount_flags = mount_flags;
                *ret_remaining_options = NULL;
                return 0;
        }
 
#if HAVE_LIBMOUNT
        const struct libmnt_optmap *map;
        _cleanup_free_ char *ret = NULL;
        int r;
 
        /* This extracts mount flags from the mount options, and stores
         * non-mount-flag options to '*ret_remaining_options'.
         * E.g.,
         * "rw,nosuid,nodev,relatime,size=1630748k,mode=0700,uid=1000,gid=1000"
         * is split to MS_NOSUID|MS_NODEV|MS_RELATIME and
         * "size=1630748k,mode=0700,uid=1000,gid=1000".
         * See more examples in test-mount-util.c.
         *
         * If 'options' does not contain any non-mount-flag options,
         * then '*ret_remaining_options' is set to NULL instead of empty string.
         * The validity of options stored in '*ret_remaining_options' is not checked.
         * If 'options' is NULL, this just copies 'mount_flags' to *ret_mount_flags. */
 
        r = dlopen_libmount(LOG_DEBUG);
        if (r < 0)
                return r;
 
        map = sym_mnt_get_builtin_optmap(MNT_LINUX_MAP);
        if (!map)
                return -EINVAL;
 
        for (const char *p = options;;) {
                _cleanup_free_ char *word = NULL;
                const struct libmnt_optmap *ent;
 
                r = extract_first_word(&p, &word, ",", EXTRACT_KEEP_QUOTE);
                if (r < 0)
                        return r;
                if (r == 0)
                        break;
 
                for (ent = map; ent->name; ent++) {
                        /* All entries in MNT_LINUX_MAP do not take any argument.
                         * Thus, ent->name does not contain "=" or "[=]". */
                        if (!streq(word, ent->name))
                                continue;
 
                        if (!(ent->mask & MNT_INVERT))
                                mount_flags |= ent->id;
                        else
                                mount_flags &= ~ent->id;
 
                        break;
                }
 
                /* If 'word' is not a mount flag, then store it in '*ret_remaining_options'. */
                if (!ent->name &&
                    !startswith_no_case(word, "x-") &&
                    !strextend_with_separator(&ret, ",", word))
                        return -ENOMEM;
        }
 
        *ret_mount_flags = mount_flags;
        *ret_remaining_options = TAKE_PTR(ret);
 
        return 0;
#else
        return log_debug_errno(SYNTHETIC_ERRNO(EOPNOTSUPP), "libmount support not compiled in");
#endif
}
 
static int mount_in_namespace_legacy(
                const char *chased_src_path,
                int chased_src_fd,
                struct stat *chased_src_st,
                const char *propagate_path,
                const char *incoming_path,
                const char *dest,
                int pidns_fd,
                int mntns_fd,
                int root_fd,
                MountInNamespaceFlags flags,
                const MountOptions *options,
                const ImagePolicy *image_policy) {
 
        _cleanup_close_pair_ int errno_pipe_fd[2] = EBADF_PAIR;
        char mount_slave[] = "/tmp/propagate.XXXXXX", *mount_tmp, *mount_outside, *p;
        bool mount_slave_created = false, mount_slave_mounted = false,
                mount_tmp_created = false, mount_tmp_mounted = false,
                mount_outside_created = false, mount_outside_mounted = false;
        _cleanup_(pidref_done) PidRef child = PIDREF_NULL;
        int r;
 
        assert(chased_src_path);
        assert(chased_src_fd >= 0);
        assert(chased_src_st);
        assert(propagate_path);
        assert(incoming_path);
        assert(dest);
        assert(pidns_fd >= 0);
        assert(mntns_fd >= 0);
        assert(root_fd >= 0);