[thirdparty/systemd.git] / src / shared / switch-root.c

/* SPDX-License-Identifier: LGPL-2.1-or-later */

#include <fcntl.h>
#include <sys/mount.h>
#include <sys/stat.h>
#include <unistd.h>

#include "alloc-util.h"
#include "base-filesystem.h"
#include "chase.h"
#include "errno-util.h"
#include "fd-util.h"
#include "log.h"
#include "missing_syscall.h"
#include "mkdir-label.h"
#include "mount-util.h"
#include "mountpoint-util.h"
#include "rm-rf.h"
#include "stat-util.h"
#include "switch-root.h"

int switch_root(const char *new_root,
                const char *old_root_after,   /* path below the new root, where to place the old root after the transition; may be NULL to unmount it */
                SwitchRootFlags flags) {

        /* Stuff mounted below /run/ we don't save on soft reboot, as it might have lost its relevance,
         * e.g. removable media and such. We rather want that the new boot mounts this fresh. But on
         * the switch from initrd we do use MS_REC, as it is expected that mounts set up in /run/ are
         * maintained. */
        static const struct {
                const char *path;
                unsigned long mount_flags;                 /* Flags to apply if SWITCH_ROOT_RECURSIVE_RUN is unset */
                unsigned long mount_flags_recursive_run;   /* Flags to apply if SWITCH_ROOT_RECURSIVE_RUN is set (0 if shall be skipped) */
        } transfer_table[] = {
                { "/dev",             MS_BIND|MS_REC,  MS_BIND|MS_REC }, /* Recursive, because we want to save the original /dev/shm/ + /dev/pts/ and similar */
                { "/sys",             MS_BIND|MS_REC,  MS_BIND|MS_REC }, /* Similar, we want to retain various API VFS, or the cgroupv1 /sys/fs/cgroup/ tree */
                { "/proc",            MS_BIND|MS_REC,  MS_BIND|MS_REC }, /* Similar */
                { "/run",             MS_BIND,         MS_BIND|MS_REC }, /* Recursive except on soft reboot, see above */
                { "/run/credentials", MS_BIND|MS_REC,  0 /* skip! */  }, /* Credential mounts should survive */
                { "/run/host",        MS_BIND|MS_REC,  0 /* skip! */  }, /* Host supplied hierarchy should also survive */
        };

        _cleanup_close_ int old_root_fd = -EBADF, new_root_fd = -EBADF;
        _cleanup_free_ char *resolved_old_root_after = NULL;
        int r, istmp;

        assert(new_root);

        /* Check if we shall remove the contents of the old root */
        old_root_fd = open("/", O_DIRECTORY|O_CLOEXEC);
        if (old_root_fd < 0)
                return log_error_errno(errno, "Failed to open root directory: %m");

        new_root_fd = open(new_root, O_PATH|O_DIRECTORY|O_CLOEXEC);
        if (new_root_fd < 0)
                return log_error_errno(errno, "Failed to open target directory '%s': %m", new_root);

        r = fds_are_same_mount(old_root_fd, new_root_fd);
        if (r < 0)
                return log_error_errno(r, "Failed to check if old and new root directory/mount are the same: %m");
        if (r > 0) {
                log_debug("Skipping switch root, as old and new root directories/mounts are the same.");
                return 0;
        }

        /* Make the new root directory a mount point if it isn't */
        r = fd_make_mount_point(new_root_fd);
        if (r < 0)
                return log_error_errno(r, "Failed to make new root directory a mount point: %m");
        if (r > 0) {
                int fd;

                /* When the path was not a mount point, then we need to reopen the path, otherwise, it still
                 * points to the underlying directory. */

                fd = open(new_root, O_DIRECTORY|O_CLOEXEC);
                if (fd < 0)
                        return log_error_errno(errno, "Failed to reopen target directory '%s': %m", new_root);

                close_and_replace(new_root_fd, fd);
        }

        if (FLAGS_SET(flags, SWITCH_ROOT_DESTROY_OLD_ROOT)) {
                istmp = fd_is_temporary_fs(old_root_fd);
                if (istmp < 0)
                        return log_error_errno(istmp, "Failed to stat root directory: %m");
                if (istmp > 0)
                        log_debug("Root directory is on tmpfs, will do cleanup later.");
        } else
                istmp = -1; /* don't know */

        if (old_root_after) {
                /* Determine where we shall place the old root after the transition */
                r = chase(old_root_after, new_root, CHASE_PREFIX_ROOT|CHASE_NONEXISTENT, &resolved_old_root_after, NULL);
                if (r < 0)
                        return log_error_errno(r, "Failed to resolve %s/%s: %m", new_root, old_root_after);
                if (r == 0) /* Doesn't exist yet. Let's create it */
                        (void) mkdir_p_label(resolved_old_root_after, 0755);
        }

        /* We are about to unmount various file systems with MNT_DETACH (either explicitly via umount() or
         * indirectly via pivot_root()), and thus do not synchronously wait for them to be fully sync'ed —
         * all while making them invisible/inaccessible in the file system tree for later code. That makes
         * sync'ing them then difficult. Let's hence issue a manual sync() here, so that we at least can
         * guarantee all file systems are an a good state before entering this state. */
        if (!FLAGS_SET(flags, SWITCH_ROOT_DONT_SYNC))
                sync();

        /* Work-around for kernel design: the kernel refuses MS_MOVE if any file systems are mounted
         * MS_SHARED. Hence remount them MS_PRIVATE here as a work-around.
         *
         * https://bugzilla.redhat.com/show_bug.cgi?id=847418 */
        if (mount(NULL, "/", NULL, MS_REC|MS_PRIVATE, NULL) < 0)
                return log_error_errno(errno, "Failed to set \"/\" mount propagation to private: %m");

        /* Do not fail if base_filesystem_create() fails. Not all switch roots are like base_filesystem_create() wants
         * them to look like. They might even boot, if they are RO and don't have the FS layout. Just ignore the error
         * and switch_root() nevertheless. */
        (void) base_filesystem_create_fd(new_root_fd, new_root, UID_INVALID, GID_INVALID);

        FOREACH_ELEMENT(transfer, transfer_table) {
                _cleanup_free_ char *chased = NULL;
                unsigned long mount_flags;

                mount_flags = FLAGS_SET(flags, SWITCH_ROOT_RECURSIVE_RUN) ? transfer->mount_flags_recursive_run : transfer->mount_flags;
                if (mount_flags == 0) /* skip if zero */
                        continue;

                if (access(transfer->path, F_OK) < 0) {
                        log_debug_errno(errno, "Path '%s' to move to target root directory, not found, ignoring: %m", transfer->path);
                        continue;
                }

                r = chase(transfer->path, new_root, CHASE_PREFIX_ROOT, &chased, NULL);
                if (r < 0)
                        return log_error_errno(r, "Failed to resolve %s/%s: %m", new_root, transfer->path);

                /* Let's see if it is a mount point already. */
                r = path_is_mount_point(chased);
                if (r < 0)
                        return log_error_errno(r, "Failed to determine whether %s is a mount point: %m", chased);
                if (r > 0) /* If it is already mounted, then do nothing */
                        continue;

                r = mount_nofollow_verbose(LOG_ERR, transfer->path, chased, NULL, mount_flags, NULL);
                if (r < 0)
                        return r;
        }

        if (fchdir(new_root_fd) < 0)
                return log_error_errno(errno, "Failed to change directory to %s: %m", new_root);

        /* We first try a pivot_root() so that we can umount the old root dir. In many cases (i.e. where rootfs is /),
         * that's not possible however, and hence we simply overmount root */
        if (resolved_old_root_after)
                r = RET_NERRNO(pivot_root(".", resolved_old_root_after));
        else {
                r = RET_NERRNO(pivot_root(".", "."));
                if (r >= 0) {
                        /* Now unmount the upper of the two stacked file systems */
                        if (umount2(".", MNT_DETACH) < 0)
                                return log_error_errno(errno, "Failed to unmount the old root: %m");
                }
        }
        if (r < 0) {
                log_debug_errno(r, "Pivoting root file system failed, moving mounts instead: %m");

                if (resolved_old_root_after) {
                        r = mount_nofollow_verbose(LOG_ERR, "/", resolved_old_root_after, NULL, MS_BIND|MS_REC, NULL);
                        if (r < 0)
                                return r;
                }

                /* If we have to use MS_MOVE let's first try to get rid of *all* mounts we can, with the
                 * exception of the path we want to switch to, plus everything leading to it and within
                 * it. This is necessary because unlike pivot_root() just moving the mount to the root via
                 * MS_MOVE won't magically unmount anything below it. Once the chroot() succeeds the mounts
                 * below would still be around but invisible to us, because not accessible via
                 * /proc/self/mountinfo. Hence, let's clean everything up first, as long as we still can. */
                (void) umount_recursive_full(NULL, MNT_DETACH, STRV_MAKE(new_root));

                if (mount(".", "/", NULL, MS_MOVE, NULL) < 0)
                        return log_error_errno(errno, "Failed to move %s to /: %m", new_root);

                if (chroot(".") < 0)
                        return log_error_errno(errno, "Failed to change root: %m");

                if (chdir(".") < 0)
                        return log_error_errno(errno, "Failed to change directory: %m");
        }

        if (istmp > 0) {
                struct stat rb;

                if (fstat(old_root_fd, &rb) < 0)
                        return log_error_errno(errno, "Failed to stat old root directory: %m");

                /* Note: the below won't operate on non-memory file systems (i.e. only on tmpfs, ramfs), and
                 * it will stop at mount boundaries */
                (void) rm_rf_children(TAKE_FD(old_root_fd), 0, &rb); /* takes possession of the dir fd, even on failure */
        }

        return 0;
}
Commit	Line	Data
	1	/* SPDX-License-Identifier: LGPL-2.1-or-later */
	2
	3	#include <fcntl.h>
	4	#include <sys/mount.h>
	5	#include <sys/stat.h>
	6	#include <unistd.h>
	7
	8	#include "alloc-util.h"
	9	#include "base-filesystem.h"
	10	#include "chase.h"
	11	#include "errno-util.h"
	12	#include "fd-util.h"
	13	#include "log.h"
	14	#include "missing_syscall.h"
	15	#include "mkdir-label.h"
	16	#include "mount-util.h"
	17	#include "mountpoint-util.h"
	18	#include "rm-rf.h"
	19	#include "stat-util.h"
	20	#include "switch-root.h"
	21
	22	int switch_root(const char *new_root,
	23	const char old_root_after, / path below the new root, where to place the old root after the transition; may be NULL to unmount it */
	24	SwitchRootFlags flags) {
	25
	26	/* Stuff mounted below /run/ we don't save on soft reboot, as it might have lost its relevance,
	27	* e.g. removable media and such. We rather want that the new boot mounts this fresh. But on
	28	* the switch from initrd we do use MS_REC, as it is expected that mounts set up in /run/ are
	29	* maintained. */
	30	static const struct {
	31	const char *path;
	32	unsigned long mount_flags; /* Flags to apply if SWITCH_ROOT_RECURSIVE_RUN is unset */
	33	unsigned long mount_flags_recursive_run; /* Flags to apply if SWITCH_ROOT_RECURSIVE_RUN is set (0 if shall be skipped) */
	34	} transfer_table[] = {
	35	{ "/dev", MS_BIND\|MS_REC, MS_BIND\|MS_REC }, /* Recursive, because we want to save the original /dev/shm/ + /dev/pts/ and similar */
	36	{ "/sys", MS_BIND\|MS_REC, MS_BIND\|MS_REC }, /* Similar, we want to retain various API VFS, or the cgroupv1 /sys/fs/cgroup/ tree */
	37	{ "/proc", MS_BIND\|MS_REC, MS_BIND\|MS_REC }, /* Similar */
	38	{ "/run", MS_BIND, MS_BIND\|MS_REC }, /* Recursive except on soft reboot, see above */
	39	{ "/run/credentials", MS_BIND\|MS_REC, 0 /* skip! / }, / Credential mounts should survive */
	40	{ "/run/host", MS_BIND\|MS_REC, 0 /* skip! / }, / Host supplied hierarchy should also survive */
	41	};
	42
	43	_cleanup_close_ int old_root_fd = -EBADF, new_root_fd = -EBADF;
	44	_cleanup_free_ char *resolved_old_root_after = NULL;
	45	int r, istmp;
	46
	47	assert(new_root);
	48
	49	/* Check if we shall remove the contents of the old root */
	50	old_root_fd = open("/", O_DIRECTORY\|O_CLOEXEC);
	51	if (old_root_fd < 0)
	52	return log_error_errno(errno, "Failed to open root directory: %m");
	53
	54	new_root_fd = open(new_root, O_PATH\|O_DIRECTORY\|O_CLOEXEC);
	55	if (new_root_fd < 0)
	56	return log_error_errno(errno, "Failed to open target directory '%s': %m", new_root);
	57
	58	r = fds_are_same_mount(old_root_fd, new_root_fd);
	59	if (r < 0)
	60	return log_error_errno(r, "Failed to check if old and new root directory/mount are the same: %m");
	61	if (r > 0) {
	62	log_debug("Skipping switch root, as old and new root directories/mounts are the same.");
	63	return 0;
	64	}
	65
	66	/* Make the new root directory a mount point if it isn't */
	67	r = fd_make_mount_point(new_root_fd);
	68	if (r < 0)
	69	return log_error_errno(r, "Failed to make new root directory a mount point: %m");
	70	if (r > 0) {
	71	int fd;
	72
	73	/* When the path was not a mount point, then we need to reopen the path, otherwise, it still
	74	* points to the underlying directory. */
	75
	76	fd = open(new_root, O_DIRECTORY\|O_CLOEXEC);
	77	if (fd < 0)
	78	return log_error_errno(errno, "Failed to reopen target directory '%s': %m", new_root);
	79
	80	close_and_replace(new_root_fd, fd);
	81	}
	82
	83	if (FLAGS_SET(flags, SWITCH_ROOT_DESTROY_OLD_ROOT)) {
	84	istmp = fd_is_temporary_fs(old_root_fd);
	85	if (istmp < 0)
	86	return log_error_errno(istmp, "Failed to stat root directory: %m");
	87	if (istmp > 0)
	88	log_debug("Root directory is on tmpfs, will do cleanup later.");
	89	} else
	90	istmp = -1; /* don't know */
	91
	92	if (old_root_after) {
	93	/* Determine where we shall place the old root after the transition */
	94	r = chase(old_root_after, new_root, CHASE_PREFIX_ROOT\|CHASE_NONEXISTENT, &resolved_old_root_after, NULL);
	95	if (r < 0)
	96	return log_error_errno(r, "Failed to resolve %s/%s: %m", new_root, old_root_after);
	97	if (r == 0) /* Doesn't exist yet. Let's create it */
	98	(void) mkdir_p_label(resolved_old_root_after, 0755);
	99	}
	100
	101	/* We are about to unmount various file systems with MNT_DETACH (either explicitly via umount() or
	102	* indirectly via pivot_root()), and thus do not synchronously wait for them to be fully sync'ed —
	103	* all while making them invisible/inaccessible in the file system tree for later code. That makes
	104	* sync'ing them then difficult. Let's hence issue a manual sync() here, so that we at least can
	105	* guarantee all file systems are an a good state before entering this state. */
	106	if (!FLAGS_SET(flags, SWITCH_ROOT_DONT_SYNC))
	107	sync();
	108
	109	/* Work-around for kernel design: the kernel refuses MS_MOVE if any file systems are mounted
	110	* MS_SHARED. Hence remount them MS_PRIVATE here as a work-around.
	111	*
	112	* https://bugzilla.redhat.com/show_bug.cgi?id=847418 */
	113	if (mount(NULL, "/", NULL, MS_REC\|MS_PRIVATE, NULL) < 0)
	114	return log_error_errno(errno, "Failed to set \"/\" mount propagation to private: %m");
	115
	116	/* Do not fail if base_filesystem_create() fails. Not all switch roots are like base_filesystem_create() wants
	117	* them to look like. They might even boot, if they are RO and don't have the FS layout. Just ignore the error
	118	* and switch_root() nevertheless. */
	119	(void) base_filesystem_create_fd(new_root_fd, new_root, UID_INVALID, GID_INVALID);
	120
	121	FOREACH_ELEMENT(transfer, transfer_table) {
	122	_cleanup_free_ char *chased = NULL;
	123	unsigned long mount_flags;
	124
	125	mount_flags = FLAGS_SET(flags, SWITCH_ROOT_RECURSIVE_RUN) ? transfer->mount_flags_recursive_run : transfer->mount_flags;
	126	if (mount_flags == 0) /* skip if zero */
	127	continue;
	128
	129	if (access(transfer->path, F_OK) < 0) {
	130	log_debug_errno(errno, "Path '%s' to move to target root directory, not found, ignoring: %m", transfer->path);
	131	continue;
	132	}
	133
	134	r = chase(transfer->path, new_root, CHASE_PREFIX_ROOT, &chased, NULL);
	135	if (r < 0)
	136	return log_error_errno(r, "Failed to resolve %s/%s: %m", new_root, transfer->path);
	137
	138	/* Let's see if it is a mount point already. */
	139	r = path_is_mount_point(chased);
	140	if (r < 0)
	141	return log_error_errno(r, "Failed to determine whether %s is a mount point: %m", chased);
	142	if (r > 0) /* If it is already mounted, then do nothing */
	143	continue;
	144
	145	r = mount_nofollow_verbose(LOG_ERR, transfer->path, chased, NULL, mount_flags, NULL);
	146	if (r < 0)
	147	return r;
	148	}
	149
	150	if (fchdir(new_root_fd) < 0)
	151	return log_error_errno(errno, "Failed to change directory to %s: %m", new_root);
	152
	153	/* We first try a pivot_root() so that we can umount the old root dir. In many cases (i.e. where rootfs is /),
	154	* that's not possible however, and hence we simply overmount root */
	155	if (resolved_old_root_after)
	156	r = RET_NERRNO(pivot_root(".", resolved_old_root_after));
	157	else {
	158	r = RET_NERRNO(pivot_root(".", "."));
	159	if (r >= 0) {
	160	/* Now unmount the upper of the two stacked file systems */
	161	if (umount2(".", MNT_DETACH) < 0)
	162	return log_error_errno(errno, "Failed to unmount the old root: %m");
	163	}
	164	}
	165	if (r < 0) {
	166	log_debug_errno(r, "Pivoting root file system failed, moving mounts instead: %m");
	167
	168	if (resolved_old_root_after) {
	169	r = mount_nofollow_verbose(LOG_ERR, "/", resolved_old_root_after, NULL, MS_BIND\|MS_REC, NULL);
	170	if (r < 0)
	171	return r;
	172	}
	173
	174	/* If we have to use MS_MOVE let's first try to get rid of all mounts we can, with the
	175	* exception of the path we want to switch to, plus everything leading to it and within
	176	* it. This is necessary because unlike pivot_root() just moving the mount to the root via
	177	* MS_MOVE won't magically unmount anything below it. Once the chroot() succeeds the mounts
	178	* below would still be around but invisible to us, because not accessible via
	179	* /proc/self/mountinfo. Hence, let's clean everything up first, as long as we still can. */
	180	(void) umount_recursive_full(NULL, MNT_DETACH, STRV_MAKE(new_root));
	181
	182	if (mount(".", "/", NULL, MS_MOVE, NULL) < 0)
	183	return log_error_errno(errno, "Failed to move %s to /: %m", new_root);
	184
	185	if (chroot(".") < 0)
	186	return log_error_errno(errno, "Failed to change root: %m");
	187
	188	if (chdir(".") < 0)
	189	return log_error_errno(errno, "Failed to change directory: %m");
	190	}
	191
	192	if (istmp > 0) {
	193	struct stat rb;
	194
	195	if (fstat(old_root_fd, &rb) < 0)
	196	return log_error_errno(errno, "Failed to stat old root directory: %m");
	197
	198	/* Note: the below won't operate on non-memory file systems (i.e. only on tmpfs, ramfs), and
	199	* it will stop at mount boundaries */
	200	(void) rm_rf_children(TAKE_FD(old_root_fd), 0, &rb); /* takes possession of the dir fd, even on failure */
	201	}
	202
	203	return 0;
	204	}