1 /* SPDX-License-Identifier: LGPL-2.1+ */
3 This file is part of systemd.
5 Copyright 2010 Lennart Poettering
9 //#include <stdio_ext.h>
12 #include <sys/mount.h>
14 #include <sys/statvfs.h>
18 //#include <libmount.h>
20 #include "alloc-util.h"
22 //#include "extract-word.h"
27 #include "mount-util.h"
28 #include "parse-util.h"
29 #include "path-util.h"
31 #include "stdio-util.h"
32 #include "string-util.h"
35 /* This is the original MAX_HANDLE_SZ definition from the kernel, when the API was introduced. We use that in place of
36 * any more currently defined value to future-proof things: if the size is increased in the API headers, and our code
37 * is recompiled then it would cease working on old kernels, as those refuse any sizes larger than this value with
38 * EINVAL right-away. Hence, let's disconnect ourselves from any such API changes, and stick to the original definition
39 * from when it was introduced. We use it as a start value only anyway (see below), and hence should be able to deal
40 * with large file handles anyway. */
41 #define ORIGINAL_MAX_HANDLE_SZ 128
43 int name_to_handle_at_loop(
46 struct file_handle **ret_handle,
50 _cleanup_free_ struct file_handle *h = NULL;
51 size_t n = ORIGINAL_MAX_HANDLE_SZ;
53 /* We need to invoke name_to_handle_at() in a loop, given that it might return EOVERFLOW when the specified
54 * buffer is too small. Note that in contrast to what the docs might suggest, MAX_HANDLE_SZ is only good as a
55 * start value, it is not an upper bound on the buffer size required.
57 * This improves on raw name_to_handle_at() also in one other regard: ret_handle and ret_mnt_id can be passed
58 * as NULL if there's no interest in either. */
63 h = malloc0(offsetof(struct file_handle, f_handle) + n);
69 if (name_to_handle_at(fd, path, h, &mnt_id, flags) >= 0) {
72 *ret_handle = TAKE_PTR(h);
79 if (errno != EOVERFLOW)
82 if (!ret_handle && ret_mnt_id && mnt_id >= 0) {
84 /* As it appears, name_to_handle_at() fills in mnt_id even when it returns EOVERFLOW when the
85 * buffer is too small, but that's undocumented. Hence, let's make use of this if it appears to
86 * be filled in, and the caller was interested in only the mount ID an nothing else. */
92 /* If name_to_handle_at() didn't increase the byte size, then this EOVERFLOW is caused by something
93 * else (apparently EOVERFLOW is returned for untriggered nfs4 mounts sometimes), not by the too small
94 * buffer. In that case propagate EOVERFLOW */
95 if (h->handle_bytes <= n)
98 /* The buffer was too small. Size the new buffer by what name_to_handle_at() returned. */
100 if (offsetof(struct file_handle, f_handle) + n < n) /* check for addition overflow */
107 static int fd_fdinfo_mnt_id(int fd, const char *filename, int flags, int *mnt_id) {
108 char path[STRLEN("/proc/self/fdinfo/") + DECIMAL_STR_MAX(int)];
109 _cleanup_free_ char *fdinfo = NULL;
110 _cleanup_close_ int subfd = -1;
114 if ((flags & AT_EMPTY_PATH) && isempty(filename))
115 xsprintf(path, "/proc/self/fdinfo/%i", fd);
117 subfd = openat(fd, filename, O_CLOEXEC|O_PATH);
121 xsprintf(path, "/proc/self/fdinfo/%i", subfd);
124 r = read_full_file(path, &fdinfo, NULL);
125 if (r == -ENOENT) /* The fdinfo directory is a relatively new addition */
130 p = startswith(fdinfo, "mnt_id:");
132 p = strstr(fdinfo, "\nmnt_id:");
133 if (!p) /* The mnt_id field is a relatively new addition */
139 p += strspn(p, WHITESPACE);
140 p[strcspn(p, WHITESPACE)] = 0;
142 return safe_atoi(p, mnt_id);
145 int fd_is_mount_point(int fd, const char *filename, int flags) {
146 _cleanup_free_ struct file_handle *h = NULL, *h_parent = NULL;
147 int mount_id = -1, mount_id_parent = -1;
148 bool nosupp = false, check_st_dev = true;
155 /* First we will try the name_to_handle_at() syscall, which
156 * tells us the mount id and an opaque file "handle". It is
157 * not supported everywhere though (kernel compile-time
158 * option, not all file systems are hooked up). If it works
159 * the mount id is usually good enough to tell us whether
160 * something is a mount point.
162 * If that didn't work we will try to read the mount id from
163 * /proc/self/fdinfo/<fd>. This is almost as good as
164 * name_to_handle_at(), however, does not return the
165 * opaque file handle. The opaque file handle is pretty useful
166 * to detect the root directory, which we should always
167 * consider a mount point. Hence we use this only as
168 * fallback. Exporting the mnt_id in fdinfo is a pretty recent
171 * As last fallback we do traditional fstat() based st_dev
172 * comparisons. This is how things were traditionally done,
173 * but unionfs breaks this since it exposes file
174 * systems with a variety of st_dev reported. Also, btrfs
175 * subvolumes have different st_dev, even though they aren't
176 * real mounts of their own. */
178 r = name_to_handle_at_loop(fd, filename, &h, &mount_id, flags);
179 if (IN_SET(r, -ENOSYS, -EACCES, -EPERM, -EOVERFLOW, -EINVAL))
180 /* This kernel does not support name_to_handle_at() at all (ENOSYS), or the syscall was blocked
181 * (EACCES/EPERM; maybe through seccomp, because we are running inside of a container?), or the mount
182 * point is not triggered yet (EOVERFLOW, think nfs4), or some general name_to_handle_at() flakiness
183 * (EINVAL): fall back to simpler logic. */
184 goto fallback_fdinfo;
185 else if (r == -EOPNOTSUPP)
186 /* This kernel or file system does not support name_to_handle_at(), hence let's see if the upper fs
187 * supports it (in which case it is a mount point), otherwise fallback to the traditional stat()
193 r = name_to_handle_at_loop(fd, "", &h_parent, &mount_id_parent, AT_EMPTY_PATH);
194 if (r == -EOPNOTSUPP) {
196 /* Neither parent nor child do name_to_handle_at()? We have no choice but to fall back. */
197 goto fallback_fdinfo;
199 /* The parent can't do name_to_handle_at() but the directory we are interested in can? If so,
200 * it must be a mount point. */
205 /* The parent can do name_to_handle_at() but the
206 * directory we are interested in can't? If so, it
207 * must be a mount point. */
211 /* If the file handle for the directory we are
212 * interested in and its parent are identical, we
213 * assume this is the root directory, which is a mount
216 if (h->handle_bytes == h_parent->handle_bytes &&
217 h->handle_type == h_parent->handle_type &&
218 memcmp(h->f_handle, h_parent->f_handle, h->handle_bytes) == 0)
221 return mount_id != mount_id_parent;
224 r = fd_fdinfo_mnt_id(fd, filename, flags, &mount_id);
225 if (IN_SET(r, -EOPNOTSUPP, -EACCES, -EPERM))
230 r = fd_fdinfo_mnt_id(fd, "", AT_EMPTY_PATH, &mount_id_parent);
234 if (mount_id != mount_id_parent)
237 /* Hmm, so, the mount ids are the same. This leaves one
238 * special case though for the root file system. For that,
239 * let's see if the parent directory has the same inode as we
240 * are interested in. Hence, let's also do fstat() checks now,
241 * too, but avoid the st_dev comparisons, since they aren't
242 * that useful on unionfs mounts. */
243 check_st_dev = false;
246 /* yay for fstatat() taking a different set of flags than the other
248 if (flags & AT_SYMLINK_FOLLOW)
249 flags &= ~AT_SYMLINK_FOLLOW;
251 flags |= AT_SYMLINK_NOFOLLOW;
252 if (fstatat(fd, filename, &a, flags) < 0)
255 if (fstatat(fd, "", &b, AT_EMPTY_PATH) < 0)
258 /* A directory with same device and inode as its parent? Must
259 * be the root directory */
260 if (a.st_dev == b.st_dev &&
261 a.st_ino == b.st_ino)
264 return check_st_dev && (a.st_dev != b.st_dev);
267 /* flags can be AT_SYMLINK_FOLLOW or 0 */
268 int path_is_mount_point(const char *t, const char *root, int flags) {
269 _cleanup_free_ char *canonical = NULL, *parent = NULL;
270 _cleanup_close_ int fd = -1;
274 assert((flags & ~AT_SYMLINK_FOLLOW) == 0);
276 if (path_equal(t, "/"))
279 /* we need to resolve symlinks manually, we can't just rely on
280 * fd_is_mount_point() to do that for us; if we have a structure like
281 * /bin -> /usr/bin/ and /usr is a mount point, then the parent that we
282 * look at needs to be /usr, not /. */
283 if (flags & AT_SYMLINK_FOLLOW) {
284 r = chase_symlinks(t, root, CHASE_TRAIL_SLASH, &canonical);
291 parent = dirname_malloc(t);
295 fd = openat(AT_FDCWD, parent, O_DIRECTORY|O_CLOEXEC|O_PATH);
299 return fd_is_mount_point(fd, last_path_component(t), flags);
302 int path_get_mnt_id(const char *path, int *ret) {
305 r = name_to_handle_at_loop(AT_FDCWD, path, NULL, ret, 0);
306 if (IN_SET(r, -EOPNOTSUPP, -ENOSYS, -EACCES, -EPERM, -EOVERFLOW, -EINVAL)) /* kernel/fs don't support this, or seccomp blocks access, or untriggered mount, or name_to_handle_at() is flaky */
307 return fd_fdinfo_mnt_id(AT_FDCWD, path, 0, ret);
312 #if 0 /// UNNEEDED by elogind
313 int umount_recursive(const char *prefix, int flags) {
317 /* Try to umount everything recursively below a
318 * directory. Also, take care of stacked mounts, and keep
319 * unmounting them until they are gone. */
322 _cleanup_fclose_ FILE *proc_self_mountinfo = NULL;
327 proc_self_mountinfo = fopen("/proc/self/mountinfo", "re");
328 if (!proc_self_mountinfo)
331 (void) __fsetlocking(proc_self_mountinfo, FSETLOCKING_BYCALLER);
334 _cleanup_free_ char *path = NULL, *p = NULL;
337 k = fscanf(proc_self_mountinfo,
338 "%*s " /* (1) mount id */
339 "%*s " /* (2) parent id */
340 "%*s " /* (3) major:minor */
341 "%*s " /* (4) root */
342 "%ms " /* (5) mount point */
343 "%*s" /* (6) mount options */
344 "%*[^-]" /* (7) optional fields */
345 "- " /* (8) separator */
346 "%*s " /* (9) file system type */
347 "%*s" /* (10) mount source */
348 "%*s" /* (11) mount options 2 */
349 "%*[^\n]", /* some rubbish at the end */
358 r = cunescape(path, UNESCAPE_RELAX, &p);
362 if (!path_startswith(p, prefix))
365 if (umount2(p, flags) < 0) {
366 r = log_debug_errno(errno, "Failed to umount %s: %m", p);
370 log_debug("Successfully unmounted %s", p);
383 static int get_mount_flags(const char *path, unsigned long *flags) {
386 if (statvfs(path, &buf) < 0)
392 /* Use this function only if do you have direct access to /proc/self/mountinfo
393 * and need the caller to open it for you. This is the case when /proc is
394 * masked or not mounted. Otherwise, use bind_remount_recursive. */
395 int bind_remount_recursive_with_mountinfo(const char *prefix, bool ro, char **blacklist, FILE *proc_self_mountinfo) {
396 _cleanup_set_free_free_ Set *done = NULL;
397 _cleanup_free_ char *cleaned = NULL;
400 assert(proc_self_mountinfo);
402 /* Recursively remount a directory (and all its submounts) read-only or read-write. If the directory is already
403 * mounted, we reuse the mount and simply mark it MS_BIND|MS_RDONLY (or remove the MS_RDONLY for read-write
404 * operation). If it isn't we first make it one. Afterwards we apply MS_BIND|MS_RDONLY (or remove MS_RDONLY) to
405 * all submounts we can access, too. When mounts are stacked on the same mount point we only care for each
406 * individual "top-level" mount on each point, as we cannot influence/access the underlying mounts anyway. We
407 * do not have any effect on future submounts that might get propagated, they migt be writable. This includes
408 * future submounts that have been triggered via autofs.
410 * If the "blacklist" parameter is specified it may contain a list of subtrees to exclude from the
411 * remount operation. Note that we'll ignore the blacklist for the top-level path. */
413 cleaned = strdup(prefix);
417 path_kill_slashes(cleaned);
419 done = set_new(&path_hash_ops);
424 _cleanup_set_free_free_ Set *todo = NULL;
425 bool top_autofs = false;
427 unsigned long orig_flags;
429 todo = set_new(&path_hash_ops);
433 rewind(proc_self_mountinfo);
436 _cleanup_free_ char *path = NULL, *p = NULL, *type = NULL;
439 k = fscanf(proc_self_mountinfo,
440 "%*s " /* (1) mount id */
441 "%*s " /* (2) parent id */
442 "%*s " /* (3) major:minor */
443 "%*s " /* (4) root */
444 "%ms " /* (5) mount point */
445 "%*s" /* (6) mount options (superblock) */
446 "%*[^-]" /* (7) optional fields */
447 "- " /* (8) separator */
448 "%ms " /* (9) file system type */
449 "%*s" /* (10) mount source */
450 "%*s" /* (11) mount options (bind mount) */
451 "%*[^\n]", /* some rubbish at the end */
461 r = cunescape(path, UNESCAPE_RELAX, &p);
465 if (!path_startswith(p, cleaned))
468 /* Ignore this mount if it is blacklisted, but only if it isn't the top-level mount we shall
470 if (!path_equal(cleaned, p)) {
471 bool blacklisted = false;
474 STRV_FOREACH(i, blacklist) {
476 if (path_equal(*i, cleaned))
479 if (!path_startswith(*i, cleaned))
482 if (path_startswith(p, *i)) {
484 log_debug("Not remounting %s, because blacklisted by %s, called for %s", p, *i, cleaned);
492 /* Let's ignore autofs mounts. If they aren't
493 * triggered yet, we want to avoid triggering
494 * them, as we don't make any guarantees for
495 * future submounts anyway. If they are
496 * already triggered, then we will find
497 * another entry for this. */
498 if (streq(type, "autofs")) {
499 top_autofs = top_autofs || path_equal(cleaned, p);
503 if (!set_contains(done, p)) {
504 r = set_consume(todo, p);
513 /* If we have no submounts to process anymore and if
514 * the root is either already done, or an autofs, we
516 if (set_isempty(todo) &&
517 (top_autofs || set_contains(done, cleaned)))
520 if (!set_contains(done, cleaned) &&
521 !set_contains(todo, cleaned)) {
522 /* The prefix directory itself is not yet a mount, make it one. */
523 if (mount(cleaned, cleaned, NULL, MS_BIND|MS_REC, NULL) < 0)
527 (void) get_mount_flags(cleaned, &orig_flags);
528 orig_flags &= ~MS_RDONLY;
530 if (mount(NULL, prefix, NULL, orig_flags|MS_BIND|MS_REMOUNT|(ro ? MS_RDONLY : 0), NULL) < 0)
533 log_debug("Made top-level directory %s a mount point.", prefix);
539 r = set_consume(done, x);
544 while ((x = set_steal_first(todo))) {
546 r = set_consume(done, x);
547 if (IN_SET(r, 0, -EEXIST))
552 /* Deal with mount points that are obstructed by a later mount */
553 r = path_is_mount_point(x, NULL, 0);
554 if (IN_SET(r, 0, -ENOENT))
559 /* Try to reuse the original flag set */
561 (void) get_mount_flags(x, &orig_flags);
562 orig_flags &= ~MS_RDONLY;
564 if (mount(NULL, x, NULL, orig_flags|MS_BIND|MS_REMOUNT|(ro ? MS_RDONLY : 0), NULL) < 0)
567 log_debug("Remounted %s read-only.", x);
572 int bind_remount_recursive(const char *prefix, bool ro, char **blacklist) {
573 _cleanup_fclose_ FILE *proc_self_mountinfo = NULL;
575 proc_self_mountinfo = fopen("/proc/self/mountinfo", "re");
576 if (!proc_self_mountinfo)
579 (void) __fsetlocking(proc_self_mountinfo, FSETLOCKING_BYCALLER);
581 return bind_remount_recursive_with_mountinfo(prefix, ro, blacklist, proc_self_mountinfo);
584 int mount_move_root(const char *path) {
590 if (mount(path, "/", NULL, MS_MOVE, NULL) < 0)
602 bool fstype_is_network(const char *fstype) {
605 x = startswith(fstype, "fuse.");
609 return STR_IN_SET(fstype,
621 "pvfs2", /* OrangeFS */
626 bool fstype_is_api_vfs(const char *fstype) {
627 return STR_IN_SET(fstype,
650 bool fstype_is_ro(const char *fstype) {
651 /* All Linux file systems that are necessarily read-only */
652 return STR_IN_SET(fstype,
658 bool fstype_can_discard(const char *fstype) {
659 return STR_IN_SET(fstype,
666 bool fstype_can_uid_gid(const char *fstype) {
668 /* All file systems that have a uid=/gid= mount option that fixates the owners of all files and directories,
669 * current and future. */
671 return STR_IN_SET(fstype,
682 int repeat_unmount(const char *path, int flags) {
687 /* If there are multiple mounts on a mount point, this
688 * removes them all */
691 if (umount2(path, flags) < 0) {
704 const char* mode_to_inaccessible_node(mode_t mode) {
705 /* This function maps a node type to the correspondent inaccessible node type.
706 * Character and block inaccessible devices may not be created (because major=0 and minor=0),
707 * in such case we map character and block devices to the inaccessible node type socket. */
708 switch(mode & S_IFMT) {
710 return "/run/systemd/inaccessible/reg";
712 return "/run/systemd/inaccessible/dir";
714 if (access("/run/systemd/inaccessible/chr", F_OK) == 0)
715 return "/run/systemd/inaccessible/chr";
716 return "/run/systemd/inaccessible/sock";
718 if (access("/run/systemd/inaccessible/blk", F_OK) == 0)
719 return "/run/systemd/inaccessible/blk";
720 return "/run/systemd/inaccessible/sock";
722 return "/run/systemd/inaccessible/fifo";
724 return "/run/systemd/inaccessible/sock";
729 #if 0 /// UNNEEDED by elogind
730 #define FLAG(name) (flags & name ? STRINGIFY(name) "|" : "")
731 static char* mount_flags_to_string(long unsigned flags) {
733 _cleanup_free_ char *y = NULL;
734 long unsigned overflow;
736 overflow = flags & ~(MS_RDONLY |
761 if (flags == 0 || overflow != 0)
762 if (asprintf(&y, "%lx", overflow) < 0)
765 x = strjoin(FLAG(MS_RDONLY),
769 FLAG(MS_SYNCHRONOUS),
787 FLAG(MS_STRICTATIME),
793 x[strlen(x) - 1] = '\0'; /* truncate the last | */
803 const char *options) {
805 _cleanup_free_ char *fl = NULL, *o = NULL;
809 r = mount_option_mangle(options, flags, &f, &o);
811 return log_full_errno(error_log_level, r,
812 "Failed to mangle mount options %s: %m",
815 fl = mount_flags_to_string(f);
817 if ((f & MS_REMOUNT) && !what && !type)
818 log_debug("Remounting %s (%s \"%s\")...",
819 where, strnull(fl), strempty(o));
820 else if (!what && !type)
821 log_debug("Mounting %s (%s \"%s\")...",
822 where, strnull(fl), strempty(o));
823 else if ((f & MS_BIND) && !type)
824 log_debug("Bind-mounting %s on %s (%s \"%s\")...",
825 what, where, strnull(fl), strempty(o));
826 else if (f & MS_MOVE)
827 log_debug("Moving mount %s → %s (%s \"%s\")...",
828 what, where, strnull(fl), strempty(o));
830 log_debug("Mounting %s on %s (%s \"%s\")...",
831 strna(type), where, strnull(fl), strempty(o));
832 if (mount(what, where, type, f, o) < 0)
833 return log_full_errno(error_log_level, errno,
834 "Failed to mount %s on %s (%s \"%s\"): %m",
835 strna(type), where, strnull(fl), strempty(o));
839 int umount_verbose(const char *what) {
840 log_debug("Umounting %s...", what);
841 if (umount(what) < 0)
842 return log_error_errno(errno, "Failed to unmount %s: %m", what);
847 const char *mount_propagation_flags_to_string(unsigned long flags) {
849 switch (flags & (MS_SHARED|MS_SLAVE|MS_PRIVATE)) {
863 int mount_propagation_flags_from_string(const char *name, unsigned long *ret) {
867 else if (streq(name, "shared"))
869 else if (streq(name, "slave"))
871 else if (streq(name, "private"))
878 #if 0 /// UNNEEDED by elogind
879 int mount_option_mangle(
881 unsigned long mount_flags,
882 unsigned long *ret_mount_flags,
883 char **ret_remaining_options) {
885 const struct libmnt_optmap *map;
886 _cleanup_free_ char *ret = NULL;
890 /* This extracts mount flags from the mount options, and store
891 * non-mount-flag options to '*ret_remaining_options'.
893 * "rw,nosuid,nodev,relatime,size=1630748k,mode=700,uid=1000,gid=1000"
894 * is split to MS_NOSUID|MS_NODEV|MS_RELATIME and
895 * "size=1630748k,mode=700,uid=1000,gid=1000".
896 * See more examples in test-mount-utils.c.
898 * Note that if 'options' does not contain any non-mount-flag options,
899 * then '*ret_remaining_options' is set to NULL instread of empty string.
900 * Note that this does not check validity of options stored in
901 * '*ret_remaining_options'.
902 * Note that if 'options' is NULL, then this just copies 'mount_flags'
903 * to '*ret_mount_flags'. */
905 assert(ret_mount_flags);
906 assert(ret_remaining_options);
908 map = mnt_get_builtin_optmap(MNT_LINUX_MAP);
914 _cleanup_free_ char *word = NULL;
915 const struct libmnt_optmap *ent;
917 r = extract_first_word(&p, &word, ",", EXTRACT_QUOTES);
923 for (ent = map; ent->name; ent++) {
924 /* All entries in MNT_LINUX_MAP do not take any argument.
925 * Thus, ent->name does not contain "=" or "[=]". */
926 if (!streq(word, ent->name))
929 if (!(ent->mask & MNT_INVERT))
930 mount_flags |= ent->id;
931 else if (mount_flags & ent->id)
932 mount_flags ^= ent->id;
937 /* If 'word' is not a mount flag, then store it in '*ret_remaining_options'. */
938 if (!ent->name && !strextend_with_separator(&ret, ",", word, NULL))
942 *ret_mount_flags = mount_flags;
943 *ret_remaining_options = TAKE_PTR(ret);