Merge pull request #21077 from poettering/mount-setattr

optimize remounting with mount_setattr() syscall

4 files changed, 90 insertions, 19 deletions

diff --git a/src/basic/missing_syscall.h b/src/basic/missing_syscall.h
index 5e80fa79fd..2f67adaec3 100644
--- a/src/basic/missing_syscall.h
+++ b/src/basic/missing_syscall.h
@@ -465,10 +465,18 @@ struct mount_attr;
 #define MOUNT_ATTR_IDMAP 0x00100000
 #endif
 
+#ifndef MOUNT_ATTR_NOSYMFOLLOW
+#define MOUNT_ATTR_NOSYMFOLLOW 0x00200000
+#endif
+
 #ifndef AT_RECURSIVE
 #define AT_RECURSIVE 0x8000
 #endif
 
+#ifndef MOUNT_ATTR_SIZE_VER0
+#define MOUNT_ATTR_SIZE_VER0 32
+#endif
+
 static inline int missing_mount_setattr(
                 int dfd,
                 const char *path,
diff --git a/src/shared/mount-util.c b/src/shared/mount-util.c
index a7217adfa1..813fdc1664 100644
--- a/src/shared/mount-util.c
+++ b/src/shared/mount-util.c
@@ -134,6 +134,31 @@ int umount_recursive(const char *prefix, int flags) {
         return n;
 }
 
+#define MS_CONVERTIBLE_FLAGS (MS_RDONLY|MS_NOSUID|MS_NODEV|MS_NOEXEC|MS_NOSYMFOLLOW)
+
+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;
+
+        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(
@@ -143,12 +168,44 @@ int bind_remount_recursive_with_mountinfo(
                 char **deny_list,
                 FILE *proc_self_mountinfo) {
 
+        _cleanup_fclose_ FILE *proc_self_mountinfo_opened = NULL;
         _cleanup_set_free_ Set *done = NULL;
         unsigned n_tries = 0;
         int r;
 
         assert(prefix);
-        assert(proc_self_mountinfo);
+
+        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 (!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
@@ -343,22 +400,6 @@ int bind_remount_recursive_with_mountinfo(
         }
 }
 
-int bind_remount_recursive(
-                const char *prefix,
-                unsigned long new_flags,
-                unsigned long flags_mask,
-                char **deny_list) {
-
-        _cleanup_fclose_ FILE *proc_self_mountinfo = NULL;
-        int r;
-
-        r = fopen_unlocked("/proc/self/mountinfo", "re", &proc_self_mountinfo);
-        if (r < 0)
-                return r;
-
-        return bind_remount_recursive_with_mountinfo(prefix, new_flags, flags_mask, deny_list, proc_self_mountinfo);
-}
-
 int bind_remount_one_with_mountinfo(
                 const char *path,
                 unsigned long new_flags,
@@ -374,6 +415,23 @@ int bind_remount_one_with_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! */
+        }
+
         rewind(proc_self_mountinfo);
 
         table = mnt_new_table();
diff --git a/src/shared/mount-util.h b/src/shared/mount-util.h
index 36501c2c4a..d9fb801e61 100644
--- a/src/shared/mount-util.h
+++ b/src/shared/mount-util.h
@@ -40,8 +40,12 @@ int mount_nofollow(const char *source, const char *target, const char *filesyste
 
 int repeat_unmount(const char *path, int flags);
 int umount_recursive(const char *target, int flags);
-int bind_remount_recursive(const char *prefix, unsigned long new_flags, unsigned long flags_mask, char **deny_list);
+
 int bind_remount_recursive_with_mountinfo(const char *prefix, unsigned long new_flags, unsigned long flags_mask, char **deny_list, FILE *proc_self_mountinfo);
+static inline int bind_remount_recursive(const char *prefix, unsigned long new_flags, unsigned long flags_mask, char **deny_list) {
+        return bind_remount_recursive_with_mountinfo(prefix, new_flags, flags_mask, deny_list, NULL);
+}
+
 int bind_remount_one_with_mountinfo(const char *path, unsigned long new_flags, unsigned long flags_mask, FILE *proc_self_mountinfo);
 
 int mount_move_root(const char *path);
diff --git a/src/test/test-mount-util.c b/src/test/test-mount-util.c
index d3d004071b..fdf2a249ed 100644
--- a/src/test/test-mount-util.c
+++ b/src/test/test-mount-util.c
@@ -169,7 +169,7 @@ static void test_bind_remount_recursive(void) {
                         assert_se(!FLAGS_SET(svfs.f_flag, ST_RDONLY));
 
                         /* Now mark the path we currently run for read-only */
-                        assert_se(bind_remount_recursive(p, MS_RDONLY, MS_RDONLY, STRV_MAKE("/sys/kernel")) >= 0);
+                        assert_se(bind_remount_recursive(p, MS_RDONLY, MS_RDONLY, path_equal(p, "/sys") ? STRV_MAKE("/sys/kernel") : NULL) >= 0);
 
                         /* Ensure that this worked on the top-level */
                         assert_se(statvfs(p, &svfs) >= 0);
@@ -209,6 +209,7 @@ static void test_bind_remount_one(void) {
                 assert_se(fopen_unlocked("/proc/self/mountinfo", "re", &proc_self_mountinfo) >= 0);
 
                 assert_se(bind_remount_one_with_mountinfo("/run", MS_RDONLY, MS_RDONLY, proc_self_mountinfo) >= 0);
+                assert_se(bind_remount_one_with_mountinfo("/run", MS_NOEXEC, MS_RDONLY|MS_NOEXEC, proc_self_mountinfo) >= 0);
                 assert_se(bind_remount_one_with_mountinfo("/proc/idontexist", MS_RDONLY, MS_RDONLY, proc_self_mountinfo) == -ENOENT);
                 assert_se(bind_remount_one_with_mountinfo("/proc/self", MS_RDONLY, MS_RDONLY, proc_self_mountinfo) == -EINVAL);
                 assert_se(bind_remount_one_with_mountinfo("/", MS_RDONLY, MS_RDONLY, proc_self_mountinfo) >= 0);