linux: new mount option "idmap"

when the "idmap" mount option is specified, create the mount outside
of the container user namespace context and pass the mount fd to the
container init process.

Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>

crun.1

@@ -638,6 +638,14 @@ itself.
 .PP
 If the \fB\fCrro\fR option is specified then the mount is made recursively read-only.
 
+.SH rro mount options
+.PP
+If the \fB\fCrro\fR option is specified then the mount is made recursively read-only.
+
+.SH idmap mount options
+.PP
+If the \fB\fCidmap\fR option is specified then the mount is ID mapped using the container target user namespace.
+
 .SH Automatically create user namespace
 .PP
 When running as user different than root, an user namespace is

crun.1.md

@@ -507,6 +507,14 @@ itself.
 
 If the `rro` option is specified then the mount is made recursively read-only.
 
+## rro mount options
+
+If the `rro` option is specified then the mount is made recursively read-only.
+
+## idmap mount options
+
+If the `idmap` option is specified then the mount is ID mapped using the container target user namespace.
+
 ## Automatically create user namespace
 
 When running as user different than root, an user namespace is

src/libcrun/linux.c

@@ -291,6 +291,10 @@ struct mount_attr_s
 #  define MOUNT_ATTR_RDONLY 0x00000001 /* Mount read-only */
 #endif
 
+#ifndef MOUNT_ATTR_IDMAP
+#  define MOUNT_ATTR_IDMAP 0x00100000 /* Idmap mount to @userns_fd in struct mount_attr. */
+#endif
+
 static int
 syscall_mount_setattr (int dfd, const char *path, unsigned int flags,
                        struct mount_attr_s *attr)
@@ -359,6 +363,42 @@ make_mount_rro (const char *target, int targetfd, libcrun_error_t *err)
   return 0;
 }
 
+static int
+get_idmapped_mount (const char *src, pid_t pid, libcrun_error_t *err)
+{
+  cleanup_close int open_tree_fd = -1;
+  cleanup_close int fd = -1;
+  int ret;
+  char proc_path[64];
+  struct mount_attr_s attr = {
+    0,
+  };
+
+  sprintf (proc_path, "/proc/%d/ns/user", pid);
+  fd = open (proc_path, O_RDONLY);
+  if (UNLIKELY (fd < 0))
+    return crun_make_error (err, errno, "open `%s`", proc_path);
+
+  open_tree_fd = syscall_open_tree (-1, src,
+                                    AT_NO_AUTOMOUNT | AT_SYMLINK_NOFOLLOW | OPEN_TREE_CLOEXEC | OPEN_TREE_CLONE);
+  if (UNLIKELY (open_tree_fd < 0))
+    return crun_make_error (err, errno, "open `/%s`", src);
+
+  attr.attr_set = MOUNT_ATTR_IDMAP;
+  attr.userns_fd = fd;
+
+  ret = syscall_mount_setattr (open_tree_fd, "", AT_EMPTY_PATH, &attr);
+  if (UNLIKELY (ret < 0))
+    return crun_make_error (err, errno, "mount_setattr `%s`", src);
+
+  /*
+  ret = syscall_fsmount (open_tree_fd, FSMOUNT_CLOEXEC, 0);
+  if (UNLIKELY (ret < 0))
+    return crun_make_error (err, errno, "fsmount `%s`", src);
+*/
+  return get_and_reset (&open_tree_fd);
+}
+
 int
 libcrun_create_keyring (const char *name, const char *label, libcrun_error_t *err)
 {
@@ -432,8 +472,11 @@ enum
 {
   OPTION_TMPCOPYUP = (1 << 0),
   OPTION_RRO = (1 << 1),
+  OPTION_IDMAP = (1 << 2),
 };
 
+#define IDMAP "idmap"
+
 static struct propagation_flags_s propagation_flags[] = { { "defaults", 0, 0, 0 },
                                                           { "bind", 0, MS_BIND, 0 },
                                                           { "rbind", 0, MS_REC | MS_BIND, 0 },
@@ -470,6 +513,7 @@ static struct propagation_flags_s propagation_flags[] = { { "defaults", 0, 0, 0
 
                                                           { "tmpcopyup", 0, 0, OPTION_TMPCOPYUP },
                                                           { "rro", 0, 0, OPTION_RRO },
+                                                          { IDMAP, 0, 0, OPTION_IDMAP },
 
                                                           { NULL, 0, 0, 0 } };
 
@@ -1658,7 +1702,8 @@ do_mounts (libcrun_container_t *container, int rootfsfd, const char *rootfs, con
           ret = fs_move_mount_to (mfd, targetfd, NULL);
           if (LIKELY (ret == 0))
             {
-              ret = do_mount (container, NULL, mfd, target, NULL, flags, data, LABEL_NONE, err);
+              /* Force no MS_BIND flag to not attempt again the bind mount.  */
+              ret = do_mount (container, NULL, mfd, target, NULL, flags & ~MS_BIND, data, LABEL_NONE, err);
               if (UNLIKELY (ret < 0))
                 return ret;
               mounted = true;
@@ -1711,17 +1756,20 @@ do_mounts (libcrun_container_t *container, int rootfsfd, const char *rootfs, con
             return ret;
         }
 
-      if (extra_flags & OPTION_RRO)
+      if (extra_flags & (OPTION_RRO | OPTION_IDMAP))
         {
           cleanup_close int dfd = -1;
 
           dfd = safe_openat (rootfsfd, rootfs, rootfs_len, target, O_DIRECTORY, 0, err);
           if (UNLIKELY (dfd < 0))
             return crun_make_error (err, errno, "open mount target `/%s`", target);
 
-          ret = make_mount_rro (target, dfd, err);
-          if (UNLIKELY (ret < 0))
-            return ret;
+          if (extra_flags & OPTION_RRO)
+            {
+              ret = make_mount_rro (target, dfd, err);
+              if (UNLIKELY (ret < 0))
+                return ret;
+            }
         }
     }
   return 0;
@@ -3257,6 +3305,17 @@ get_fd_map (libcrun_container_t *container)
   return mount_fds;
 }
 
+static bool
+is_idmapped (runtime_spec_schema_defs_mount *mnt)
+{
+  size_t i;
+
+  for (i = 0; i < mnt->options_len; i++)
+    if (strcmp (mnt->options[i], IDMAP) == 0)
+      return true;
+  return false;
+}
+
 static int
 prepare_and_send_mounts (libcrun_container_t *container, struct init_status_s *init_status,
                          pid_t pid, int sync_socket_host, libcrun_error_t *err)
@@ -3274,6 +3333,17 @@ prepare_and_send_mounts (libcrun_container_t *container, struct init_status_s *i
 
   for (i = 0; i < def->mounts_len; i++)
     {
+      if (is_idmapped (def->mounts[i]))
+        {
+          int fd;
+
+          fd = get_idmapped_mount (def->mounts[i]->source, pid, err);
+          if (UNLIKELY (fd < 0))
+            return fd;
+
+          mount_fds->fds[i] = fd;
+        }
+
       if (mount_fds->fds[i] >= 0)
         how_many++;
     }