cgroup: avoid attaching a cgroup root to two different superblocks, take 2

Commit bfb0b80d ("cgroup: avoid attaching a cgroup root to two
different superblocks") is broken.  Now we try to fix the race by
delaying the initialization of cgroup root refcnt until a superblock
has been allocated.
Reported-by: NDmitry Vyukov <dvyukov@google.com>
Reported-by: NAndrei Vagin <avagin@virtuozzo.com>
Tested-by: NAndrei Vagin <avagin@virtuozzo.com>
Signed-off-by: NZefan Li <lizefan@huawei.com>
Signed-off-by: NTejun Heo <tj@kernel.org>

kernel/cgroup/cgroup-internal.h

@@ -164,7 +164,7 @@ int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
 
 void cgroup_free_root(struct cgroup_root *root);
 void init_cgroup_root(struct cgroup_root *root, struct cgroup_sb_opts *opts);
-int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask);
+int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask, int ref_flags);
 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask);
 struct dentry *cgroup_do_mount(struct file_system_type *fs_type, int flags,
 			       struct cgroup_root *root, unsigned long magic,

kernel/cgroup/cgroup-v1.c

@@ -1072,6 +1072,7 @@ struct dentry *cgroup1_mount(struct file_system_type *fs_type, int flags,
 	struct cgroup_subsys *ss;
 	struct dentry *dentry;
 	int i, ret;
+	bool new_root = false;
 
 	cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
 
@@ -1181,10 +1182,11 @@ struct dentry *cgroup1_mount(struct file_system_type *fs_type, int flags,
 		ret = -ENOMEM;
 		goto out_unlock;
 	}
+	new_root = true;
 
 	init_cgroup_root(root, &opts);
 
-	ret = cgroup_setup_root(root, opts.subsys_mask);
+	ret = cgroup_setup_root(root, opts.subsys_mask, PERCPU_REF_INIT_DEAD);
 	if (ret)
 		cgroup_free_root(root);
 
@@ -1200,6 +1202,18 @@ struct dentry *cgroup1_mount(struct file_system_type *fs_type, int flags,
 	dentry = cgroup_do_mount(&cgroup_fs_type, flags, root,
 				 CGROUP_SUPER_MAGIC, ns);
 
+	/*
+	 * There's a race window after we release cgroup_mutex and before
+	 * allocating a superblock. Make sure a concurrent process won't
+	 * be able to re-use the root during this window by delaying the
+	 * initialization of root refcnt.
+	 */
+	if (new_root) {
+		mutex_lock(&cgroup_mutex);
+		percpu_ref_reinit(&root->cgrp.self.refcnt);
+		mutex_unlock(&cgroup_mutex);
+	}
+
 	/*
 	 * If @pinned_sb, we're reusing an existing root and holding an
 	 * extra ref on its sb.  Mount is complete.  Put the extra ref.

kernel/cgroup/cgroup.c

@@ -1645,7 +1645,7 @@ void init_cgroup_root(struct cgroup_root *root, struct cgroup_sb_opts *opts)
 		set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
 }
 
-int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
+int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask, int ref_flags)
 {
 	LIST_HEAD(tmp_links);
 	struct cgroup *root_cgrp = &root->cgrp;
@@ -1661,8 +1661,8 @@ int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
 	root_cgrp->id = ret;
 	root_cgrp->ancestor_ids[0] = ret;
 
-	ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
-			      GFP_KERNEL);
+	ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
+			      ref_flags, GFP_KERNEL);
 	if (ret)
 		goto out;
 
@@ -4517,7 +4517,7 @@ int __init cgroup_init(void)
 	hash_add(css_set_table, &init_css_set.hlist,
 		 css_set_hash(init_css_set.subsys));
 
-	BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
+	BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0, 0));
 
 	mutex_unlock(&cgroup_mutex);