cgroup: Prevent double killing of css when enabling threaded cgroup

commit 3bc0bb36 upstream.

The test_cgcore_no_internal_process_constraint_on_threads selftest when
running with subsystem controlling noise triggers two warnings:

> [  597.443115] WARNING: CPU: 1 PID: 28167 at kernel/cgroup/cgroup.c:3131 cgroup_apply_control_enable+0xe0/0x3f0
> [  597.443413] WARNING: CPU: 1 PID: 28167 at kernel/cgroup/cgroup.c:3177 cgroup_apply_control_disable+0xa6/0x160

Both stem from a call to cgroup_type_write. The first warning was also
triggered by syzkaller.

When we're switching cgroup to threaded mode shortly after a subsystem
was disabled on it, we can see the respective subsystem css dying there.

The warning in cgroup_apply_control_enable is harmless in this case
since we're not adding new subsys anyway.
The warning in cgroup_apply_control_disable indicates an attempt to kill
css of recently disabled subsystem repeatedly.

The commit prevents these situations by making cgroup_type_write wait
for all dying csses to go away before re-applying subtree controls.
When at it, the locations of WARN_ON_ONCE calls are moved so that
warning is triggered only when we are about to misuse the dying css.

Reported-by: syzbot+5493b2a54d31d6aea629@syzkaller.appspotmail.com
Reported-by: Christian Brauner <christian.brauner@ubuntu.com>
Signed-off-by: Michal Koutný <mkoutny@suse.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

kernel/cgroup/cgroup.c

@@ -2940,8 +2940,6 @@ static int cgroup_apply_control_enable(struct cgroup *cgrp)
 		for_each_subsys(ss, ssid) {
 			struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
 
-			WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
-
 			if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
 				continue;
 
@@ -2951,6 +2949,8 @@ static int cgroup_apply_control_enable(struct cgroup *cgrp)
 					return PTR_ERR(css);
 			}
 
+			WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
+
 			if (css_visible(css)) {
 				ret = css_populate_dir(css);
 				if (ret)
@@ -2986,11 +2986,11 @@ static void cgroup_apply_control_disable(struct cgroup *cgrp)
 		for_each_subsys(ss, ssid) {
 			struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
 
-			WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
-
 			if (!css)
 				continue;
 
+			WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
+
 			if (css->parent &&
 			    !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
 				kill_css(css);
@@ -3277,7 +3277,8 @@ static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
 	if (strcmp(strstrip(buf), "threaded"))
 		return -EINVAL;
 
-	cgrp = cgroup_kn_lock_live(of->kn, false);
+	/* drain dying csses before we re-apply (threaded) subtree control */
+	cgrp = cgroup_kn_lock_live(of->kn, true);
 	if (!cgrp)
 		return -ENOENT;