cgroup: Prevent double killing of css when enabling threaded cgroup
authorMichal Koutný <mkoutny@suse.com>
Thu, 9 Jan 2020 15:05:59 +0000 (16:05 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 5 Feb 2020 21:22:42 +0000 (21:22 +0000)
commit 3bc0bb36fa30e95ca829e9cf480e1ef7f7638333 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

index ef4242e..595c52d 100644 (file)
@@ -3111,8 +3111,6 @@ 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;
 
@@ -3122,6 +3120,8 @@ 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)
@@ -3157,11 +3157,11 @@ 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);
@@ -3448,7 +3448,8 @@ 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;