[RFC][PATCH] sched: Fix race in task_group()

[Peter Zijlstra <peterz@infradead.org>]

Stefan reported a crash on a kernel before a3e5d1091c1 ("sched: Don't
call task_group() too many times in set_task_rq()"), he found the reason
to be that the multiple task_group() invocations in set_task_rq()
returned different values.

Looking at all that I found a lack of serialization and plain wrong
comments.

The below tries to fix it using an extra pointer which is updated under
the appropriate scheduler locks. Its not pretty, but I can't really see
another way given how all the cgroup stuff works.

Anybody else got a better idea?


Reported-by: Stefan Bader <stefan.bader@canonical.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 include/linux/init_task.h |   12 +++++++++++-
 include/linux/sched.h     |    5 ++++-
 kernel/sched/core.c       |    9 ++++++++-
 kernel/sched/sched.h      |   23 ++++++++++-------------
 4 files changed, 33 insertions(+), 16 deletions(-)

diff --git a/include/linux/init_task.h b/include/linux/init_task.h
index 4e4bc1a..53be033 100644
--- a/include/linux/init_task.h
+++ b/include/linux/init_task.h
@@ -123,8 +123,17 @@ extern struct group_info init_groups;
 
 extern struct cred init_cred;
 
+extern struct task_group root_task_group;
+
+#ifdef CONFIG_CGROUP_SCHED
+# define INIT_CGROUP_SCHED(tsk)						\
+	.sched_task_group = &root_task_group,
+#else
+# define INIT_CGROUP_SCHED(tsk)
+#endif
+
 #ifdef CONFIG_PERF_EVENTS
-# define INIT_PERF_EVENTS(tsk)					\
+# define INIT_PERF_EVENTS(tsk)						\
 	.perf_event_mutex = 						\
 		 __MUTEX_INITIALIZER(tsk.perf_event_mutex),		\
 	.perf_event_list = LIST_HEAD_INIT(tsk.perf_event_list),
@@ -168,6 +177,7 @@ extern struct cred init_cred;
 	},								\
 	.tasks		= LIST_HEAD_INIT(tsk.tasks),			\
 	INIT_PUSHABLE_TASKS(tsk)					\
+	INIT_CGROUP_SCHED(tsk)						\
 	.ptraced	= LIST_HEAD_INIT(tsk.ptraced),			\
 	.ptrace_entry	= LIST_HEAD_INIT(tsk.ptrace_entry),		\
 	.real_parent	= &tsk,						\
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 32157b9..77437d4 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1246,6 +1246,9 @@ struct task_struct {
 	const struct sched_class *sched_class;
 	struct sched_entity se;
 	struct sched_rt_entity rt;
+#ifdef CONFIG_CGROUP_SCHED
+	struct task_struct *sched_task_group;
+#endif
 
 #ifdef CONFIG_NUMA
 	unsigned long	 numa_contrib;
@@ -2741,7 +2744,7 @@ extern int sched_group_set_rt_period(struct task_group *tg,
 extern long sched_group_rt_period(struct task_group *tg);
 extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
 #endif
-#endif
+#endif /* CONFIG_CGROUP_SCHED */
 
 extern int task_can_switch_user(struct user_struct *up,
 					struct task_struct *tsk);
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 9bb7d28..9adb9a0 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1096,7 +1096,7 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
 	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
 	 *
 	 * sched_move_task() holds both and thus holding either pins the cgroup,
-	 * see set_task_rq().
+	 * see task_group().
 	 *
 	 * Furthermore, all task_rq users should acquire both locks, see
 	 * task_rq_lock().
@@ -7581,6 +7581,8 @@ void sched_destroy_group(struct task_group *tg)
  */
 void sched_move_task(struct task_struct *tsk)
 {
+	struct cgroup_subsys_state *css;
+	struct task_group *tg;
 	int on_rq, running;
 	unsigned long flags;
 	struct rq *rq;
@@ -7595,6 +7597,11 @@ void sched_move_task(struct task_struct *tsk)
 	if (unlikely(running))
 		tsk->sched_class->put_prev_task(rq, tsk);
 
+	tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
+			  struct task_group, css);
+	tg = autogroup_task_group(p, tg);
+	tsk->sched_task_group = tg;
+
 #ifdef CONFIG_FAIR_GROUP_SCHED
 	if (tsk->sched_class->task_move_group)
 		tsk->sched_class->task_move_group(tsk, on_rq);
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 4134d37..c26378c 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -554,22 +554,19 @@ extern int group_balance_cpu(struct sched_group *sg);
 /*
  * Return the group to which this tasks belongs.
  *
- * We use task_subsys_state_check() and extend the RCU verification with
- * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
- * task it moves into the cgroup. Therefore by holding either of those locks,
- * we pin the task to the current cgroup.
+ * We cannot use task_subsys_state() and friends because the cgroup
+ * subsystem changes that value before the cgroup_subsys::attach() method
+ * is called, therefore we cannot pin it and might observe the wrong value.
+ *
+ * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
+ * core changes this before calling sched_move_task().
+ *
+ * Instead we use a 'copy' which is updated from sched_move_task() while
+ * holding both task_struct::pi_lock and rq::lock.
  */
 static inline struct task_group *task_group(struct task_struct *p)
 {
-	struct task_group *tg;
-	struct cgroup_subsys_state *css;
-
-	css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
-			lockdep_is_held(&p->pi_lock) ||
-			lockdep_is_held(&task_rq(p)->lock));
-	tg = container_of(css, struct task_group, css);
-
-	return autogroup_task_group(p, tg);
+	return p->sched_task_group;
 }
 
 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */