[PATCH -mm 3/4] cputimers: thread_group_times: make it rcu-safe

[Oleg Nesterov <oleg@redhat.com>]

thread_group_times() relies on ->siglock hold by the caller when it
updates sig->prev_Xtime members. Change the code to take cputimer.lock
around the assignment.

This makes it rcu-safe and fixes the theoretical race. wait_task_zombie()
calls it lockless and thus can race with do_task_stat().

Note: we are taking cputimer.lock under ->siglock but this dependency
is not new, thread_group_cputimer() already does this.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
---

 kernel/sched.c |    3 +++
 1 file changed, 3 insertions(+)

--- 34-rc1/kernel/sched.c~cpuacct_3_thread_group_times_dont_rely_on_siglock	2010-03-29 19:44:25.000000000 +0200
+++ 34-rc1/kernel/sched.c	2010-03-29 19:46:21.000000000 +0200
@@ -3449,6 +3449,7 @@ void thread_group_times(struct task_stru
 	struct signal_struct *sig = p->signal;
 	struct task_cputime cputime;
 	cputime_t rtime, utime, total;
+	unsigned long flags;
 
 	thread_group_cputime(p, &cputime);
 
@@ -3464,9 +3465,11 @@ void thread_group_times(struct task_stru
 	} else
 		utime = rtime;
 
+	spin_lock_irqsave(&sig->cputimer.lock, flags);
 	sig->prev_utime = max(sig->prev_utime, utime);
 	sig->prev_stime = max(sig->prev_stime,
 			      cputime_sub(rtime, sig->prev_utime));
+	spin_unlock_irqrestore(&sig->cputimer.lock, flags);
 
 	*ut = sig->prev_utime;
 	*st = sig->prev_stime;