[PATCH RT] sched: Prevent task state corruption by spurious lock wakeup

[Thomas Gleixner <tglx@linutronix.de>]

Mathias and some others reported GDB failures on RT.

The following scenario leads to task state corruption:

CPU0						CPU1

T1->state = TASK_XXX;
spin_lock(&lock)
  rt_spin_lock_slowlock(&lock->rtmutex)
    raw_spin_lock(&rtm->wait_lock);
    T1->saved_state = current->state;
    T1->state = TASK_UNINTERRUPTIBLE;
						spin_unlock(&lock)
    task_blocks_on_rt_mutex(rtm)  		  rt_spin_lock_slowunlock(&lock->rtmutex)
      queue_waiter(rtm)				    raw_spin_lock(&rtm->wait_lock);
      pi_chain_walk(rtm)
        raw_spin_unlock(&rtm->wait_lock);
						    mark_top_waiter_for_wakeup(T1)
						    raw_spin_unlock(&rtm->wait_lock);
      raw_spin_lock(&rtm->wait_lock);
						    wake_up_top_waiter()    

    for (;;) {
      if (__try_to_take_rt_mutex())  <- Succeeds
        break;
      ...
    }
						     try_to_wake_up(T1)
    T1->state = T1->saved_state;
    ==> T1->state == TASK_XXX
						       ttwu_do_wakeup(T1)
				FAIL ---->               T1->state = TASK_RUNNING;


In most cases this is harmless because waiting for some event, which is the
usual reason for TASK_[UN]INTERRUPTIBLE, has to be safe against other forms
of spurious wakeups anyway.

But in case of TASK_TRACED this is actually fatal, because the task loses
the TASK_TRACED state. In consequence it fails to consume SIGSTOP which was
sent from the debugger and actually delivers SIGSTOP to the task which
breaks the ptrace mechanics and brings the debugger into an unexpected
state.

The cure is way simpler as figuring it out:

In a lock wakeup, check whether the task is actually blocked on a lock. If
yes, deliver it. If not, consider the wakeup spurious and exit the wake up
code without touching tasks state.

Reported-by: Mathias Koehrer <mathias.koehrer@etas.com>
Reported-by: David Hauck <davidh@netacquire.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: stable-rt@vger.kernel.org
---
 kernel/sched/core.c |   12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

Index: b/kernel/sched/core.c
===================================================================
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2174,8 +2174,18 @@ try_to_wake_up(struct task_struct *p, un
 	 * If this is a regular wakeup, then we can unconditionally
 	 * clear the saved state of a "lock sleeper".
 	 */
-	if (!(wake_flags & WF_LOCK_SLEEPER))
+	if (!(wake_flags & WF_LOCK_SLEEPER)) {
 		p->saved_state = TASK_RUNNING;
+	} else {
+		/*
+		 * The task might not yet have reached schedule() and has
+		 * taken over the lock already and restored the saved
+		 * state. Prevent that this spurious wakeup destroys the saved
+		 * state.
+		 */
+		if (!tsk_is_pi_blocked(p))
+			goto out;
+	}
 
 	trace_sched_waking(p);