[-rt] Fix race condition and following BUG in PI-futex

[Esben Nielsen <nielsen.esben@googlemail.com>]

I ran into the bug on 2.6.17-rt8 with the previous posted patches which 
make pthread_timed_lock() work on UP, but the bug is there without the 
patches - I just can't trigger it - and it is also in the mainline kernel.

The problem is that rt_mutex_next_owner() is used unprotected in 
wake_futex_pi(). At least it isn't probably serialiazed against the next 
owner being signalled or getting a timeout. The only lock, which is 
good enough here, is &pi_state->pi_mutex.wait_lock, so I added this 
protection.

Esben

 kernel/futex.c |   12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)

Index: linux-2.6.17-rt8/kernel/futex.c
===================================================================
--- linux-2.6.17-rt8.orig/kernel/futex.c
+++ linux-2.6.17-rt8/kernel/futex.c
@@ -565,6 +565,7 @@ static int wake_futex_pi(u32 __user *uad
 	if (!pi_state)
 		return -EINVAL;

+	spin_lock(&pi_state->pi_mutex.wait_lock);
 	new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);

 	/*
@@ -590,15 +591,22 @@ static int wake_futex_pi(u32 __user *uad
 	curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
 	dec_preempt_count();

-	if (curval == -EFAULT)
+	if (curval == -EFAULT) {
+		spin_unlock(&pi_state->pi_mutex.wait_lock);
 		return -EFAULT;
-	if (curval != uval)
+	}
+	if (curval != uval) {
+		spin_unlock(&pi_state->pi_mutex.wait_lock);
 		return -EINVAL;
+	}

 	list_del_init(&pi_state->owner->pi_state_list);
 	list_add(&pi_state->list, &new_owner->pi_state_list);
 	pi_state->owner = new_owner;
+	atomic_inc(&pi_state->refcount);
+	spin_unlock(&pi_state->pi_mutex.wait_lock);
 	rt_mutex_unlock(&pi_state->pi_mutex);
+	free_pi_state(pi_state);

 	return 0;
 }
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/