rtmutex: Plug slow unlock race
authorThomas Gleixner <tglx@linutronix.de>
Wed, 11 Jun 2014 18:44:04 +0000 (18:44 +0000)
committerThomas Gleixner <tglx@linutronix.de>
Mon, 16 Jun 2014 08:03:09 +0000 (10:03 +0200)
When the rtmutex fast path is enabled the slow unlock function can
create the following situation:

spin_lock(foo->m->wait_lock);
foo->m->owner = NULL;
     rt_mutex_lock(foo->m); <-- fast path
free = atomic_dec_and_test(foo->refcnt);
rt_mutex_unlock(foo->m); <-- fast path
if (free)
   kfree(foo);

spin_unlock(foo->m->wait_lock); <--- Use after free.

Plug the race by changing the slow unlock to the following scheme:

     while (!rt_mutex_has_waiters(m)) {
          /* Clear the waiters bit in m->owner */
    clear_rt_mutex_waiters(m);
           owner = rt_mutex_owner(m);
           spin_unlock(m->wait_lock);
           if (cmpxchg(m->owner, owner, 0) == owner)
              return;
           spin_lock(m->wait_lock);
     }

So in case of a new waiter incoming while the owner tries the slow
path unlock we have two situations:

 unlock(wait_lock);
lock(wait_lock);
 cmpxchg(p, owner, 0) == owner
           mark_rt_mutex_waiters(lock);
  acquire(lock);

Or:

 unlock(wait_lock);
lock(wait_lock);
  mark_rt_mutex_waiters(lock);
 cmpxchg(p, owner, 0) != owner
enqueue_waiter();
unlock(wait_lock);
 lock(wait_lock);
 wakeup_next waiter();
 unlock(wait_lock);
lock(wait_lock);
acquire(lock);

If the fast path is disabled, then the simple

   m->owner = NULL;
   unlock(m->wait_lock);

is sufficient as all access to m->owner is serialized via
m->wait_lock;

Also document and clarify the wakeup_next_waiter function as suggested
by Oleg Nesterov.

Reported-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20140611183852.937945560@linutronix.de
Cc: stable@vger.kernel.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
kernel/locking/rtmutex.c

index a8a83a2..fc60594 100644 (file)
@@ -83,6 +83,47 @@ static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
                owner = *p;
        } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
 }
+
+/*
+ * Safe fastpath aware unlock:
+ * 1) Clear the waiters bit
+ * 2) Drop lock->wait_lock
+ * 3) Try to unlock the lock with cmpxchg
+ */
+static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
+       __releases(lock->wait_lock)
+{
+       struct task_struct *owner = rt_mutex_owner(lock);
+
+       clear_rt_mutex_waiters(lock);
+       raw_spin_unlock(&lock->wait_lock);
+       /*
+        * If a new waiter comes in between the unlock and the cmpxchg
+        * we have two situations:
+        *
+        * unlock(wait_lock);
+        *                                      lock(wait_lock);
+        * cmpxchg(p, owner, 0) == owner
+        *                                      mark_rt_mutex_waiters(lock);
+        *                                      acquire(lock);
+        * or:
+        *
+        * unlock(wait_lock);
+        *                                      lock(wait_lock);
+        *                                      mark_rt_mutex_waiters(lock);
+        *
+        * cmpxchg(p, owner, 0) != owner
+        *                                      enqueue_waiter();
+        *                                      unlock(wait_lock);
+        * lock(wait_lock);
+        * wake waiter();
+        * unlock(wait_lock);
+        *                                      lock(wait_lock);
+        *                                      acquire(lock);
+        */
+       return rt_mutex_cmpxchg(lock, owner, NULL);
+}
+
 #else
 # define rt_mutex_cmpxchg(l,c,n)       (0)
 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
@@ -90,6 +131,17 @@ static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
        lock->owner = (struct task_struct *)
                        ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 }
+
+/*
+ * Simple slow path only version: lock->owner is protected by lock->wait_lock.
+ */
+static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
+       __releases(lock->wait_lock)
+{
+       lock->owner = NULL;
+       raw_spin_unlock(&lock->wait_lock);
+       return true;
+}
 #endif
 
 static inline int
@@ -650,7 +702,8 @@ static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 /*
  * Wake up the next waiter on the lock.
  *
- * Remove the top waiter from the current tasks waiter list and wake it up.
+ * Remove the top waiter from the current tasks pi waiter list and
+ * wake it up.
  *
  * Called with lock->wait_lock held.
  */
@@ -671,10 +724,23 @@ static void wakeup_next_waiter(struct rt_mutex *lock)
         */
        rt_mutex_dequeue_pi(current, waiter);
 
-       rt_mutex_set_owner(lock, NULL);
+       /*
+        * As we are waking up the top waiter, and the waiter stays
+        * queued on the lock until it gets the lock, this lock
+        * obviously has waiters. Just set the bit here and this has
+        * the added benefit of forcing all new tasks into the
+        * slow path making sure no task of lower priority than
+        * the top waiter can steal this lock.
+        */
+       lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
 
        raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 
+       /*
+        * It's safe to dereference waiter as it cannot go away as
+        * long as we hold lock->wait_lock. The waiter task needs to
+        * acquire it in order to dequeue the waiter.
+        */
        wake_up_process(waiter->task);
 }
 
@@ -928,12 +994,49 @@ rt_mutex_slowunlock(struct rt_mutex *lock)
 
        rt_mutex_deadlock_account_unlock(current);
 
-       if (!rt_mutex_has_waiters(lock)) {
-               lock->owner = NULL;
-               raw_spin_unlock(&lock->wait_lock);
-               return;
+       /*
+        * We must be careful here if the fast path is enabled. If we
+        * have no waiters queued we cannot set owner to NULL here
+        * because of:
+        *
+        * foo->lock->owner = NULL;
+        *                      rtmutex_lock(foo->lock);   <- fast path
+        *                      free = atomic_dec_and_test(foo->refcnt);
+        *                      rtmutex_unlock(foo->lock); <- fast path
+        *                      if (free)
+        *                              kfree(foo);
+        * raw_spin_unlock(foo->lock->wait_lock);
+        *
+        * So for the fastpath enabled kernel:
+        *
+        * Nothing can set the waiters bit as long as we hold
+        * lock->wait_lock. So we do the following sequence:
+        *
+        *      owner = rt_mutex_owner(lock);
+        *      clear_rt_mutex_waiters(lock);
+        *      raw_spin_unlock(&lock->wait_lock);
+        *      if (cmpxchg(&lock->owner, owner, 0) == owner)
+        *              return;
+        *      goto retry;
+        *
+        * The fastpath disabled variant is simple as all access to
+        * lock->owner is serialized by lock->wait_lock:
+        *
+        *      lock->owner = NULL;
+        *      raw_spin_unlock(&lock->wait_lock);
+        */
+       while (!rt_mutex_has_waiters(lock)) {
+               /* Drops lock->wait_lock ! */
+               if (unlock_rt_mutex_safe(lock) == true)
+                       return;
+               /* Relock the rtmutex and try again */
+               raw_spin_lock(&lock->wait_lock);
        }
 
+       /*
+        * The wakeup next waiter path does not suffer from the above
+        * race. See the comments there.
+        */
        wakeup_next_waiter(lock);
 
        raw_spin_unlock(&lock->wait_lock);