locking/rtmutex: Dont dereference waiter lockless
authorThomas Gleixner <tglx@linutronix.de>
Wed, 25 Aug 2021 10:33:12 +0000 (12:33 +0200)
committerPeter Zijlstra <peterz@infradead.org>
Wed, 25 Aug 2021 13:42:32 +0000 (15:42 +0200)
The new rt_mutex_spin_on_onwer() loop checks whether the spinning waiter is
still the top waiter on the lock by utilizing rt_mutex_top_waiter(), which
is broken because that function contains a sanity check which dereferences
the top waiter pointer to check whether the waiter belongs to the
lock. That's wrong in the lockless spinwait case:

 CPU 0 CPU 1
 rt_mutex_lock(lock) rt_mutex_lock(lock);
   queue(waiter0)
   waiter0 == rt_mutex_top_waiter(lock)
   rt_mutex_spin_on_onwer(lock, waiter0) { queue(waiter1)
      waiter1 == rt_mutex_top_waiter(lock)
    ...
     top_waiter = rt_mutex_top_waiter(lock)
       leftmost = rb_first_cached(&lock->waiters);
-> signal
dequeue(waiter1)
destroy(waiter1)
       w = rb_entry(leftmost, ....)
       BUG_ON(w->lock != lock)  <- UAF

The BUG_ON() is correct for the case where the caller holds lock->wait_lock
which guarantees that the leftmost waiter entry cannot vanish. For the
lockless spinwait case it's broken.

Create a new helper function which avoids the pointer dereference and just
compares the leftmost entry pointer with current's waiter pointer to
validate that currrent is still elegible for spinning.

Fixes: 992caf7f1724 ("locking/rtmutex: Add adaptive spinwait mechanism")
Reported-by: Sebastian Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20210825102453.981720644@linutronix.de
kernel/locking/rtmutex.c
kernel/locking/rtmutex_common.h

index 8aaa352..b3c0961 100644 (file)
@@ -1329,8 +1329,9 @@ static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
                 *    for CONFIG_PREEMPT_RCU=y)
                 *  - the VCPU on which owner runs is preempted
                 */
-               if (!owner->on_cpu || waiter != rt_mutex_top_waiter(lock) ||
-                   need_resched() || vcpu_is_preempted(task_cpu(owner))) {
+               if (!owner->on_cpu || need_resched() ||
+                   rt_mutex_waiter_is_top_waiter(lock, waiter) ||
+                   vcpu_is_preempted(task_cpu(owner))) {
                        res = false;
                        break;
                }
index 61256de..c47e836 100644 (file)
@@ -95,6 +95,19 @@ static inline int rt_mutex_has_waiters(struct rt_mutex_base *lock)
        return !RB_EMPTY_ROOT(&lock->waiters.rb_root);
 }
 
+/*
+ * Lockless speculative check whether @waiter is still the top waiter on
+ * @lock. This is solely comparing pointers and not derefencing the
+ * leftmost entry which might be about to vanish.
+ */
+static inline bool rt_mutex_waiter_is_top_waiter(struct rt_mutex_base *lock,
+                                                struct rt_mutex_waiter *waiter)
+{
+       struct rb_node *leftmost = rb_first_cached(&lock->waiters);
+
+       return rb_entry(leftmost, struct rt_mutex_waiter, tree_entry) == waiter;
+}
+
 static inline struct rt_mutex_waiter *rt_mutex_top_waiter(struct rt_mutex_base *lock)
 {
        struct rb_node *leftmost = rb_first_cached(&lock->waiters);