RT has a problem when the wait on a futex/rtmutex got interrupted by a
timeout or a signal. task->pi_blocked_on is still set when returning from
rt_mutex_wait_proxy_lock(). The task must acquire the hash bucket lock
after this.

If the hash bucket lock is contended then the
BUG_ON(rt_mutex_real_waiter(task->pi_blocked_on)) in
task_blocks_on_rt_mutex() will trigger.

This can be avoided by clearing task->pi_blocked_on in the return path of
rt_mutex_wait_proxy_lock() which removes the task from the boosting chain
of the rtmutex. That's correct because the task is not longer blocked on
it.

Signed-off-by: Thomas Gleixner <[email protected]>
Reported-by: Engleder Gerhard <[email protected]>
Signed-off-by: Sebastian Andrzej Siewior <[email protected]>
---
 v1…v2: reset ->pi_blocked_on only in the error case.

 kernel/locking/rtmutex.c | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/kernel/locking/rtmutex.c b/kernel/locking/rtmutex.c
index 314fc65a35b1..4675f1197f33 100644
--- a/kernel/locking/rtmutex.c
+++ b/kernel/locking/rtmutex.c
@@ -2400,6 +2400,7 @@ int rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
                               struct hrtimer_sleeper *to,
                               struct rt_mutex_waiter *waiter)
 {
+       struct task_struct *tsk = current;
        int ret;
 
        raw_spin_lock_irq(&lock->wait_lock);
@@ -2409,6 +2410,24 @@ int rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
        /* sleep on the mutex */
        ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter, NULL);
 
+       /*
+        * RT has a problem here when the wait got interrupted by a timeout
+        * or a signal. task->pi_blocked_on is still set. The task must
+        * acquire the hash bucket lock when returning from this function.
+        *
+        * If the hash bucket lock is contended then the
+        * BUG_ON(rt_mutex_real_waiter(task->pi_blocked_on)) in
+        * task_blocks_on_rt_mutex() will trigger. This can be avoided by
+        * clearing task->pi_blocked_on which removes the task from the
+        * boosting chain of the rtmutex. That's correct because the task
+        * is not longer blocked on it.
+        */
+       if (ret) {
+               raw_spin_lock(&tsk->pi_lock);
+               tsk->pi_blocked_on = NULL;
+               raw_spin_unlock(&tsk->pi_lock);
+       }
+
        raw_spin_unlock_irq(&lock->wait_lock);
 
        return ret;
-- 
2.11.0

Reply via email to