sched/wait: Fix a kthread race with wait_woken()
authorPeter Zijlstra <peterz@infradead.org>
Fri, 31 Oct 2014 10:57:30 +0000 (11:57 +0100)
committerIngo Molnar <mingo@kernel.org>
Tue, 4 Nov 2014 06:17:44 +0000 (07:17 +0100)
There is a race between kthread_stop() and the new wait_woken() that
can result in a lack of progress.

CPU 0                                    | CPU 1
                                         |
rfcomm_run()                             | kthread_stop()
  ...                                    |
  if (!test_bit(KTHREAD_SHOULD_STOP))    |
                                         |   set_bit(KTHREAD_SHOULD_STOP)
                                         |   wake_up_process()
    wait_woken()                         |   wait_for_completion()
      set_current_state(INTERRUPTIBLE)   |
      if (!WQ_FLAG_WOKEN)                |
        schedule_timeout()               |
                                         |

After which both tasks will wait.. forever.

Fix this by having wait_woken() check for kthread_should_stop() but
only for kthreads (obviously).

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Peter Hurley <peter@hurleysoftware.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/sched/wait.c

index 4dae188..852143a 100644 (file)
@@ -9,6 +9,7 @@
 #include <linux/mm.h>
 #include <linux/wait.h>
 #include <linux/hash.h>
+#include <linux/kthread.h>
 
 void __init_waitqueue_head(wait_queue_head_t *q, const char *name, struct lock_class_key *key)
 {
@@ -297,6 +298,10 @@ int autoremove_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *
 }
 EXPORT_SYMBOL(autoremove_wake_function);
 
+static inline bool is_kthread_should_stop(void)
+{
+       return (current->flags & PF_KTHREAD) && kthread_should_stop();
+}
 
 /*
  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
@@ -326,7 +331,7 @@ long wait_woken(wait_queue_t *wait, unsigned mode, long timeout)
         * woken_wake_function() such that if we observe WQ_FLAG_WOKEN we must
         * also observe all state before the wakeup.
         */
-       if (!(wait->flags & WQ_FLAG_WOKEN))
+       if (!(wait->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
                timeout = schedule_timeout(timeout);
        __set_current_state(TASK_RUNNING);