dm crypt: fix a possible hang due to race condition on exit

A kernel thread executes __set_current_state(TASK_INTERRUPTIBLE),
__add_wait_queue, spin_unlock_irq and then tests kthread_should_stop().
It is possible that the processor reorders memory accesses so that
kthread_should_stop() is executed before __set_current_state().  If such
reordering happens, there is a possible race on thread termination:

CPU 0:
calls kthread_should_stop()
	it tests KTHREAD_SHOULD_STOP bit, returns false
CPU 1:
calls kthread_stop(cc->write_thread)
	sets the KTHREAD_SHOULD_STOP bit
	calls wake_up_process on the kernel thread, that sets the thread
	state to TASK_RUNNING
CPU 0:
sets __set_current_state(TASK_INTERRUPTIBLE)
spin_unlock_irq(&cc->write_thread_wait.lock)
schedule() - and the process is stuck and never terminates, because the
	state is TASK_INTERRUPTIBLE and wake_up_process on CPU 1 already
	terminated

Fix this race condition by using a new flag DM_CRYPT_EXIT_THREAD to
signal that the kernel thread should exit.  The flag is set and tested
while holding cc->write_thread_wait.lock, so there is no possibility of
racy access to the flag.

Also, remove the unnecessary set_task_state(current, TASK_RUNNING)
following the schedule() call.  When the process was woken up, its state
was already set to TASK_RUNNING.  Other kernel code also doesn't set the
state to TASK_RUNNING following schedule() (for example,
do_wait_for_common in completion.c doesn't do it).

Fixes: dc26762 ("dm crypt: offload writes to thread")
Signed-off-by: Mikulas Patocka <[email protected]>
Cc: [email protected] # v4.0+
Signed-off-by: Mike Snitzer <[email protected]>

Author: Mikulas Patocka

drivers/md/dm-crypt.c

@@ -112,7 +112,8 @@ struct iv_tcw_private {
  * and encrypts / decrypts at the same time.
  */
 enum flags { DM_CRYPT_SUSPENDED, DM_CRYPT_KEY_VALID,
-	     DM_CRYPT_SAME_CPU, DM_CRYPT_NO_OFFLOAD };
+	     DM_CRYPT_SAME_CPU, DM_CRYPT_NO_OFFLOAD,
+	     DM_CRYPT_EXIT_THREAD};
 
 /*
  * The fields in here must be read only after initialization.
@@ -1203,20 +1204,18 @@ static int dmcrypt_write(void *data)
 		if (!RB_EMPTY_ROOT(&cc->write_tree))
 			goto pop_from_list;
 
+		if (unlikely(test_bit(DM_CRYPT_EXIT_THREAD, &cc->flags))) {
+			spin_unlock_irq(&cc->write_thread_wait.lock);
+			break;
+		}
+
 		__set_current_state(TASK_INTERRUPTIBLE);
 		__add_wait_queue(&cc->write_thread_wait, &wait);
 
 		spin_unlock_irq(&cc->write_thread_wait.lock);
 
-		if (unlikely(kthread_should_stop())) {
-			set_task_state(current, TASK_RUNNING);
-			remove_wait_queue(&cc->write_thread_wait, &wait);
-			break;
-		}
-
 		schedule();
 
-		set_task_state(current, TASK_RUNNING);
 		spin_lock_irq(&cc->write_thread_wait.lock);
 		__remove_wait_queue(&cc->write_thread_wait, &wait);
 		goto continue_locked;
@@ -1531,8 +1530,13 @@ static void crypt_dtr(struct dm_target *ti)
 	if (!cc)
 		return;
 
-	if (cc->write_thread)
+	if (cc->write_thread) {
+		spin_lock_irq(&cc->write_thread_wait.lock);
+		set_bit(DM_CRYPT_EXIT_THREAD, &cc->flags);
+		wake_up_locked(&cc->write_thread_wait);
+		spin_unlock_irq(&cc->write_thread_wait.lock);
 		kthread_stop(cc->write_thread);
+	}
 
 	if (cc->io_queue)
 		destroy_workqueue(cc->io_queue);