]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
dm crypt: fix crash on exit
authorRabin Vincent <rabinv@axis.com>
Wed, 21 Sep 2016 14:22:29 +0000 (16:22 +0200)
committerMike Snitzer <snitzer@redhat.com>
Thu, 22 Sep 2016 15:15:06 +0000 (11:15 -0400)
As the documentation for kthread_stop() says, "if threadfn() may call
do_exit() itself, the caller must ensure task_struct can't go away".
dm-crypt does not ensure this and therefore crashes when crypt_dtr()
calls kthread_stop().  The crash is trivially reproducible by adding a
delay before the call to kthread_stop() and just opening and closing a
dm-crypt device.

 general protection fault: 0000 [#1] PREEMPT SMP
 CPU: 0 PID: 533 Comm: cryptsetup Not tainted 4.8.0-rc7+ #7
 task: ffff88003bd0df40 task.stack: ffff8800375b4000
 RIP: 0010: kthread_stop+0x52/0x300
 Call Trace:
  crypt_dtr+0x77/0x120
  dm_table_destroy+0x6f/0x120
  __dm_destroy+0x130/0x250
  dm_destroy+0x13/0x20
  dev_remove+0xe6/0x120
  ? dev_suspend+0x250/0x250
  ctl_ioctl+0x1fc/0x530
  ? __lock_acquire+0x24f/0x1b10
  dm_ctl_ioctl+0x13/0x20
  do_vfs_ioctl+0x91/0x6a0
  ? ____fput+0xe/0x10
  ? entry_SYSCALL_64_fastpath+0x5/0xbd
  ? trace_hardirqs_on_caller+0x151/0x1e0
  SyS_ioctl+0x41/0x70
  entry_SYSCALL_64_fastpath+0x1f/0xbd

This problem was introduced by bcbd94ff481e ("dm crypt: fix a possible
hang due to race condition on exit").

Looking at the description of that patch (excerpted below), it seems
like the problem it addresses can be solved by just using
set_current_state instead of __set_current_state, since we obviously
need the memory barrier.

| 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: [...]

So this patch just reverts the aforementioned patch and changes the
__set_current_state(TASK_INTERRUPTIBLE) to set_current_state(...).  This
fixes the crash and should also fix the potential hang.

Fixes: bcbd94ff481e ("dm crypt: fix a possible hang due to race condition on exit")
Cc: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org # v4.0+
Signed-off-by: Rabin Vincent <rabinv@axis.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
drivers/md/dm-crypt.c

index 9ba0f0724d28fea0821b871dd62de5ec65868b7f..bcba462a7d14f1a0a77934a131379739645cffb8 100644 (file)
@@ -113,8 +113,7 @@ 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_EXIT_THREAD};
+            DM_CRYPT_SAME_CPU, DM_CRYPT_NO_OFFLOAD };
 
 /*
  * The fields in here must be read only after initialization.
@@ -1207,18 +1206,20 @@ continue_locked:
                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);
+               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;
@@ -1533,13 +1534,8 @@ static void crypt_dtr(struct dm_target *ti)
        if (!cc)
                return;
 
-       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);
+       if (cc->write_thread)
                kthread_stop(cc->write_thread);
-       }
 
        if (cc->io_queue)
                destroy_workqueue(cc->io_queue);