[PATCH] Bug Fix: 2.6.{5,6}: i8042 module unload bug

[Sau Dan Lee <danlee@informatik.uni-freiburg.de>]

I've reported  this bug  3 days ago.   But since  it is still  not yet
fixed in the new 2.6.6, I'm reposting this.


    Dan> I've successfully and repeated hanged vanilla 2.6.5 kernel.
    Dan> Here is how I do it:

    Dan> 1) Configure and compile kernel (on Intel x86) so that
    Dan>    'serio', 'i8042', 'atkbd' are all modules.  i.e.
    Dan>    CONFIG_SERIO_I8042=m CONFIG_SERIO=m CONFIG_KEYBOARD_ATKBD=m

    Dan> 2) Boot the compiled kernel.  
    Dan> 3) Log in from network.  
    Dan> 4) modprobe the following modules in order: serio, i8042, atkbd
    Dan> 5) Now, keyboard works, as expected.  
    Dan> 6) rmmod i8042 ===> hangs.  CAPSLOCK doesn't respond (of
    Dan>    course, i8042 is rmoved), machine doesn't respond to ping
    Dan>    (why?), although my PCMCIA network card still blinks and the
    Dan>    hardware indicate of PCMCIA activity also blinks.  But
    Dan>    (un)plugging the PCMCIA card gives no response.

    Dan> The same is observed in 2.6.6-rc3.

The same is also observed in 2.6.6.


I've found and fixed the bug.

It has  nothing to  do with module  dependencies.  Rather, it's  a bug
with improper shutdown of a timer.


Cause: When the i8042 module  is unloaded, the i8042_exit() shuts down
       it  timer  by calling  del_timer_sync().   However, after  that
       call, there can still be interrupts from the i8042 chip (due to
       keyboard  or  mouse activities).   Such  interrupts invoke  the
       interrupt   handler   i8042_interrupt(),   which   would   call
       mod_timer() to reschedule a  new timer event.  But i8042_exit()
       won't want  for this new timer  event.  It may  exit before the
       event arrives.   The module is unloaded  as i8042_exit() exits,
       but this  can happen  before the new  timer event comes.   As a
       result,  when  the timer  expires,  the  kernel  tries to  call
       i8042_timer_func(), which has been unloaded!  This results in a
       complete freeze.

The Fix: is  to make sure interrupts from the  i8042 chip are disabled
         before calling  del_timer_sync() to  clean up the  timer.  In
         i8042_exit(), this  is as easy  as swapping the order  of the
         calls to del_timer_sync() and i8042_controller_cleanup().


The Patch:



--- linux-2.6.6-vanilla/drivers/input/serio/i8042.c	2004/05/06 13:24:46	1.1
+++ linux-2.6.6/drivers/input/serio/i8042.c	2004/05/06 15:17:50	1.2
@@ -990,24 +990,28 @@
 	unregister_reboot_notifier(&i8042_notifier);
 
 	if (i8042_pm_dev)
 		pm_unregister(i8042_pm_dev);
 
 	if (i8042_sysdev_initialized) {
 		sysdev_unregister(&device_i8042);
 		sysdev_class_unregister(&kbc_sysclass);
 	}
 
-	del_timer_sync(&i8042_timer);
-
 	i8042_controller_cleanup();
 	
+	/* we must delete the timer AFTER the i8042 chip is cleaned up,
+	   so as to prevent any more interrupts to invoke i8042_interrupt()
+	   which then schedules newer timer events.
+	*/
+	del_timer_sync(&i8042_timer);
+
 	if (i8042_kbd_values.exists)
 		serio_unregister_port(&i8042_kbd_port);
 
 	if (i8042_aux_values.exists)
 		serio_unregister_port(&i8042_aux_port);
 	
 	for (i = 0; i < 4; i++)
 		if (i8042_mux_values[i].exists)
 			serio_unregister_port(i8042_mux_port + i);
 
 



-- 
Sau Dan LEE                     李守敦(Big5)                    ~{@nJX6X~}(HZ) 

E-mail: danlee@informatik.uni-freiburg.de
Home page: http://www.informatik.uni-freiburg.de/~danlee