Re: [PATCH] Bluetooth: Fix hidp disconnect deadlock

On Wed, 2011-06-29 at 16:52 -0400, Gustavo F. Padovan wrote:

* Gustavo F. Padovan [off-list ref] [2011-06-29 17:24:56 -0300]:

quoted

* Ilia, Kolominsky [off-list ref] [2011-06-26 09:16:58 +0200]:

quoted

Hi!
IMHO the fix isnt good due to possible race condition which 
will destroy session/task objects - either by a call to kthread_stop
from the timer func or reentry to hidp_del_connection() on 
smp platforms.

....

This should fix the timer issue. Please test.

	Gustavo

Hi Ilia & Gustavo,

After Ilia pointed out the problem with the timer function, I went back
and reviewed *all* the synchronization code relevant to the hid session
thread.

A number of problems were introduced with commit aabf6f89 - when the
session thread was converted from a kernel_thread to a kthread. Although
a kthread is a better choice for representing the session thread, the
naive conversion of atomic/wakeup to kthread_stop() was inappropriate.

kthread_stop() has usage semantics that different significantly from
atomic/wakeup. As we already know, because kthread_stop() blocks on
thread completion, it can introduce deadlocks in code that already uses
exclusion mechanisms. Even with Ilia's new patch, consider the following
sequence:

Thread 0                 Thread 1                 Thread 2
in hidp_del_connection                            in hidp session
  claim r/w sem                                     .
                         timer triggers             .
                           kthread_stop() --------->.
                           *blocks on thread 2*     exits loop
                                                    *blocks for r/w sem*
  in hidp_del_timer
    del_timer_sync()
      *blocks on thread 1*

Deadlock occurs because:
+ thread 0 holds reader lock but is waiting for the timer function on
thread 1 to finish
+ thread 1 has stopped kthread and is waiting for thread completion
+ thread 2 (aka kthread) is waiting to claim writer lock held by thread
0

In addition to the deadlocks and races, kthread_stop() is being called
by hidp_process_hid_control() which is *in the session thread context* -
kthread_stop() cannot be called on itself!

I've been testing patch v2 since Monday which continues to use kthread
but reverts back to the old behavior of atomic/wakeup. It also fixes the
potential for a lost wakeup. Of course, "testing" means running it and
looking at it carefully - as someone on IRC pointed out, kthread really
needs to get instrumented with lockdep.

Regards,
Peter Hurley