tree 99962489f95ad6219a5db52a45acc0e168fb4f5d
parent ea0de4ec5489da0fe738b274effac4f950e93d76
author James Hogan <james.hogan@imgtec.com> 1417438510 -0300
committer Mauro Carvalho Chehab <mchehab@osg.samsung.com> 1417714772 -0200

[media] img-ir/hw: Fix potential deadlock stopping timer

The end timer is used for switching back from repeat code timings when
no repeat codes have been received for a certain amount of time. When
the protocol is changed, the end timer is deleted synchronously with
del_timer_sync(), however this takes place while holding the main spin
lock, and the timer handler also needs to acquire the spin lock.

This opens the possibility of a deadlock on an SMP system if the
protocol is changed just as the repeat timer is expiring. One CPU could
end up in img_ir_set_decoder() holding the lock and waiting for the end
timer to complete, while the other CPU is stuck in the timer handler
spinning on the lock held by the first CPU.

Lockdep also spots a possible lock inversion in the same code, since
img_ir_set_decoder() acquires the img-ir lock before the timer lock, but
the timer handler will try and acquire them the other way around:

=========================================================
[ INFO: possible irq lock inversion dependency detected ]
3.18.0-rc5+ #957 Not tainted
---------------------------------------------------------
swapper/0/0 just changed the state of lock:
 (((&hw->end_timer))){+.-...}, at: [<4006ae5c>] _call_timer_fn+0x0/0xfc
but this lock was taken by another, HARDIRQ-safe lock in the past:
 (&(&priv->lock)->rlock#2){-.....}

and interrupts could create inverse lock ordering between them.

other info that might help us debug this:
 Possible interrupt unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(((&hw->end_timer)));
                               local_irq_disable();
                               lock(&(&priv->lock)->rlock#2);
                               lock(((&hw->end_timer)));
  <Interrupt>
    lock(&(&priv->lock)->rlock#2);

 *** DEADLOCK ***

This is fixed by releasing the main spin lock while performing the
del_timer_sync() call. The timer is prevented from restarting before the
lock is reacquired by a new "stopping" flag which img_ir_handle_data()
checks before updating the timer.

---------------------------------------------------------
swapper/0/0 just changed the state of lock:
 (((&hw->end_timer))){+.-...}, at: [<4006ae5c>] _call_timer_fn+0x0/0xfc
but this lock was taken by another, HARDIRQ-safe lock in the past:
 (&(&priv->lock)->rlock#2){-.....}
and interrupts could create inverse lock ordering between them.
other info that might help us debug this:
 Possible interrupt unsafe locking scenario:
       CPU0                    CPU1
       ----                    ----
  lock(((&hw->end_timer)));
                               local_irq_disable();
                               lock(&(&priv->lock)->rlock#2);
                               lock(((&hw->end_timer)));
  <Interrupt>
    lock(&(&priv->lock)->rlock#2);
 *** DEADLOCK ***
This is fixed by releasing the main spin lock while performing the
del_timer_sync() call. The timer is prevented from restarting before the
lock is reacquired by a new "stopping" flag which img_ir_handle_data()
checks before updating the timer.

Signed-off-by: James Hogan <james.hogan@imgtec.com>
Cc: Sifan Naeem <sifan.naeem@imgtec.com>
Cc: <stable@vger.kernel.org> # v3.15+
Signed-off-by: Mauro Carvalho Chehab <mchehab@osg.samsung.com>