[PATCH] serial: fix possible imx deadlock

[Andrew Morton]

On Mon, 15 Mar 2010 14:45:57 +0100
Sascha Hauer <s.hauer at pengutronix.de> wrote:

> On Mon, Mar 15, 2010 at 10:08:58AM +0100, John Ogness wrote:
> > This patch fixes a possible deadlock situation with the port and timer
> > locks. The calling paths for the deadlock situation are:
> > 
> > 1. imx_set_termios() -> LOCK(port.lock)
> >       -> del_timer_sync() -> LOCK(timer.base.lock)
> > 
> > 2. run_timer_softirq() -> LOCK(timer.base.lock)
> >       -> imx_timeout() -> LOCK(port.lock)
> > 
> > This patch is against 2.6.34-rc1.
> > 
> > Signed-off-by: John Ogness <john.ogness at linutronix.de>
> > ---
> >  drivers/serial/imx.c |    8 ++++++--
> >  1 file changed, 6 insertions(+), 2 deletions(-)
> > diff -Naurp a/drivers/serial/imx.c b/drivers/serial/imx.c
> > --- a/drivers/serial/imx.c	2010-03-08 19:45:44.000000000 +0100
> > +++ b/drivers/serial/imx.c	2010-03-15 09:52:39.855261060 +0100
> > @@ -836,6 +836,12 @@ imx_set_termios(struct uart_port *port, 
> >  	baud = uart_get_baud_rate(port, termios, old, 50, port->uartclk / 16);
> >  	quot = uart_get_divisor(port, baud);
> >  
> > +	/*
> > +	 * Stop our timer. This is done with the port unlocked
> > +	 * to avoid possible deadlock against the timer function.
> > +	 */
> > +	del_timer_sync(&sport->timer);
> > +
> 
> Is this call needed here anyway? Only few drivers have it. Does anyone
> have more insights on this?

The other drivers might be buggy ;)

I think the del_timer can probably be removed, as the later
imx_enable_ms() will do a mod_timer() on it.  Unless that
imx_enable_ms() doesn't get performed, of course..

Plus it could be that we want to kill the timer because we're about to
change various pieces of state and we don't want the timer handler
running while that state is in flux.  Presumably that's why it's all
done under the spinlock.

Moving the del_timer_sync() to outside the spinlock might introduce
races though - perhaps userspace can cause the timer to get re-enabled
after it's been del_timer()'ed.