Re: UART/TTY console deadlock

[Sergey Senozhatsky <sergey.senozhatsky@gmail.com>]

On (20/06/30 12:21), Petr Mladek wrote:
> > So... Do we need to hold uart->port when we disable port->irq? What do we
> > race with? Module removal? The function bumps device PM counter (albeit
> > for UART_CAP_RPM ports only).
> 
> Honestly, I do not see where a PM counter gets incremented.

serial8250_do_startup()
 serial8250_rpm_get()
  pm_runtime_get_sync(p->port.dev)

But this does not happen for all ports, just for UART_CAP_RPM ones.

> Anyway, __disable_irq_nosync() does nothing when
> irq_get_desc_buslock() returns NULL. And irq_get_desc_buslock()
> takes desc->lock when desc exist. This should be enough to
> synchronize any calls.
> 
> > But, at the same time, we do a whole bunch
> > of unprotected port->FOO accesses in serial8250_do_startup(). We even set
> > the IRQF_SHARED up->port.irqflags without grabbing the port->lock:
> > 
> > 	 up->port.irqflags |= IRQF_SHARED;
> > 	 spin_lock_irqsave(&port->lock, flags);
> > 	 if (up->port.irqflags & IRQF_SHARED)
> > 	    disable_irq_nosync(port->irq);
> 
> Yup, this looks suspicious. We set a flag in port.irqflags and take the lock
> only when the flag was set. Either everything needs to be done under
> the lock or the lock is not needed.
> 
> Well, I might have missed something. I do not fully understand meaning
> and relation of all the structures.
> 
> Anyway, I believe that this is a false positive. If I get it correctly
> serial8250_do_startup() must be called before the serial port could
> be registered as a console. It means that it could not be called
> from inside printk().

From my understanding, I'm afraid we are talking about actual deadlock
here, not about false positive report. Quoting the original email:

 : We are trying an S3 suspend stress test and occasionally while
 : entering S3 we get a console deadlock.

[..]

> >  drivers/tty/serial/8250/8250_port.c | 11 +++++++----
> >  1 file changed, 7 insertions(+), 4 deletions(-)
> > 
> > diff --git a/drivers/tty/serial/8250/8250_port.c b/drivers/tty/serial/8250/8250_port.c
> > index d64ca77d9cfa..ad30991e1b3b 100644
> > --- a/drivers/tty/serial/8250/8250_port.c
> > +++ b/drivers/tty/serial/8250/8250_port.c
> > @@ -2275,6 +2275,11 @@ int serial8250_do_startup(struct uart_port *port)
> >  
> >  	if (port->irq && !(up->port.flags & UPF_NO_THRE_TEST)) {
> >  		unsigned char iir1;
> > +		bool irq_shared = up->port.irqflags & IRQF_SHARED;
> > +
> > +		if (irq_shared)
> > +			disable_irq_nosync(port->irq);
> > +
> >  		/*
> >  		 * Test for UARTs that do not reassert THRE when the
> >  		 * transmitter is idle and the interrupt has already
> > @@ -2284,8 +2289,6 @@ int serial8250_do_startup(struct uart_port *port)
> >  		 * allow register changes to become visible.
> >  		 */
> >  		spin_lock_irqsave(&port->lock, flags);
> > -		if (up->port.irqflags & IRQF_SHARED)
> > -			disable_irq_nosync(port->irq);
> >  
> >  		wait_for_xmitr(up, UART_LSR_THRE);
> >  		serial_port_out_sync(port, UART_IER, UART_IER_THRI);
> > @@ -2297,9 +2300,9 @@ int serial8250_do_startup(struct uart_port *port)
> >  		iir = serial_port_in(port, UART_IIR);
> >  		serial_port_out(port, UART_IER, 0);
> >  
> > -		if (port->irqflags & IRQF_SHARED)
> > -			enable_irq(port->irq);
> >  		spin_unlock_irqrestore(&port->lock, flags);
> > +		if (irq_shared)
> > +			enable_irq(port->irq);
> >  
> >  		/*
> >  		 * If the interrupt is not reasserted, or we otherwise
> 
> I think that it might be safe but I am not 100% sure, sigh.

Yeah, I'm not 100%, but I'd give it a try.

	-ss