On Sun, 26 Jun 2022, Lino Sanfilippo wrote: > On 25.06.22 at 21:58, Lukas Wunner wrote: > > On Wed, Jun 22, 2022 at 05:46:53PM +0200, Lino Sanfilippo wrote: > >> From: Lino Sanfilippo <l.sanfilippo@xxxxxxxxxx> > >> > >> In serial8250_em485_config() the termination GPIO is set with the uart_port > >> spinlock held. This is an issue if setting the GPIO line can sleep (e.g. > >> since the concerning GPIO expander is connected via SPI or I2C). > >> > >> Fix this by setting the termination line outside of the uart_port spinlock > >> in the serial core. > > [...] > >> --- a/drivers/tty/serial/serial_core.c > >> +++ b/drivers/tty/serial/serial_core.c > >> @@ -1400,6 +1411,7 @@ static int uart_set_rs485_config(struct uart_port *port, > >> if (ret) > >> return ret; > >> uart_sanitize_serial_rs485(port, &rs485); > >> + uart_set_rs485_termination(port, &rs485); > >> > >> spin_lock_irqsave(&port->lock, flags); > >> ret = port->rs485_config(port, &rs485); > > > > That's one way to solve the issue. Another would be to push > > acquisition of the port spinlock down into drivers. > > > > I think in most drivers we don't need to take the port spinlock at all > > or only for a few specific register accesses. So taking the lock here > > in the midlayer is likely unwarranted. However, changing that requires > > going through every single driver's ->rs485_config() callback and > > checking whether it needs the lock or not. > > As a first step its sufficient to take the lock in each drivers rs485_config() > function and remove it from uart_set_rs485_config(). Then after time sort out > the drivers that dont require the lock and remove it from their function. > > However the point of this patch was also to generalize the handling of the > termination GPIO, so I would still see this placed in uart_set_rs485_config(). Additional thing to consider is that core currently handles also the port->rs485 assignment under spinlock when ->rs485_config() was successful. TIOCSRS485 ioctl calls are synchronized by other primitives wrt. each other (port mutex and now also termios_rwsem) but drivers probably would like to see consistent rs485 which can only be realized by holding port->lock. -- i.
