Re: [PATCH] serial: sifive: Switch to nbcon console

[John Ogness]

Hi Ryo,

On 2025-03-23, Ryo Takakura <ryotkkr98@xxxxxxxxx> wrote:
> Add the necessary callbacks(write_atomic, write_thread, device_lock
> and device_unlock) and CON_NBCON flag to switch the sifive console
> driver to perform as nbcon console.
>
> Both ->write_atomic() and ->write_thread() will check for console
> ownership whenever they are accessing registers.
>
> The ->device_lock()/unlock() will provide the additional serilization
> necessary for ->write_thread() which is called from dedicated printing
> thread.
>
> Signed-off-by: Ryo Takakura <ryotkkr98@xxxxxxxxx>

This driver has the same issue that the 8250 previously had. The
->startup() and ->shutdown() callbacks are called without the port
lock. However, the sifive driver is accessing SIFIVE_SERIAL_IE_OFFS in
these callbacks and this register is also accessed by the ->write()
callback. This needs to be synchronized.

The related 8250 patches fixing this are startup [0] and shutdown [1]. I
am assuming the following change would be sufficient:

diff --git a/drivers/tty/serial/sifive.c b/drivers/tty/serial/sifive.c
index d032de6199af..1de1b2a5833d 100644
--- a/drivers/tty/serial/sifive.c
+++ b/drivers/tty/serial/sifive.c
@@ -564,8 +564,11 @@ static void sifive_serial_break_ctl(struct uart_port *port, int break_state)
 static int sifive_serial_startup(struct uart_port *port)
 {
 	struct sifive_serial_port *ssp = port_to_sifive_serial_port(port);
+	unsigned long flags;
 
+	uart_port_lock_irqsave(&ssp->port, &flags);
 	__ssp_enable_rxwm(ssp);
+	uart_port_unlock_irqrestore(&ssp->port, flags);
 
 	return 0;
 }
@@ -573,9 +576,12 @@ static int sifive_serial_startup(struct uart_port *port)
 static void sifive_serial_shutdown(struct uart_port *port)
 {
 	struct sifive_serial_port *ssp = port_to_sifive_serial_port(port);
+	unsigned long flags;
 
+	uart_port_lock_irqsave(&ssp->port, &flags);
 	__ssp_disable_rxwm(ssp);
 	__ssp_disable_txwm(ssp);
+	uart_port_unlock_irqrestore(&ssp->port, flags);
 }
 
 /**

The fix should be applied first (and likely Cc stable) since it is
fixing an existing mainline problem.

Your patch also needs the synchronization. The ->write_atomic() callback
does not use the port lock. However, the uart_port_*() functions also
take the nbcon console ownership, so they synchronize against
->write_atomic() callbacks.

Otherwise, this patch looks good. If the ->startup() and ->shutdown()
callbacks are fixed in a previous patch, feel free to add:

Reviewed-by: John Ogness <john.ogness@xxxxxxxxxxxxx>

to this patch.

John Ogness

[0] https://lore.kernel.org/lkml/20230525093159.223817-2-john.ogness@xxxxxxxxxxxxx
[1] https://lore.kernel.org/lkml/20230525093159.223817-9-john.ogness@xxxxxxxxxxxxx