Re: [PATCH] serial: 8250_of: set UART_CAP_NOTEMT for rts-gpios RS485 direction

[Greg KH <gregkh@linuxfoundation.org>]

On Mon, Sep 07, 2026 at 07:30:00PM +0200, Nicolas Thibert wrote:
> __stop_tx() (8250_port.c) only calls the RS485 rs485_stop_tx() hook
> (which de-asserts the direction GPIO/RTS line) once it has observed
> both UART_LSR_THRE and UART_LSR_TEMT for the last byte. If TEMT is
> never seen and the driver hasn't set UART_CAP_NOTEMT, the function
> returns without scheduling any retry -- the direction line is left
> asserted (driver enabled) forever, with nothing to un-stick it short
> of another kernel-visible LSR event.
> 
> of_platform_serial_setup()/of_platform_serial_probe() unconditionally
> wire up the generic em485 GPIO-RTS RS485 support
> (rs485_config/rs485_start_tx/rs485_stop_tx) for every port they
> register, but never set UART_CAP_NOTEMT, so any board using this
> driver whose 16550-compatible core doesn't reliably surface TEMT for
> its shift register hits the stuck-direction-GPIO case above.
> 
> Confirmed live on an ath79 QCA9531 board (SoC-internal ns16550a-
> compatible UART, RS485 transceiver DE/RE tied together on a GPIO via
> rts-gpios, linux,rs485-enabled-at-boot-time): the direction GPIO
> correctly asserts for the duration of a transmit, but never
> de-asserts afterwards -- confirmed by sampling the GPIO's debugfs
> state through and after a multi-hundred-byte write, on both the first
> transmit and repeated back-to-back transmits. Setting
> UART_CAP_NOTEMT, which makes __stop_tx() fall back to a frame-time-
> based timer instead of waiting indefinitely on TEMT, makes the
> direction GPIO reliably return low right after each transmit
> completes.
> 
> Scope the fix to ports that declare a GPIO-controlled direction line
> (rts-gpios), rather than setting it unconditionally for every port
> this driver registers: this is the class of hardware actually
> affected (RTS state has to be explicitly un-stuck by software, unlike
> a UART's native RTS pin), and it avoids adding the extra frame-time
> margin to ports relying on the native RTS pin, which this has not
> been observed to need.
> 
> Set it in of_platform_serial_probe(), after the existing
> "if (port8250.port.fifosize) port8250.capabilities = UART_CAP_FIFO;"
> assignment, rather than in of_platform_serial_setup(): that later
> plain assignment (not "|=") unconditionally overwrites
> port8250.capabilities on any port whose "fifo-size" DT property is
> set, silently discarding a capability bit set earlier in setup(). Not
> observed on the reporter's own board (it has no "fifo-size" property,
> so port.fifosize stays 0 and that assignment is skipped), but a real
> regression waiting to happen on any other of_platform_serial user
> that does set "fifo-size" -- which is a common, documented property
> for this binding.

Please rewrite this to be in english and says things properly.

> 
> Signed-off-by: Nicolas Thibert <nithibert@gmail.com>
> Assisted-by: LLM (Claude Sonnet 5, Anthropic)

signed-off-by goes last.

> ---
>  drivers/tty/serial/8250/8250_of.c | 18 ++++++++++++++++++
>  1 file changed, 18 insertions(+)
> 
> --- a/drivers/tty/serial/8250/8250_of.c
> +++ b/drivers/tty/serial/8250/8250_of.c
> @@ -233,6 +233,24 @@ static int of_platform_serial_probe(struct platform_device *ofdev)
>  			&port8250.overrun_backoff_time_ms) != 0)
>  		port8250.overrun_backoff_time_ms = 0;
> 
> +	/*
> +	 * This generic driver never enables a dedicated line-status
> +	 * interrupt on TEMT, so for ports whose RS485 direction is
> +	 * controlled via a GPIO (rts-gpios) rather than the native RTS
> +	 * pin, __stop_tx() (8250_port.c) can see THRE without TEMT on the
> +	 * last byte and bail out without ever retrying -- leaving the
> +	 * direction GPIO stuck asserted after the last byte sent, on
> +	 * hardware whose shift register doesn't reliably surface TEMT.
> +	 * UART_CAP_NOTEMT makes it fall back to a frame-time-based timer
> +	 * instead of waiting on that interrupt. Scoped to rts-gpios users
> +	 * only, to avoid changing timing for ports relying on the native
> +	 * RTS pin, which this has not been observed to affect. Set here,
> +	 * after the fifosize-based capabilities assignment above, so it
> +	 * isn't clobbered by it.
> +	 */

Is this comment really needed?


> +	if (of_property_present(ofdev->dev.of_node, "rts-gpios"))

What will change from this now on existing systems?

How was this tested?

And why was this sent twice?

thanks,

greg k-h