diff mbox series

[1/9] serial: core: move RS485 configuration tasks from drivers into core

Message ID 20220213222737.15709-2-LinoSanfilippo@gmx.de
State New
Headers show
Series [1/9] serial: core: move RS485 configuration tasks from drivers into core | expand

Commit Message

Lino Sanfilippo Feb. 13, 2022, 10:27 p.m. UTC
Several drivers that support setting the RS485 configuration via userspace
implement on or more of the following tasks:

- in case of an invalid RTS configuration (both RTS after send and RTS on
  send set or both unset) fall back to enable RTS on send and disable RTS
  after send

- nullify the padding field of the returned serial_rs485 struct

- copy the configuration into the uart port struct

- limit RTS delays to 100 ms

Move these tasks into the serial core to make them generic and to provide
a consistent beheviour among all drivers.

Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
---
 drivers/tty/serial/serial_core.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)


base-commit: ad30d108a5135af584ff47f5ff81be971b6c26f1

Comments

Lino Sanfilippo Feb. 14, 2022, 3:09 p.m. UTC | #1
Hi,

On 14.02.22 at 08:06, Uwe Kleine-König wrote:
> On Sun, Feb 13, 2022 at 11:27:29PM +0100, Lino Sanfilippo wrote:
>> Several drivers that support setting the RS485 configuration via userspace
>> implement on or more of the following tasks:
>
> s/on/one/

Ok

>
>>
>> - in case of an invalid RTS configuration (both RTS after send and RTS on
>>   send set or both unset) fall back to enable RTS on send and disable RTS
>>   after send
>>
>> - nullify the padding field of the returned serial_rs485 struct
>>
>> - copy the configuration into the uart port struct
>>
>> - limit RTS delays to 100 ms
>>
>> Move these tasks into the serial core to make them generic and to provide
>> a consistent beheviour among all drivers.
>
> s/beheviour/behaviour/
>

Ok

>> Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
>> ---
>>  drivers/tty/serial/serial_core.c | 13 +++++++++++++
>>  1 file changed, 13 insertions(+)
>>
>> diff --git a/drivers/tty/serial/serial_core.c b/drivers/tty/serial/serial_core.c
>> index 846192a7b4bf..3fab4070359c 100644
>> --- a/drivers/tty/serial/serial_core.c
>> +++ b/drivers/tty/serial/serial_core.c
>> @@ -1282,8 +1282,21 @@ static int uart_set_rs485_config(struct uart_port *port,
>>  	if (copy_from_user(&rs485, rs485_user, sizeof(*rs485_user)))
>>  		return -EFAULT;
>>
>> +	/* pick sane settings if the user hasn't */
>> +	if (!(rs485.flags & SER_RS485_RTS_ON_SEND) ==
>> +	    !(rs485.flags & SER_RS485_RTS_AFTER_SEND)) {
>> +		rs485.flags |= SER_RS485_RTS_ON_SEND;
>> +		rs485.flags &= ~SER_RS485_RTS_AFTER_SEND;
>> +	}
>> +	/* clamp the delays to [0, 100ms] */
>> +	rs485.delay_rts_before_send = min(rs485.delay_rts_before_send, 100U);
>> +	rs485.delay_rts_after_send = min(rs485.delay_rts_after_send, 100U);
>> +	memset(rs485.padding, 0, sizeof(rs485.padding));
>> +
>>  	spin_lock_irqsave(&port->lock, flags);
>>  	ret = port->rs485_config(port, &rs485);
>> +	if (!ret)
>> +		port->rs485 = rs485;
>
> I was only Cc:d for the imx patch (patch #7) and tried to verify the
> claim there that "the serial core already assigns the passed
> configuration to the uart port". That failed when I looked at my kernel
> tree.
>
> So it would be great, if you sent dependencies (or at least a cover
> letter) to all recipients of a given patch to ease review. Also I want
> to suggest to mention uart_set_rs485_config() in the commit log of the
> imx patch (and probably the others) to simplify verifying the claim
> there.
>

Thanks for the review, I will correct the typos in the next version.
I will also cc you directly for the next version if you dont mind.
get_maintainers only spit out "Pengutronix Kernel Team" so I used that
address for the whole series (including the cover letter).


Regards,
Lino
Lino Sanfilippo Feb. 14, 2022, 4:13 p.m. UTC | #2
Hi,

On 14.02.22 at 06:41, Jiri Slaby wrote:
> On 13. 02. 22, 23:27, Lino Sanfilippo wrote:
>> Several drivers that support setting the RS485 configuration via userspace
>> implement on or more of the following tasks:
>>
>> - in case of an invalid RTS configuration (both RTS after send and RTS on
>>    send set or both unset) fall back to enable RTS on send and disable RTS
>>    after send
>>
>> - nullify the padding field of the returned serial_rs485 struct
>>
>> - copy the configuration into the uart port struct
>>
>> - limit RTS delays to 100 ms
>>
>> Move these tasks into the serial core to make them generic and to provide
>> a consistent beheviour among all drivers.
>>
>> Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
>> ---
>>   drivers/tty/serial/serial_core.c | 13 +++++++++++++
>>   1 file changed, 13 insertions(+)
>>
>> diff --git a/drivers/tty/serial/serial_core.c b/drivers/tty/serial/serial_core.c
>> index 846192a7b4bf..3fab4070359c 100644
>> --- a/drivers/tty/serial/serial_core.c
>> +++ b/drivers/tty/serial/serial_core.c
>> @@ -1282,8 +1282,21 @@ static int uart_set_rs485_config(struct uart_port *port,
>>       if (copy_from_user(&rs485, rs485_user, sizeof(*rs485_user)))
>>           return -EFAULT;
>>   +    /* pick sane settings if the user hasn't */
>> +    if (!(rs485.flags & SER_RS485_RTS_ON_SEND) ==
>> +        !(rs485.flags & SER_RS485_RTS_AFTER_SEND)) {
>> +        rs485.flags |= SER_RS485_RTS_ON_SEND;
>> +        rs485.flags &= ~SER_RS485_RTS_AFTER_SEND;
>> +    }
>> +    /* clamp the delays to [0, 100ms] */
>> +    rs485.delay_rts_before_send = min(rs485.delay_rts_before_send, 100U);
>> +    rs485.delay_rts_after_send = min(rs485.delay_rts_after_send, 100U);
>
> Why is this magic 100?

The only drivers that seem to care about a max value for the RTS delays use 100 ms
(omap-serial, amba pl011, 8250) so I chose this to stay compatible with the current
driver implementations. 100 ms also seems large enough to be used as a general max value.

> Can we have that number somehow documented? You should define a macro for that anyway.

Ok, I will do so.

>
>> +    memset(rs485.padding, 0, sizeof(rs485.padding));
>
> What is this memset good for?

Drivers like max310x, amba-pl011, 8250_pci, 8250_fintek, 8250_lpc18xx seem to care about
returning a serial_rs485 struct with cleared padding field to userspace. So they all clear
that field on their own. Although not really necessary, to me this seems to be a good
default behavior, so I added it to the serial core.

>
>>       spin_lock_irqsave(&port->lock, flags);
>>       ret = port->rs485_config(port, &rs485);
>> +    if (!ret)
>> +        port->rs485 = rs485;
>>       spin_unlock_irqrestore(&port->lock, flags);
>>       if (ret)
>>           return ret;
>
> thanks,

Thanks for the review!

Regards,
Lino
Uwe Kleine-König Feb. 14, 2022, 6:23 p.m. UTC | #3
Hello Lino,

On Mon, Feb 14, 2022 at 04:09:53PM +0100, Lino Sanfilippo wrote:
> On 14.02.22 at 08:06, Uwe Kleine-König wrote:
> > I was only Cc:d for the imx patch (patch #7) and tried to verify the
> > claim there that "the serial core already assigns the passed
> > configuration to the uart port". That failed when I looked at my kernel
> > tree.
> >
> > So it would be great, if you sent dependencies (or at least a cover
> > letter) to all recipients of a given patch to ease review. Also I want
> > to suggest to mention uart_set_rs485_config() in the commit log of the
> > imx patch (and probably the others) to simplify verifying the claim
> > there.
> 
> Thanks for the review, I will correct the typos in the next version.
> I will also cc you directly for the next version if you dont mind.

I don't mind. I get so many patches by mail, I'm good at ignoring them
;-)

> get_maintainers only spit out "Pengutronix Kernel Team" so I used that
> address for the whole series (including the cover letter).

That's why I eventually found the whole series and could reply to patch
#1.

Best regards
Uwe
diff mbox series

Patch

diff --git a/drivers/tty/serial/serial_core.c b/drivers/tty/serial/serial_core.c
index 846192a7b4bf..3fab4070359c 100644
--- a/drivers/tty/serial/serial_core.c
+++ b/drivers/tty/serial/serial_core.c
@@ -1282,8 +1282,21 @@  static int uart_set_rs485_config(struct uart_port *port,
 	if (copy_from_user(&rs485, rs485_user, sizeof(*rs485_user)))
 		return -EFAULT;
 
+	/* pick sane settings if the user hasn't */
+	if (!(rs485.flags & SER_RS485_RTS_ON_SEND) ==
+	    !(rs485.flags & SER_RS485_RTS_AFTER_SEND)) {
+		rs485.flags |= SER_RS485_RTS_ON_SEND;
+		rs485.flags &= ~SER_RS485_RTS_AFTER_SEND;
+	}
+	/* clamp the delays to [0, 100ms] */
+	rs485.delay_rts_before_send = min(rs485.delay_rts_before_send, 100U);
+	rs485.delay_rts_after_send = min(rs485.delay_rts_after_send, 100U);
+	memset(rs485.padding, 0, sizeof(rs485.padding));
+
 	spin_lock_irqsave(&port->lock, flags);
 	ret = port->rs485_config(port, &rs485);
+	if (!ret)
+		port->rs485 = rs485;
 	spin_unlock_irqrestore(&port->lock, flags);
 	if (ret)
 		return ret;