[v3,13/14] serial: liteuart: add IRQ support for the TX path

  Modify the TX path to operate in an IRQ-compatible way, while
maintaining support for polling mode via the poll timer.

Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
---
 drivers/tty/serial/liteuart.c | 67 ++++++++++++++++++++++++-----------
 1 file changed, 47 insertions(+), 20 deletions(-)
  

On Sat, Nov 12, 2022 at 04:21:24PM -0500, Gabriel Somlo wrote:
> Modify the TX path to operate in an IRQ-compatible way, while
> maintaining support for polling mode via the poll timer.
> 
> Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
> ---
>  drivers/tty/serial/liteuart.c | 67 ++++++++++++++++++++++++-----------
>  1 file changed, 47 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/tty/serial/liteuart.c b/drivers/tty/serial/liteuart.c
> index e30adb30277f..307c27398e30 100644
> --- a/drivers/tty/serial/liteuart.c
> +++ b/drivers/tty/serial/liteuart.c
> @@ -46,6 +46,7 @@ struct liteuart_port {
>  	struct uart_port port;
>  	struct timer_list timer;
>  	u32 id;
> +	bool poll_tx_started;
>  };
>  
>  #define to_liteuart_port(port)	container_of(port, struct liteuart_port, port)
> @@ -78,29 +79,24 @@ static void liteuart_putchar(struct uart_port *port, unsigned char ch)
>  
>  static void liteuart_stop_tx(struct uart_port *port)
>  {
> -	/* not used in LiteUART, but called unconditionally from serial_core */
> +	if (port->irq) {
> +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask & ~EV_TX);
> +	} else {
> +		struct liteuart_port *uart = to_liteuart_port(port);
> +		uart->poll_tx_started = false;
> +	}
>  }
>  
>  static void liteuart_start_tx(struct uart_port *port)
>  {
> -	struct circ_buf *xmit = &port->state->xmit;
> -	unsigned char ch;
> -
> -	if (unlikely(port->x_char)) {
> -		litex_write8(port->membase + OFF_RXTX, port->x_char);
> -		port->icount.tx++;
> -		port->x_char = 0;
> -	} else if (!uart_circ_empty(xmit)) {
> -		while (xmit->head != xmit->tail) {
> -			ch = xmit->buf[xmit->tail];

I just realized that this:

> -			xmit->tail = (xmit->tail + 1) & (UART_XMIT_SIZE - 1);
> -			port->icount.tx++;

... conflicts with another accepted patch (by Ilpo) that's currently
making its way to being released:
https://lore.kernel.org/all/20221019091151.6692-20-ilpo.jarvinen@linux.intel.com/

> -			liteuart_putchar(port, ch);
> -		}
> +	if (port->irq) {
> +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask | EV_TX);
> +	} else {
> +		struct liteuart_port *uart = to_liteuart_port(port);
> +		uart->poll_tx_started = true;
>  	}
> -
> -	if (uart_circ_chars_pending(xmit) < WAKEUP_CHARS)
> -		uart_write_wakeup(port);
>  }
>  
>  static void liteuart_stop_rx(struct uart_port *port)
> @@ -131,18 +127,47 @@ static void liteuart_rx_chars(struct uart_port *port)
>  	tty_flip_buffer_push(&port->state->port);
>  }
>  
> +static void liteuart_tx_chars(struct uart_port *port)
> +{
> +	struct circ_buf *xmit = &port->state->xmit;
> +
> +	if (unlikely(port->x_char)) {
> +		litex_write8(port->membase + OFF_RXTX, port->x_char);
> +		port->x_char = 0;
> +		port->icount.tx++;
> +		return;
> +	}
> +
> +	while (!litex_read8(port->membase + OFF_TXFULL)) {
> +		if (xmit->head == xmit->tail)
> +			break;
> +		litex_write8(port->membase + OFF_RXTX, xmit->buf[xmit->tail]);

Also, this:

> +		xmit->tail = (xmit->tail + 1) & (UART_XMIT_SIZE - 1);
> +		port->icount.tx++;

should now be `uart_xmit_advance(port, 1);` instead.
I'm going to take that into account in v4, in addition to any extra
feedback I'll receive.

Thanks!

> +	}
> +
> +	if (uart_circ_chars_pending(xmit) < WAKEUP_CHARS)
> +		uart_write_wakeup(port);
> +
> +	if (uart_circ_empty(xmit))
> +		liteuart_stop_tx(port);
> +}
> +
>  static irqreturn_t liteuart_interrupt(int irq, void *data)
>  {
>  	struct liteuart_port *uart = data;
>  	struct uart_port *port = &uart->port;
>  	u8 isr = litex_read8(port->membase + OFF_EV_PENDING);
>  
> -	/* for now, only rx path triggers interrupts */
> -	isr &= EV_RX;
> +	if (!(port->irq || uart->poll_tx_started))
> +		isr &= ~EV_TX;	/* polling mode with tx stopped */
>  
>  	spin_lock(&port->lock);
>  	if (isr & EV_RX)
>  		liteuart_rx_chars(port);
> +	if (isr & EV_TX) {
> +		liteuart_tx_chars(port);
> +	}
>  	spin_unlock(&port->lock);
>  
>  	return IRQ_RETVAL(isr);
> @@ -196,6 +221,7 @@ static int liteuart_startup(struct uart_port *port)
>  	}
>  
>  	if (!port->irq) {
> +		uart->poll_tx_started = false;
>  		timer_setup(&uart->timer, liteuart_timer, 0);
>  		mod_timer(&uart->timer, jiffies + uart_poll_timeout(port));
>  	}
> @@ -210,6 +236,7 @@ static void liteuart_shutdown(struct uart_port *port)
>  	struct liteuart_port *uart = to_liteuart_port(port);
>  
>  	litex_write8(port->membase + OFF_EV_ENABLE, 0);
> +	uart->poll_tx_started = false;
>  
>  	if (port->irq)
>  		free_irq(port->irq, port);
> -- 
> 2.37.3
>
  

On Sat, 12 Nov 2022, Gabriel Somlo wrote:

> Modify the TX path to operate in an IRQ-compatible way, while
> maintaining support for polling mode via the poll timer.
> 
> Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
> ---
>  drivers/tty/serial/liteuart.c | 67 ++++++++++++++++++++++++-----------
>  1 file changed, 47 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/tty/serial/liteuart.c b/drivers/tty/serial/liteuart.c
> index e30adb30277f..307c27398e30 100644
> --- a/drivers/tty/serial/liteuart.c
> +++ b/drivers/tty/serial/liteuart.c
> @@ -46,6 +46,7 @@ struct liteuart_port {
>  	struct uart_port port;
>  	struct timer_list timer;
>  	u32 id;
> +	bool poll_tx_started;
>  };
>  
>  #define to_liteuart_port(port)	container_of(port, struct liteuart_port, port)
> @@ -78,29 +79,24 @@ static void liteuart_putchar(struct uart_port *port, unsigned char ch)
>  
>  static void liteuart_stop_tx(struct uart_port *port)
>  {
> -	/* not used in LiteUART, but called unconditionally from serial_core */

Drop this comment from the earlier patch too given you remove it here. It 
just adds useless churn in diff for no useful reason.

> +	if (port->irq) {
> +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask & ~EV_TX);

If you put irq_mask into liteuart_port you wouldn't need to read it 
back here?

> +	} else {
> +		struct liteuart_port *uart = to_liteuart_port(port);
> +		uart->poll_tx_started = false;
> +	}
>  }
>  
>  static void liteuart_start_tx(struct uart_port *port)
>  {
> -	struct circ_buf *xmit = &port->state->xmit;
> -	unsigned char ch;
> -
> -	if (unlikely(port->x_char)) {
> -		litex_write8(port->membase + OFF_RXTX, port->x_char);
> -		port->icount.tx++;
> -		port->x_char = 0;
> -	} else if (!uart_circ_empty(xmit)) {
> -		while (xmit->head != xmit->tail) {
> -			ch = xmit->buf[xmit->tail];
> -			xmit->tail = (xmit->tail + 1) & (UART_XMIT_SIZE - 1);
> -			port->icount.tx++;

This is not based on tty-next tree. Please rebase on top of it (you 
might have noted it already, IIRC, somebody noted uart_xmit_advance
conflict in some patch, perhaps it was you :-)).

> -			liteuart_putchar(port, ch);
> -		}
> +	if (port->irq) {
> +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask | EV_TX);

->irq_mask?

> +	} else {
> +		struct liteuart_port *uart = to_liteuart_port(port);
> +		uart->poll_tx_started = true;
>  	}
> -
> -	if (uart_circ_chars_pending(xmit) < WAKEUP_CHARS)
> -		uart_write_wakeup(port);
>  }
>  
>  static void liteuart_stop_rx(struct uart_port *port)
> @@ -131,18 +127,47 @@ static void liteuart_rx_chars(struct uart_port *port)
>  	tty_flip_buffer_push(&port->state->port);
>  }
>  
> +static void liteuart_tx_chars(struct uart_port *port)
> +{
> +	struct circ_buf *xmit = &port->state->xmit;
> +
> +	if (unlikely(port->x_char)) {
> +		litex_write8(port->membase + OFF_RXTX, port->x_char);
> +		port->x_char = 0;
> +		port->icount.tx++;
> +		return;
> +	}
> +
> +	while (!litex_read8(port->membase + OFF_TXFULL)) {
> +		if (xmit->head == xmit->tail)

There exists a helper for this condition.

> +			break;
> +		litex_write8(port->membase + OFF_RXTX, xmit->buf[xmit->tail]);
> +		xmit->tail = (xmit->tail + 1) & (UART_XMIT_SIZE - 1);
> +		port->icount.tx++;

uart_xmit_advance()

> +	}
> +
> +	if (uart_circ_chars_pending(xmit) < WAKEUP_CHARS)
> +		uart_write_wakeup(port);
> +
> +	if (uart_circ_empty(xmit))
> +		liteuart_stop_tx(port);
> +}

You might want to check if you can generate this whole function with 
Jiri's tx helpers (IIRC, they're only in tty-next tree currently).

> +
>  static irqreturn_t liteuart_interrupt(int irq, void *data)
>  {
>  	struct liteuart_port *uart = data;
>  	struct uart_port *port = &uart->port;
>  	u8 isr = litex_read8(port->membase + OFF_EV_PENDING);
>  
> -	/* for now, only rx path triggers interrupts */
> -	isr &= EV_RX;
> +	if (!(port->irq || uart->poll_tx_started))
> +		isr &= ~EV_TX;	/* polling mode with tx stopped */
>  
>  	spin_lock(&port->lock);
>  	if (isr & EV_RX)
>  		liteuart_rx_chars(port);
> +	if (isr & EV_TX) {
> +		liteuart_tx_chars(port);
> +	}

Extra braces.

>  	spin_unlock(&port->lock);
>  
>  	return IRQ_RETVAL(isr);
> @@ -196,6 +221,7 @@ static int liteuart_startup(struct uart_port *port)
>  	}
>  
>  	if (!port->irq) {
> +		uart->poll_tx_started = false;

Can poll_tx_started ever be true here?

>  		timer_setup(&uart->timer, liteuart_timer, 0);
>  		mod_timer(&uart->timer, jiffies + uart_poll_timeout(port));
>  	}
> @@ -210,6 +236,7 @@ static void liteuart_shutdown(struct uart_port *port)
>  	struct liteuart_port *uart = to_liteuart_port(port);
>  
>  	litex_write8(port->membase + OFF_EV_ENABLE, 0);
> +	uart->poll_tx_started = false;
>  
>  	if (port->irq)
>  		free_irq(port->irq, port);
>
  

On Tue, Nov 15, 2022 at 06:14:50PM +0200, Ilpo Järvinen wrote:
> On Sat, 12 Nov 2022, Gabriel Somlo wrote:
> 
> > Modify the TX path to operate in an IRQ-compatible way, while
> > maintaining support for polling mode via the poll timer.
> > 
> > Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
> > ---
> >  drivers/tty/serial/liteuart.c | 67 ++++++++++++++++++++++++-----------
> >  1 file changed, 47 insertions(+), 20 deletions(-)
> > 
> > diff --git a/drivers/tty/serial/liteuart.c b/drivers/tty/serial/liteuart.c
> > index e30adb30277f..307c27398e30 100644
> > --- a/drivers/tty/serial/liteuart.c
> > +++ b/drivers/tty/serial/liteuart.c
> > @@ -46,6 +46,7 @@ struct liteuart_port {
> >  	struct uart_port port;
> >  	struct timer_list timer;
> >  	u32 id;
> > +	bool poll_tx_started;
> >  };
> >  
> >  #define to_liteuart_port(port)	container_of(port, struct liteuart_port, port)
> > @@ -78,29 +79,24 @@ static void liteuart_putchar(struct uart_port *port, unsigned char ch)
> >  
> >  static void liteuart_stop_tx(struct uart_port *port)
> >  {
> > -	/* not used in LiteUART, but called unconditionally from serial_core */
> 
> Drop this comment from the earlier patch too given you remove it here. It 
> just adds useless churn in diff for no useful reason.

Right -- I already had this lined up for v4 :)
 
> > +	if (port->irq) {
> > +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> > +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask & ~EV_TX);
> 
> If you put irq_mask into liteuart_port you wouldn't need to read it 
> back here?

So, instead of `bool poll_tx_started` I should just keep a copy of the
irq_mask there, and take `&port->lock` whenever I need to *both* update
the mask *and* write it out to the actual device register?

> > +	} else {
> > +		struct liteuart_port *uart = to_liteuart_port(port);
> > +		uart->poll_tx_started = false;
> > +	}
> >  }
> >  
> >  static void liteuart_start_tx(struct uart_port *port)
> >  {
> > -	struct circ_buf *xmit = &port->state->xmit;
> > -	unsigned char ch;
> > -
> > -	if (unlikely(port->x_char)) {
> > -		litex_write8(port->membase + OFF_RXTX, port->x_char);
> > -		port->icount.tx++;
> > -		port->x_char = 0;
> > -	} else if (!uart_circ_empty(xmit)) {
> > -		while (xmit->head != xmit->tail) {
> > -			ch = xmit->buf[xmit->tail];
> > -			xmit->tail = (xmit->tail + 1) & (UART_XMIT_SIZE - 1);
> > -			port->icount.tx++;
> 
> This is not based on tty-next tree. Please rebase on top of it (you 
> might have noted it already, IIRC, somebody noted uart_xmit_advance
> conflict in some patch, perhaps it was you :-)).

Yeah, I did notice that right after I sent out v3. I've already
rebased it on top of your patch using uart_xmit_advance.

> > -			liteuart_putchar(port, ch);
> > -		}
> > +	if (port->irq) {
> > +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> > +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask | EV_TX);
> 
> ->irq_mask?

I'll switch to s/bool poll_tx_started/u8 irq_mask/g in v4, hopefully
it should make it all look cleaner.

> > +	} else {
> > +		struct liteuart_port *uart = to_liteuart_port(port);
> > +		uart->poll_tx_started = true;
> >  	}
> > -
> > -	if (uart_circ_chars_pending(xmit) < WAKEUP_CHARS)
> > -		uart_write_wakeup(port);
> >  }
> >  
> >  static void liteuart_stop_rx(struct uart_port *port)
> > @@ -131,18 +127,47 @@ static void liteuart_rx_chars(struct uart_port *port)
> >  	tty_flip_buffer_push(&port->state->port);
> >  }
> >  
> > +static void liteuart_tx_chars(struct uart_port *port)
> > +{
> > +	struct circ_buf *xmit = &port->state->xmit;
> > +
> > +	if (unlikely(port->x_char)) {
> > +		litex_write8(port->membase + OFF_RXTX, port->x_char);
> > +		port->x_char = 0;
> > +		port->icount.tx++;
> > +		return;
> > +	}
> > +
> > +	while (!litex_read8(port->membase + OFF_TXFULL)) {
> > +		if (xmit->head == xmit->tail)
> 
> There exists a helper for this condition.

Is that in the released linus tree, or still only in tty-next?

> > +			break;
> > +		litex_write8(port->membase + OFF_RXTX, xmit->buf[xmit->tail]);
> > +		xmit->tail = (xmit->tail + 1) & (UART_XMIT_SIZE - 1);
> > +		port->icount.tx++;
> 
> uart_xmit_advance()

Already lined up for v4.

> 
> > +	}
> > +
> > +	if (uart_circ_chars_pending(xmit) < WAKEUP_CHARS)
> > +		uart_write_wakeup(port);
> > +
> > +	if (uart_circ_empty(xmit))
> > +		liteuart_stop_tx(port);
> > +}
> 
> You might want to check if you can generate this whole function with 
> Jiri's tx helpers (IIRC, they're only in tty-next tree currently).

Looks like I should switch to tty-next for this whole series, which
makes sense, since it's a tty I'm working on :)

I'll rebase on top of that before I send out v4, hopefully later this
afternoon.
 
> > +
> >  static irqreturn_t liteuart_interrupt(int irq, void *data)
> >  {
> >  	struct liteuart_port *uart = data;
> >  	struct uart_port *port = &uart->port;
> >  	u8 isr = litex_read8(port->membase + OFF_EV_PENDING);
> >  
> > -	/* for now, only rx path triggers interrupts */
> > -	isr &= EV_RX;
> > +	if (!(port->irq || uart->poll_tx_started))
> > +		isr &= ~EV_TX;	/* polling mode with tx stopped */
> >  
> >  	spin_lock(&port->lock);
> >  	if (isr & EV_RX)
> >  		liteuart_rx_chars(port);
> > +	if (isr & EV_TX) {
> > +		liteuart_tx_chars(port);
> > +	}
> 
> Extra braces.

Got it, thanks!

> >  	spin_unlock(&port->lock);
> >  
> >  	return IRQ_RETVAL(isr);
> > @@ -196,6 +221,7 @@ static int liteuart_startup(struct uart_port *port)
> >  	}
> >  
> >  	if (!port->irq) {
> > +		uart->poll_tx_started = false;
> 
> Can poll_tx_started ever be true here?

Proably not, but it shouldn't matter if I switch to using `u8 irq_mask`,
instead, which should be initialized to 0 during probe().

Thanks again for the feedback!

Best,
--Gabriel

> >  		timer_setup(&uart->timer, liteuart_timer, 0);
> >  		mod_timer(&uart->timer, jiffies + uart_poll_timeout(port));
> >  	}
> > @@ -210,6 +236,7 @@ static void liteuart_shutdown(struct uart_port *port)
> >  	struct liteuart_port *uart = to_liteuart_port(port);
> >  
> >  	litex_write8(port->membase + OFF_EV_ENABLE, 0);
> > +	uart->poll_tx_started = false;
> >  
> >  	if (port->irq)
> >  		free_irq(port->irq, port);
> > 
> 
> -- 
>  i.
>
  

On Tue, 15 Nov 2022, Gabriel L. Somlo wrote:

> On Tue, Nov 15, 2022 at 06:14:50PM +0200, Ilpo Järvinen wrote:
> > On Sat, 12 Nov 2022, Gabriel Somlo wrote:
> > 
> > > Modify the TX path to operate in an IRQ-compatible way, while
> > > maintaining support for polling mode via the poll timer.
> > > 
> > > Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
> > > ---
> > >  drivers/tty/serial/liteuart.c | 67 ++++++++++++++++++++++++-----------
> > >  1 file changed, 47 insertions(+), 20 deletions(-)
> > > 
> > > diff --git a/drivers/tty/serial/liteuart.c b/drivers/tty/serial/liteuart.c
> > > index e30adb30277f..307c27398e30 100644
> > > --- a/drivers/tty/serial/liteuart.c
> > > +++ b/drivers/tty/serial/liteuart.c

> > > +	if (port->irq) {
> > > +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> > > +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask & ~EV_TX);
> > 
> > If you put irq_mask into liteuart_port you wouldn't need to read it 
> > back here?
> 
> So, instead of `bool poll_tx_started` I should just keep a copy of the
> irq_mask there, and take `&port->lock` whenever I need to *both* update
> the mask *and* write it out to the actual device register?

I was mostly thinking of storing EV_RX there but then it could be derived 
from port->irq that is checked by all paths already.

> > > +	if (unlikely(port->x_char)) {
> > > +		litex_write8(port->membase + OFF_RXTX, port->x_char);
> > > +		port->x_char = 0;
> > > +		port->icount.tx++;
> > > +		return;
> > > +	}
> > > +
> > > +	while (!litex_read8(port->membase + OFF_TXFULL)) {
> > > +		if (xmit->head == xmit->tail)
> > 
> > There exists a helper for this condition.
> 
> Is that in the released linus tree, or still only in tty-next?

uart_circ_empty() has been around for ages.

> > > +	if (uart_circ_chars_pending(xmit) < WAKEUP_CHARS)
> > > +		uart_write_wakeup(port);
> > > +
> > > +	if (uart_circ_empty(xmit))
> > > +		liteuart_stop_tx(port);
> > > +}
> > 
> > You might want to check if you can generate this whole function with 
> > Jiri's tx helpers (IIRC, they're only in tty-next tree currently).
> 
> Looks like I should switch to tty-next for this whole series, which
> makes sense, since it's a tty I'm working on :)
> 
> I'll rebase on top of that before I send out v4, hopefully later this
> afternoon.

Ok.

As I now looked it up, Jiri's tx helpers is 
8275b48b278096edc1e3ea5aa9cf946a10022f79 and you'll find some example 
conversions in the changes following it.
  

On Tue, Nov 15, 2022 at 07:30:09PM +0200, Ilpo Järvinen wrote:
> On Tue, 15 Nov 2022, Gabriel L. Somlo wrote:
> 
> > On Tue, Nov 15, 2022 at 06:14:50PM +0200, Ilpo Järvinen wrote:
> > > On Sat, 12 Nov 2022, Gabriel Somlo wrote:
> > > 
> > > > Modify the TX path to operate in an IRQ-compatible way, while
> > > > maintaining support for polling mode via the poll timer.
> > > > 
> > > > Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
> > > > ---
> > > >  drivers/tty/serial/liteuart.c | 67 ++++++++++++++++++++++++-----------
> > > >  1 file changed, 47 insertions(+), 20 deletions(-)
> > > > 
> > > > diff --git a/drivers/tty/serial/liteuart.c b/drivers/tty/serial/liteuart.c
> > > > index e30adb30277f..307c27398e30 100644
> > > > --- a/drivers/tty/serial/liteuart.c
> > > > +++ b/drivers/tty/serial/liteuart.c
> 
> > > > +	if (port->irq) {
> > > > +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> > > > +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask & ~EV_TX);
> > > 
> > > If you put irq_mask into liteuart_port you wouldn't need to read it 
> > > back here?
> > 
> > So, instead of `bool poll_tx_started` I should just keep a copy of the
> > irq_mask there, and take `&port->lock` whenever I need to *both* update
> > the mask *and* write it out to the actual device register?
> 
> I was mostly thinking of storing EV_RX there but then it could be derived 
> from port->irq that is checked by all paths already.

I'm a bit confused here -- the reason I was reading in irq_mask from
the mmio port and then flipping the EV_TX bit before writing it back
out was to preserve the current value of the EV_RX bit in that
register, whatever it may happen to be at the time. If I shadowed the
entire register value (with both EV_RX and EV_TX bits reflecting their
current value), I could get away with only ever *writing* the MMIO
register each time the shadow register value is modified (one or both
of the bits are flipped), and only when port->irq is non-zero (i.e.,
the shadow register would work for polling-mode also).

I think that's how altera_uart.c is doing it (I've been using that as
a source of inspiration).

I thought that's what you were suggesting earlie, but based on your
response above I'm no longer sure I follow.

> > > > +	if (unlikely(port->x_char)) {
> > > > +		litex_write8(port->membase + OFF_RXTX, port->x_char);
> > > > +		port->x_char = 0;
> > > > +		port->icount.tx++;
> > > > +		return;
> > > > +	}
> > > > +
> > > > +	while (!litex_read8(port->membase + OFF_TXFULL)) {
> > > > +		if (xmit->head == xmit->tail)
> > > 
> > > There exists a helper for this condition.
> > 
> > Is that in the released linus tree, or still only in tty-next?
> 
> uart_circ_empty() has been around for ages.

Oh, I misunderstood what you meant here originally -- just use
`uart_circ_empty()` like I'm already doing elsewhere in the code, got
it!

> > > > +	if (uart_circ_chars_pending(xmit) < WAKEUP_CHARS)
> > > > +		uart_write_wakeup(port);
> > > > +
> > > > +	if (uart_circ_empty(xmit))
> > > > +		liteuart_stop_tx(port);
> > > > +}
> > > 
> > > You might want to check if you can generate this whole function with 
> > > Jiri's tx helpers (IIRC, they're only in tty-next tree currently).
> > 
> > Looks like I should switch to tty-next for this whole series, which
> > makes sense, since it's a tty I'm working on :)
> > 
> > I'll rebase on top of that before I send out v4, hopefully later this
> > afternoon.
> 
> Ok.
> 
> As I now looked it up, Jiri's tx helpers is 
> 8275b48b278096edc1e3ea5aa9cf946a10022f79 and you'll find some example 
> conversions in the changes following it.

I'll check that out and follow the examples.

Thanks again,
--G
  

On Tue, Nov 15, 2022 at 07:30:09PM +0200, Ilpo Järvinen wrote:
> On Tue, 15 Nov 2022, Gabriel L. Somlo wrote:
> 
> > On Tue, Nov 15, 2022 at 06:14:50PM +0200, Ilpo Järvinen wrote:
> > > On Sat, 12 Nov 2022, Gabriel Somlo wrote:
> > > 
> > > > Modify the TX path to operate in an IRQ-compatible way, while
> > > > maintaining support for polling mode via the poll timer.
> > > > 
> > > > Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
> > > > ---
> > > >  drivers/tty/serial/liteuart.c | 67 ++++++++++++++++++++++++-----------
> > > >  1 file changed, 47 insertions(+), 20 deletions(-)
> > > > 
> > > > diff --git a/drivers/tty/serial/liteuart.c b/drivers/tty/serial/liteuart.c
> > > > index e30adb30277f..307c27398e30 100644
> > > > --- a/drivers/tty/serial/liteuart.c
> > > > +++ b/drivers/tty/serial/liteuart.c
> 
> > > > +	if (port->irq) {
> > > > +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> > > > +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask & ~EV_TX);
> > > 
> > > If you put irq_mask into liteuart_port you wouldn't need to read it 
> > > back here?
> > 
> > So, instead of `bool poll_tx_started` I should just keep a copy of the
> > irq_mask there, and take `&port->lock` whenever I need to *both* update
> > the mask *and* write it out to the actual device register?
> 
> I was mostly thinking of storing EV_RX there but then it could be derived 
> from port->irq that is checked by all paths already.

So, just to clear up the best course of action here (before I rebase
everything on top of tty-next: How about the patch below (currently
applied separately at the end of the entire series, but I can respin
it as part of the rx path (12/14) and tx path (13/14) as appropriate.

Let me know what you think.

Thanks,
--Gabriel


diff --git a/drivers/tty/serial/liteuart.c b/drivers/tty/serial/liteuart.c
index eb0ae8c8bd94..185db423c65f 100644
--- a/drivers/tty/serial/liteuart.c
+++ b/drivers/tty/serial/liteuart.c
@@ -47,7 +47,7 @@ struct liteuart_port {
 	struct uart_port port;
 	struct timer_list timer;
 	u32 id;
-	bool poll_tx_started;
+	u8 irq_reg;
 };
 
 #define to_liteuart_port(port)	container_of(port, struct liteuart_port, port)
@@ -70,26 +70,27 @@ static struct uart_driver liteuart_driver = {
 #endif
 };
 
+static void liteuart_update_irq_reg(struct uart_port *port, bool set, u8 bits)
+{
+	struct liteuart_port *uart = to_liteuart_port(port);
+
+	if (set)
+		uart->irq_reg |= bits;
+	else
+		uart->irq_reg &= ~bits;
+
+	if (port->irq)
+		litex_write8(port->membase + OFF_EV_ENABLE, uart->irq_reg);
+}
+
 static void liteuart_stop_tx(struct uart_port *port)
 {
-	if (port->irq) {
-		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
-		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask & ~EV_TX);
-	} else {
-		struct liteuart_port *uart = to_liteuart_port(port);
-		uart->poll_tx_started = false;
-	}
+	liteuart_update_irq_reg(port, false, EV_TX);
 }
 
 static void liteuart_start_tx(struct uart_port *port)
 {
-	if (port->irq) {
-		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
-		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask | EV_TX);
-	} else {
-		struct liteuart_port *uart = to_liteuart_port(port);
-		uart->poll_tx_started = true;
-	}
+	liteuart_update_irq_reg(port, true, EV_TX);
 }
 
 static void liteuart_stop_rx(struct uart_port *port)
@@ -149,12 +150,10 @@ static irqreturn_t liteuart_interrupt(int irq, void *data)
 {
 	struct liteuart_port *uart = data;
 	struct uart_port *port = &uart->port;
-	u8 isr = litex_read8(port->membase + OFF_EV_PENDING);
-
-	if (!(port->irq || uart->poll_tx_started))
-		isr &= ~EV_TX;	/* polling mode with tx stopped */
+	u8 isr;
 
 	spin_lock(&port->lock);
+	isr = litex_read8(port->membase + OFF_EV_PENDING) & uart->irq_reg;
 	if (isr & EV_RX)
 		liteuart_rx_chars(port);
 	if (isr & EV_TX)
@@ -195,39 +194,40 @@ static unsigned int liteuart_get_mctrl(struct uart_port *port)
 static int liteuart_startup(struct uart_port *port)
 {
 	struct liteuart_port *uart = to_liteuart_port(port);
+	unsigned long flags;
 	int ret;
-	u8 irq_mask = 0;
 
 	if (port->irq) {
 		ret = request_irq(port->irq, liteuart_interrupt, 0,
 				  KBUILD_MODNAME, uart);
-		if (ret == 0) {
-			/* only enabling rx interrupts at startup */
-			irq_mask = EV_RX;
-		} else {
+		if (ret) {
 			pr_err(pr_fmt("line %d irq %d failed: using polling\n"),
 				port->line, port->irq);
 			port->irq = 0;
 		}
 	}
 
+	spin_lock_irqsave(&port->lock, flags);
+	/* only enabling rx irqs during startup */
+	liteuart_update_irq_reg(port, true, EV_RX);
+	spin_unlock_irqrestore(&port->lock, flags);
+
 	if (!port->irq) {
-		uart->poll_tx_started = false;
 		timer_setup(&uart->timer, liteuart_timer, 0);
 		mod_timer(&uart->timer, jiffies + uart_poll_timeout(port));
 	}
 
-	litex_write8(port->membase + OFF_EV_ENABLE, irq_mask);
-
 	return 0;
 }
 
 static void liteuart_shutdown(struct uart_port *port)
 {
 	struct liteuart_port *uart = to_liteuart_port(port);
+	unsigned long flags;
 
-	litex_write8(port->membase + OFF_EV_ENABLE, 0);
-	uart->poll_tx_started = false;
+	spin_lock_irqsave(&port->lock, flags);
+	liteuart_update_irq_reg(port, false, EV_RX | EV_TX);
+	spin_unlock_irqrestore(&port->lock, flags);
 
 	if (port->irq)
 		free_irq(port->irq, port);
  

On Tue, 15 Nov 2022, Gabriel L. Somlo wrote:

> On Tue, Nov 15, 2022 at 07:30:09PM +0200, Ilpo Järvinen wrote:
> > On Tue, 15 Nov 2022, Gabriel L. Somlo wrote:
> > 
> > > On Tue, Nov 15, 2022 at 06:14:50PM +0200, Ilpo Järvinen wrote:
> > > > On Sat, 12 Nov 2022, Gabriel Somlo wrote:
> > > > 
> > > > > Modify the TX path to operate in an IRQ-compatible way, while
> > > > > maintaining support for polling mode via the poll timer.
> > > > > 
> > > > > Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
> > > > > ---
> > > > >  drivers/tty/serial/liteuart.c | 67 ++++++++++++++++++++++++-----------
> > > > >  1 file changed, 47 insertions(+), 20 deletions(-)
> > > > > 
> > > > > diff --git a/drivers/tty/serial/liteuart.c b/drivers/tty/serial/liteuart.c
> > > > > index e30adb30277f..307c27398e30 100644
> > > > > --- a/drivers/tty/serial/liteuart.c
> > > > > +++ b/drivers/tty/serial/liteuart.c
> > 
> > > > > +	if (port->irq) {
> > > > > +		u8 irq_mask = litex_read8(port->membase + OFF_EV_ENABLE);
> > > > > +		litex_write8(port->membase + OFF_EV_ENABLE, irq_mask & ~EV_TX);
> > > > 
> > > > If you put irq_mask into liteuart_port you wouldn't need to read it 
> > > > back here?
> > > 
> > > So, instead of `bool poll_tx_started` I should just keep a copy of the
> > > irq_mask there, and take `&port->lock` whenever I need to *both* update
> > > the mask *and* write it out to the actual device register?
> > 
> > I was mostly thinking of storing EV_RX there but then it could be derived 
> > from port->irq that is checked by all paths already.
> 
> So, just to clear up the best course of action here (before I rebase
> everything on top of tty-next: How about the patch below (currently
> applied separately at the end of the entire series, but I can respin
> it as part of the rx path (12/14) and tx path (13/14) as appropriate.
> 
> Let me know what you think.

I'm fine with that I think. I'd change to bits parameter name to something 
more meaningful though, I guess the irq_mask could do ok there.