[RFC PATCH] tty: serial: core: Only invoke ->start_tx() if there is data to send

[RFC PATCH] tty: serial: core: Only invoke ->start_tx() if there is data to send

[Sebastian Andrzej Siewior]

I noticed that the serial8250_tx_dma() is invoked sometimes while
uart_circ_empty() says that the buffer is empty.

I tracked one occuring down to:

n_tty_write()
	 => O_OPOST(tty))
	   => the while loop did something but neither tty's ->write()
	      nor its ->uart_put_char() callback was invoked().
	   => tty->ops->flush_chars() is invoked with an empty buffer.

For the 8250 uart driver we end up with:
- DMA enabled
  nothing, just return (except there is DMA_TX bug or runtime-PM then we
  behave like in the no DMA case)

- no DMA
  enable THRI interrupt, wait for it, disable THRI interrupt again
  because there is nothing to be done.

While I don't know if it safe to drop that flush in n_tty if the buffer
is empty, it should not do any harm in serial's core part to not invoke
->start_tx() if the buffer is empty.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 drivers/tty/serial/serial_core.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/tty/serial/serial_core.c b/drivers/tty/serial/serial_core.c
index 5a78f6940760..e55724a911d5 100644
--- a/drivers/tty/serial/serial_core.c
+++ b/drivers/tty/serial/serial_core.c
@@ -90,7 +90,8 @@ static void __uart_start(struct tty_struct *tty)
 	struct uart_state *state = tty->driver_data;
 	struct uart_port *port = state->uart_port;
 
-	if (!tty->stopped && !tty->hw_stopped)
+	if (!tty->stopped && !tty->hw_stopped &&
+			!uart_circ_empty(&port->state->xmit))
 		port->ops->start_tx(port);
 }
 
-- 
2.1.0

Re: [RFC PATCH] tty: serial: core: Only invoke ->start_tx() if there is data to send

[Peter Hurley]

On 09/10/2014 03:33 PM, Sebastian Andrzej Siewior wrote:
> I noticed that the serial8250_tx_dma() is invoked sometimes while
> uart_circ_empty() says that the buffer is empty.
> 
> I tracked one occuring down to:
> 
> n_tty_write()
> 	 => O_OPOST(tty))
> 	   => the while loop did something but neither tty's ->write()
> 	      nor its ->uart_put_char() callback was invoked().
> 	   => tty->ops->flush_chars() is invoked with an empty buffer.
> 
> For the 8250 uart driver we end up with:
> - DMA enabled
>   nothing, just return (except there is DMA_TX bug or runtime-PM then we
>   behave like in the no DMA case)
> 
> - no DMA
>   enable THRI interrupt, wait for it, disable THRI interrupt again
>   because there is nothing to be done.
> 
> While I don't know if it safe to drop that flush in n_tty if the buffer
> is empty, it should not do any harm in serial's core part to not invoke
> ->start_tx() if the buffer is empty.

The serial core can't assume that start_tx() does not need invoking
because hardware that can stop_tx() with data in the transmitter
won't restart if the ring buffer is empty but data is still in the
transmitter. [Note that the 16C950 port type does this in the 8250 driver.]

So this has to be handled in the 8250 driver.

What is the actual issue? Are you trying not to unnecessarily wake
the omap hardware if runtime-PM is on?

Regards,
Peter Hurley

> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> ---
>  drivers/tty/serial/serial_core.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/tty/serial/serial_core.c b/drivers/tty/serial/serial_core.c
> index 5a78f6940760..e55724a911d5 100644
> --- a/drivers/tty/serial/serial_core.c
> +++ b/drivers/tty/serial/serial_core.c
> @@ -90,7 +90,8 @@ static void __uart_start(struct tty_struct *tty)
>  	struct uart_state *state = tty->driver_data;
>  	struct uart_port *port = state->uart_port;
>  
> -	if (!tty->stopped && !tty->hw_stopped)
> +	if (!tty->stopped && !tty->hw_stopped &&
> +			!uart_circ_empty(&port->state->xmit))
>  		port->ops->start_tx(port);
>  }
>  
> 

Re: [RFC PATCH] tty: serial: core: Only invoke ->start_tx() if there is data to send

[Sebastian Andrzej Siewior]

On 09/10/2014 09:52 PM, Peter Hurley wrote:
> The serial core can't assume that start_tx() does not need invoking
> because hardware that can stop_tx() with data in the transmitter
> won't restart if the ring buffer is empty but data is still in the
> transmitter. [Note that the 16C950 port type does this in the 8250 driver.]

oh, not sure how I missed this… But now that I look at this, it is also
that ->x_char that could be use for flow control which would have to be
sent even with an empty xmit buffer.
And 8250 in DMA mode does not look at x_char at all. But it would be
better if it would, right? However if the TX side does a 2 KiB transfer
not sure what should be done…

> So this has to be handled in the 8250 driver.
> 
> What is the actual issue? Are you trying not to unnecessarily wake
> the omap hardware if runtime-PM is on?

no, it is actually the extra interrupt for no reason that looked like
not needed at all. But I guess it is only during "startup" while the
init-script do fancy things and not during "normal" operations.

> Regards,
> Peter Hurley

Sebastian

Re: [RFC PATCH] tty: serial: core: Only invoke ->start_tx() if there is data to send

[Peter Hurley]

On 09/11/2014 04:03 AM, Sebastian Andrzej Siewior wrote:
> On 09/10/2014 09:52 PM, Peter Hurley wrote:
>> The serial core can't assume that start_tx() does not need invoking
>> because hardware that can stop_tx() with data in the transmitter
>> won't restart if the ring buffer is empty but data is still in the
>> transmitter. [Note that the 16C950 port type does this in the 8250 driver.]
> 
> oh, not sure how I missed this… But now that I look at this, it is also
> that ->x_char that could be use for flow control which would have to be
> sent even with an empty xmit buffer.
> And 8250 in DMA mode does not look at x_char at all. But it would be
> better if it would, right? However if the TX side does a 2 KiB transfer
> not sure what should be done…

Yeah, 8250 dma doesn't send x_char at all; at a minimum 8250 dma should
at least send the x_char.

The preferred solution is to:
1. Stop DMA
2. Enable interrupt mode, which will automatically send the x_char
3. Restart DMA (presumably from the interrupt handler if UART_LSR_THRE)

The issue is whether 8250 dma hardware in general can stop and restart
dma without losing where the dma transfer was.

Regards,
Peter Hurley