Re: [PATCH] serial: mxs-auart: wait for DMA buffer to flush before shutdown

[Hector Palacios <hector.palacios@digi.com>]

Hello Uwe,

On 10/02/2013 09:44 AM, Uwe Kleine-König wrote:
> Hello,
>
> On Wed, Oct 02, 2013 at 09:31:16AM +0200, Hector Palacios wrote:
>> On 10/01/2013 09:48 PM, Uwe Kleine-König wrote:
>>> On Tue, Oct 01, 2013 at 06:18:35PM +0200, Hector Palacios wrote:
>>>> The shutdown function was not waiting for the DMA buffer to flush
>>>> before disabling the AUART. This lead to many bytes not being
>>>> transferred (specially at low baudrates), as they were still in the
>>>> DMA buffer when the AUART was shutdown.
>>>> This patch also adds the check for the BUSY flag.
>>>>
>>>> Signed-off-by: Hector Palacios <hector.palacios@digi.com>
>>>>
>>>> [...]
>>>> ---
>>>>   drivers/tty/serial/mxs-auart.c | 16 +++++++++++++++-
>>>>   1 file changed, 15 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/tty/serial/mxs-auart.c b/drivers/tty/serial/mxs-auart.c
>>>> index f85b8e6..0d8b2ca 100644
>>>> --- a/drivers/tty/serial/mxs-auart.c
>>>> +++ b/drivers/tty/serial/mxs-auart.c
>>>> @@ -750,12 +750,26 @@ static int mxs_auart_startup(struct uart_port *u)
>>>>   	return 0;
>>>>   }
>>>>
>>>> +static unsigned int mxs_auart_tx_empty(struct uart_port *u);
>>>> +
>>>>   static void mxs_auart_shutdown(struct uart_port *u)
>>>>   {
>>>>   	struct mxs_auart_port *s = to_auart_port(u);
>>>> +	unsigned int to;
>>>> +
>>>> +	if (auart_dma_enabled(s)) {
>>>> +		/* Wait enough time to flush DMA buffer completely */
>>>> +		to = u->timeout * UART_XMIT_SIZE / u->fifosize;
>>> u->timeout is the time needed to send one char, right? UART_XMIT_SIZE is
>>> the size of the circular buffer, fifosize is the size of the fifo. I
>>> don't get what you get by dividing by the fifosize. I would have
>>> expected something like:
>>>
>>> 	u->timeout * min(UART_XMIT_SIZE, u->fifosize)
>>>
>>> Can you explain, maybe in a comment along the code for the next person
>>> not understanding?
>>
>> u->timeout is *not* the time needed to send one char but the time
>> needed to flush the complete port fifo (see uart_update_timeout() in
>> serial_core.c where it is set).
> Ah, right. (BTW, uart_update_timeout should better round up instead of
> round down, i.e.
> -	port->timeout = (HZ * bits) / baud + HZ/50;
> +	port->timeout = DIV_ROUND_UP(HZ * bits, baud) + HZ/50;
> )
>
>> UART_XMIT_SIZE is the number of bytes in the DMA buffer, so I divide
>> the max number of bytes in the FIFO by the FIFO size and multiply by
>> the FIFO timeout.
> ditto here, better round up. Although I think using a completion could
> shorten the timeout considerably because there are often less than
> UART_XMIT_SIZE chars in the buffer?!

I'm not really waiting the full timeout. I'm checking for TX fifo empty every 1ms and 
for a maximum of 'to' ms.


>>> I wonder if the individual driver is the right place this must be fixed
>>> in or if there should be (or even exists) some help form the serial
>>> framework?!
>>
>> Good question, the issue should exit in other DMA drivers as well.
> Added Jiri and lakml to Cc:. Maybe someone can say something about that
> topic?
>
> Best regards
> Uwe
>


Best regards,
--
Hector Palacios
--
To unsubscribe from this list: send the line "unsubscribe linux-serial" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html