Hello,

after receiving a special character on a UART, I would like to send out
another character in less than 2ms. I can have good latencies of about some
50µs, but often I get 3ms, which is too long. I think, this is because
of the fifo. How could I configure the fifo, to get triggered at every
character?
Here is my setup:

struct termios options;
struct serial_struct serial;
int fd = open("/dev/ttyS0", O_RDWR | O_NOCTTY);

ioctl(fd, TIOCGSERIAL, &serial);
serial.flags |= ASYNC_LOW_LATENCY;
serial.xmit_fifo_size = ???; // a value of 1 here does not change a lot
ioctl(fd, TIOCSSERIAL, &serial);

tcgetattr(fd, &options);
cfsetispeed(&options, B19200);
cfsetospeed(&options, B19200);
options.c_cflag &= ~(CSIZE | PARODD | CSTOPB | CRTSCTS);
options.c_cflag |= CLOCAL | CREAD | CS8 | PARENB;
options.c_lflag &= ~(ICANON | ECHO | ECHOE | ISIG);
options.c_iflag |= INPCK | ISTRIP;
options.c_oflag &= ~OPOST;
options.c_cc[VTIME] = 0;
options.c_cc[VMIN] = 1;
tcsetattr(fd, TCSANOW, &options);

I've discovered the file linux/serial_reg.h with a lot of definitions,
for example:
#define UART_FCR_TRIGGER_1 0x00 /* Mask for trigger set at 1 */

But I don't know where I could find some documentation how to use these
defines...

I only know the "Serial Programming Guide for POSIX" but perhaps I need
something more specific to Linux...

Thanks in advance for any help!
Cheers, Peter

--
email: pmrb at free.fr
http://pmrb.free.fr/contact/

On Thu, 3 Nov 2005, Alain Mosnier wrote:

> Can Linux guarantee a maximum latency for a task in the first place?

No, but I don't need guarantee, I only need good latency for about 99,9%.

> Do you have other processes running? Have you played with the NICE
> values?

I use sched_setscheduler(0, SCHED_RR, ...), that should be enough for some
soft real-time.

When I get 3ms latency, there is no load on the system, so I think it's
just because not every character triggers an interrupt. Of course, I could
modify the driver for the UART (for example drivers/serial/8250.c), to
disable the fifo, but I prefer a solution without modification of the
kernel.

Greetings, Peter

--
email: pmrb at free.fr
http://pmrb.free.fr/contact/