PROBLEM: serial port FIONREAD from realtime thread

Problem Conditions:

1) linux 2.6.11.7, but possibly other kernels too
1) realtime thread
2) serial port open()ed *either* with NON_BLOCKING or not
3) ioctl( FIONREAD ) always returns zero

int ncharin;
ioctl( fd, FIONREAD, &ncharin )

The above lines of code operate incorrectly from a real time thread against a normal PC serial port. The "ncharin" is always set to zero if called from a loop. The work around is to block the calling realtime thread with a sleep of some kind. That workaround is ugly. We need this to work from a realtime thread.

The program below illustrates the problem, but requires two computers. On one computer run minicom and connect an RS-232 cross over cable to the test box (at COM2, per the test source). Compile and run the program below on the test box. On the 2nd computer, say a non-test box, run minicom at 38400 kbaud to match the baudrate in the test box and program below. Run the test program as root so it can escalate the priority to realtime. Minicom should be used to send chars to the test program during the 20 seconds that the test program runs.

Test program will only run for 20 seconds. While it's running, you should send characters from minicom to the test box and you will see the buffered count of characters increment by 1. That is how it is supposed to work. Then comment out the usleep() call and recompile and do the same test again. Here you see the test program does not operate properly. The buffered character count remains at 0.

What causes this? What is the usleep() enabling?

-----------------------------------

#include <unistd.h>
#include <fcntl.h>
#include <termios.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/ioctl.h>
#include <linux/serial.h>
#include <sched.h>
#include <time.h>

int fd;

int openPort();
void setRealtimePriority();

int main( int argc, char** argv )
{

  int         ncharin;
  int         last;
  time_t      start;
    if( openPort() != 0 )
  {
      fprintf( stderr, "unable to open port\n" );
      exit(1);
  }
    setRealtimePriority();
    // poll the FIONREAD and print the number of chars in the recv buffers

start = time(NULL);
// run only for 20 seconds, since realtime priority steals cpu from
consoles.
while( time(NULL) - start < 20 )

  {
      if( -1 == ioctl( fd, FIONREAD, &ncharin ) )
      {
          printf( "error from ioctl\n" );
          ncharin = 0;
      }

      // if this line is commented out, then ioctl FIONREAD does not work.
      usleep(1);
        if( last != ncharin )
      {
          // this should show a monotonically increasing number of 
characters.
          printf(" %d", ncharin );
                fflush(stdout);
                    last = ncharin;
      }

}
}

int openPort()
{
struct termios t;

/* The O_NDELAY flag tells UNIX that this program doesn't care what
state the DCD signal line is in - whether the other end of the port is up and running. If you do not specify this flag, your process
will
be put to sleep until the DCD signal line is the space voltage.
*/
fd = open( "/dev/ttyS1", O_RDWR

                          | O_NDELAY
                          | O_NOCTTY );

if( fd == -1 )

      return -1;

  memset( &t, 0, sizeof(t) );

  cfsetispeed( &t, B38400 );
  cfsetospeed( &t, B38400 );

// Enable the receiver and set local mode...

  //  CLOCAL  : local connection, no modem contol
  //  CREAD   : enable receiving characters

t.c_cflag |= (CLOCAL | CREAD);

  t.c_cflag &= ~CSTOPB;

t.c_cflag &= ~CSIZE; // Mask the character size bits
t.c_cflag |= CS8; // Select 8 data bits

  t.c_cflag &= ~PARENB;
  t.c_cflag &= ~CSTOPB;

// Raw input is unprocessed.
// Input characters are passed through exactly as they are received, // when they are received. Generally you'll deselect the ICANON, ECHO, // ECHOE, and ISIG options when using raw input:
t.c_lflag &= ~(ICANON | ECHO | ECHOE | ISIG);

// Raw output is selected by resetting the OPOST option in the
c_oflag member:
t.c_oflag &= ~OPOST;

  t.c_cc[VMIN]  = 1;
  t.c_cc[VTIME] = 0;

// Set the new options for the port...
tcsetattr( fd, TCSANOW, &t );

return 0;
}

void setRealtimePriority()
{
int ec;

  printf("setRealtimePriority()\n");
    struct sched_param     p;

  memset( &p, 0, sizeof(p) );

  p.sched_priority = 2;

ec = sched_setscheduler( 0, SCHED_RR, &p );
if( ec != 0 )
{
fprintf(stderr, "pthread_setschedparam ec=%d, %s\n", ec,

strerror(ec) );
      fflush(stderr);
      exit(2);

}
}

The tty flip buffering uses a workqueue to push received data to the line discipline.

Your polling loop at raised priority is probably preventing the workqueue from being processed by the events kernel thread.

When you call usleep() your process yields so receive data can be processed.

Have you tried (just for testing) lowering your process priority below that of the events kernel thread?

Paul Fulghum
Microgate Systems, Ltd.

Thanks Paul. This sounds right. No, I have not dialed in at such a fine granularity. Obviously at a non realtime priority things work OK, and obviously I need to run at a realtime priority. There may be some room to fit in between.

1) Can you point me at the source file(s) where I can explore this further? (source file the workqueue written to).

2) How do I determine the priority of the events kernel thread (where is it created, and/or priority set).

Thanks much.

drivers/char/tty_io.c has the flip buffer code

The function tty_flip_buffer_push() is called by the hardware driver when data has been added to the flip buffer. This function calls flush_to_ldisc() either directly (if low_latency is set) or indirectly through the workqueue.

There is a problem in setting low_latency if tty_flip_buffer_push() is called in interrupt context, which many drivers do. You can try setting this and seeing if your machine crashes and burns.

This code is a moving target at this time. Alan Cox is working on bringing
a small measure of sanity to the code.

You should be able to look at the output of 'top' or 'ps' for the 'events/0' entry.

My machine shows a priority of 10.

Paul Fulghum
Microgate Systems, Ltd.