Subj : pthread_cond_timedwait problems
To   : comp.programming.threads
From : markh@compro.net
Date : Sat Aug 13 2005 01:27 pm

I sometimes use pthread_cond_timedwait in an application for the sole
purpose of
delay. It does not delay (ETIMEDOUT) correctly. The sample program
below shows my
problem with this call. I have a hardware timer from a special pci card
mapped into
my task space that has microsecond resolution. This timer verifies the
problem with
this funtion. I have however included in the sample program below an
alternative
method of timing the (ETIMEDOUT) of this call using the gettimeofday
call. It is
not as accurate as the hardware timer but show the problem non the
less.


#include <stdlib.h>
#include <errno.h>
#include <sys/mman.h>
#include <string.h>
#include <signal.h>
#include <pthread.h>
#include <unistd.h>
#include <stdio.h>
#include <sys/resource.h>
#include <linux/threads.h>
#include <linux/unistd.h>

#include <sys/types.h>
#include <sys/wait.h>
#include <sys/time.h>
#include <rtom.h>

int main()
{
    volatile unsigned long *rtom_microseconds;

    unsigned int usecs;
    unsigned int rtom_han;
    unsigned int elapsed_usecs;

    struct timeval  tod_end_time;
    struct timeval  tod_start_time;
    struct timezone tod_zone;

    struct timespec delay = { 0, 10000 }; // 10 micro second delay
    struct timespec delay_tim;

    pthread_mutex_t lck = PTHREAD_MUTEX_INITIALIZER;
    pthread_cond_t  cv  = PTHREAD_COND_INITIALIZER;

    rtom_han = rtom_usec_map(&rtom_microseconds, 0); // mmap usec timer

    gettimeofday(&tod_start_time,&tod_zone); // get start time

    delay_tim.tv_sec   = tod_start_time.tv_sec  ;
    delay_tim.tv_nsec  = (tod_start_time.tv_usec * 1000);
    delay_tim.tv_sec  += delay.tv_sec;
    delay_tim.tv_nsec += delay.tv_nsec;

    printf("Attempting  delay of %d nsec (%d usecs)\n",
            (uint)delay.tv_nsec,
            ((uint)delay.tv_nsec / 1000));

    usecs = *rtom_microseconds; // get real start time
    while (pthread_cond_timedwait(&cv, &lck, &delay_tim) != ETIMEDOUT);
    elapsed_usecs = (*rtom_microseconds - usecs);

    gettimeofday(&tod_end_time,&tod_zone);

    printf("gettimeofday reported delay  =  %d usecs\n",
            (int)(tod_end_time.tv_usec - tod_start_time.tv_usec));

    printf("Actual usec delay from timer =  %d usecs\n",
            elapsed_usecs);

    exit(0);

    rtom_usec_unmap(rtom_han);

}


Here is the output of this program as is for a delay of 10000ns (10us).

Attempting  delay of 10000 nsec (10 usecs)
gettimeofday reported delay  =  73 usecs
Actual usec delay from timer =  12 usecs

That looks ok. Now if I try 100000ns (100us) by changing

struct timespec delay = { 0, 100000 }; // 100 micro second delay

Attempting  delay of 100000 nsec (100 usecs)
gettimeofday reported delay  =  2120 usecs
Actual usec delay from timer =  2059 usecs

Its obvious it takes gettimeofday around 70 usecs to complete but
both the timings are bogus. The rtom should show 100 and the
gettimeofday
should be around 170.

Ok lets try 1000000ns (1000us)

Attempting  delay of 1000000 nsec (1000 usecs)
gettimeofday reported delay  =  2225 usecs
Actual usec delay from timer =  2164 usecs


Whats going on here???? I don't see anything wrong with the program.
Is there something wrong here with this code or is a BUG??

Thanks
Mark

.