[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-kernel
Subject:    Re: [PATCH] 2.3.18: siginfo data available for all signals
From:       Robert de Vries <rhdv () rhdv ! cistron ! nl>
Date:       1999-09-15 23:43:00
[Download RAW message or body]

On Wed, 15 Sep 1999, Robert de Vries wrote:

> > In principle I approve, as long as it does not slow signal delivery.  It
> > was obviously intended to work this way all along.  Sorry I haven't time
> > to examine the code.
> 
> This all depends on the speed of the slab allocator for the siginfo
> structure and the initialisation of the structure. I guess that it's fast
> because that was the reason for putting in the allocator in the first
> place and making a special slab for the siginfo structures.

Now doing a small benchmark reveals that RT signals take a little bit
longer to arrive than normal signals. I have written a small program to
measure this in a 2.2.5-22 redhat std. kernel.

This gives me for 1000 signals the following numbers on a 166 MHz Pentium:

normal signal takes on average 7.9 microsecs
RT signal takes on average 10.9 microsecs

The difference must be due to the queueing of the signal.

What does the community think?

Is the patch ok as is and is the 3 extra microsec a problem? (please run
the test prog on your platform to give me an idea what kind of latencies
you have on other archs/faster CPUs)

Or should I be more conservative in the patch and only put siginfo's in
the queue when they are either RT or if they pass siginfo in the first
place (all signals)?

	Robert

-- 
Robert de Vries
rhdv@rhdv.cistron.nl

["sigperf.c" (TEXT/PLAIN)]

#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#include <unistd.h>
#include <signal.h>


#if 1
#define SIGNO SIGUSR1		/* normal signal */
#else
#define SIGNO SIGRTMIN		/* RT signal */
#endif

#define NOF_DELAY 1000

int delay[1000];
int idel;
struct timeval start;

void handle_signal(int signal)
{
  struct timeval end;
  int delta;

  gettimeofday(&end, NULL);
  delta = (end.tv_sec - start.tv_sec) * 1000000 +
    (end.tv_usec - start.tv_usec);

  delay[idel++] = delta;
}

void print_hist(void)
{
#define HISTSIZE 1000

  int hist[1000 + 1];
  int i;

  memset(hist, 0, sizeof(hist));

  for (i = 0; i < NOF_DELAY; i++) {
    if (delay[i] >= HISTSIZE) {
      printf("sample %d over max hist: %d\n", i, delay[i]);
      hist[HISTSIZE]++;
    }
    else {
      hist[delay[i]]++;
    }
  }
  for (i = 0; i < HISTSIZE; i++) {
    if (hist[i]) {
      printf("%d: %d\n", i, hist[i]);
    }
  }
  printf("HC: %d\n", hist[HISTSIZE]);
}

void print_avg(void)
{
  double sum;
  int i;

  sum = 0;
  for (i = 0; i < NOF_DELAY; i++) {
    sum += delay[i];
  }

  printf("avg. delay: %f\n", sum/NOF_DELAY);
}

int main(void)
{
  struct sigaction act;
  int i;
  pid_t p;

  p = getpid();

  act.sa_flags = 0;
  act.sa_handler = handle_signal;
  sigemptyset(&act.sa_mask);

  if (sigaction(SIGNO, &act, NULL)) {
    perror("sigaction failed");
    exit(1);
  }

  for (i = 0; i < NOF_DELAY; i++) {
    gettimeofday(&start, NULL);
    kill(p, SIGNO);
  }

  print_hist();

  print_avg();

  return 0;
}

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/


[prev in list] [next in list] [prev in thread] [next in thread]