unix/c gurus - Starting a process from gaya without it hanging

[lordy]

Hi,

I have a few cgi applications where I want to start a long running process.
However on nmt100 this causes Gaya to hang.

So I thought I'd write a daemon launcher, but still gaya waits for it to finish.
Can anyone see what the problem is...
(note that gaya is invoking the daemon process directly and not going via wget/sybhttpd)

References : here and here

The following program should allow gaya to launch any command as

Code:

daemon command args

but gaya still waits for termination. Maybe it is a signalling issue ?

(ignore the php tag )

PHP Code:

#include <sys/types.h>
#include <sys/wait.h>
#include <sys/stat.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <string.h>

/*
 * from http://www.steve.org.uk/Reference/Unix/faq_2.html#SEC16
 * Should allow safe forking of processes from gaya.
 * The directory change step has been omitted.
 */

#define SLEEP_USECS 1000000

static void set_logs();
static void close_all();

int main(int argc, char **argv) {

    //int status;
    int pid = fork();

    switch(pid) {

        case -1:
            fprintf(stderr,"Unable to fork - error %d",errno);
            exit(-1);
            break;
        case 0:
            // fall through
            break;
        default:
            _exit(0);
    }

    // child 1 ---------------------------------------------
    
    // Give parent time to die. - grasping at straws

    usleep(SLEEP_USECS);

    // NMT gaya still detects the child process. - trying both calls
    setpgrp();
    setsid();

    pid = fork();
    switch(pid) {

        case -1:
            fprintf(stderr,"Unable to fork child - error %d",errno);
            exit(-1);
            break;
        case 0:
            // fall through
            break;
        default:
            _exit(0);
    }

    // child 2 ---------------------------------------------

    // Give parent time to die. - grasping at straws
    usleep(SLEEP_USECS);

    setpgrp();
    setsid();

    // chdir("/");
    umask(0);

    set_logs();

    close_all();

    char *env;

    if ((env = getenv("DAEMON_DIR")) != NULL) {
        chdir(env);
    }

    char *d =  get_current_dir_name();


    int i;
    for(i = 0 ; i < argc ; i++ ) {
        printf("'%s' ",argv[i]);
    }
    printf(" [%s] \n",d);
    fflush(stdout);
    free(d);

    if (execvp(argv[1],argv+1) == -1) {
        fprintf(stderr,"Unable to exec [%s] - error %d",argv[1],errno);
    }
    _exit(-1);
}

static void close_all()
{
    long i;
    long fc_max = sysconf(_SC_OPEN_MAX);
    for( i = 3 ; i < fc_max ; i++ ) {
        close(i);
    }
}

static void set_logs()
{
    char *log_dir = "/share/tmp";
    char *out_log_name = "daemon.out";
    char *err_log_name = "daemon.out";

    char *env;

    if ((env = getenv("DAEMON_LOG_DIR")) != NULL) {
        log_dir = env;
    }
    char *out_path = malloc(strlen(log_dir) + strlen(out_log_name) + 2);
    char *err_path = malloc(strlen(log_dir) + strlen(err_log_name) + 2);

    mkdir(log_dir,0777); // ignore error for now

    if (out_path != NULL ) {
        sprintf(out_path,"%s/%s",log_dir,out_log_name);
        freopen(out_path,"w",stdout);
    }

    if (err_path != NULL ) {
        sprintf(err_path,"%s/%s",log_dir,err_log_name);
        freopen(err_path,"w",stderr);
    }
    freopen("/dev/null","r",stdin);
} 


[lordy]

Hi,

I sussed it. The first of the three processes should call exit() to signal back to gaya to move on The others should use _exit().

If anyone is interested the daemon launcher will be a small executable bundled with the next version of oversight.
They are welcome to use it, to stop gaya waiting for long running processes.

It also means I can remove the cron job hack.

[accident]

It might be better for your application to just run in the background on startup and then signal it to do something instead of launch it to run.

[lordy]

Hi, accident, thanks for the suggestion, I try to minimise the number of long running processes on the NMT.

I do have regular scans that are started by a cron job. ,
but I just need this for when the user initiates a manual scan. (which hopefully is not often).

Now I have this working, it means oversight does not add any overhead to NMT resources. Overall its easier now to just say..

Code:

daemon command args

rather than develop an additional server, and manage its installation/startup/shutdown to avoid this issue.

I did have a horrible hack in place before this to launch user scans, back when Oversight was just bash/awk scripts. I had a cron job that check for new commands every minute. I didn't have the toolchain in place at the time to write proper code! It actually worked better than expected but now I can get rid of it!

However, somewhere down the line, for top performance , I may need to introduce a proper server process for performance reasons.
At the moment oversight scans the entire video inventory during every page load. (400 items in about 10ms on c200. 250 items in 30ms on hdx1000 ) but performance will degrade with bigger collections, so a full client/server architecture *may* help here.

[lundman]

Look about right, you fork(), child->setsid() close(0,1,2) and execve a new process.

Your double fork should not be required (any more, legacy)

[lordy]

Tx, gaya was waiting for one of the sub-processes to call exit() ( rather than _exit() ).
Interesting about the double fork - I'll go do some reading...