On Tue, Oct 26, 2010 at 11:52 AM, Gabriel Petrovay
<[email protected]> wrote:
> Thanks for the answer!
>
> Or, is a solution like the one we already wrote good (see below or
> attached)? At least it does the job for us. Your solution (if
> implemented like you said) will wait always 4 times.
no, that was a description of what the MPMs use on Unix, and as stated
I don't think fcgid needs as many iterations or as long a sleep; some
key characteristics missing from that description:
* it processes the entire table between sleeps instead of focusing on
one before moving to the next
* it stops waiting when there is nothing more to wait for
> In the one below,
> it waits only as long as the process is still running (for maximum 30
> secs). And maybe you want to define this timeout in the configuration
> file.
Since the MPMs have gotten away with this being unconfigurable for so
many years (back to httpd 1.3 or earlier), we should be able to avoid
configuration for mod_fcgid as well.
>
> Does this make sense?
>
>
> Add instead of lines 379-380 the following lines:
>
> // original code
> //////////////////////////////////////////////////////////////
> // else
> // proc_kill_force(&proc_table[i], main_server);
> // new code
> ///////////////////////////////////////////////////////////////////
> else {
> // don't kill the process immediately, wait 30 sec
> // every 0.5s test if the process has finished
> int seconds = 30; // get from the config file or default
> int steps = 2 * seconds;
> int chunk = seconds * 1000000 / steps;
> while (steps > 0) {
> // if the process exited, break the loop
> if (apr_proc_wait(&(proc_table[i].proc_id), &exitcode,
> &exitwhy, APR_NOWAIT) != APR_CHILD_NOTDONE) {
> break;
> }
> // chunk must be in micro-seconds
> apr_sleep(chunk);
> steps--;
> }
> if (steps != 0) {
> proc_kill_force(&proc_table[i], main_server);
> }
> }
The biggest problem here is potentially waiting a long time for a
specific child process before moving on to the next. That's
potentially 30 seconds * huge-number. Also, I think it is worthwhile
setting up the code so that it is easy to tweak the sleeps and
actions, like the MPMs have.
Look for ap_reclaim_child_processes() at
http://svn.apache.org/viewvc/httpd/httpd/trunk/server/mpm_unix.c?view=markup
