Re: [PATCH] gssd: Improve scalability by not waiting for child processes

[Steve Dickson <SteveD@redhat.com>]

On 09/25/2015 06:53 AM, Jeff Layton wrote:
> On Wed, 23 Sep 2015 17:20:50 -0400
> Steve Dickson <steved@redhat.com> wrote:
> 
>> Instead of waiting on every fork, which would
>> become a bottle neck during a mount storm, simply
>> set a SIGCHLD signal handler to do the wait on
>> the child process
>>
>> Signed-off-by: Steve Dickson <steved@redhat.com>
>> ---
>>  utils/gssd/gssd.c      | 18 ++++++++++++++++++
>>  utils/gssd/gssd_proc.c | 11 ++---------
>>  2 files changed, 20 insertions(+), 9 deletions(-)
>>
>> diff --git a/utils/gssd/gssd.c b/utils/gssd/gssd.c
>> index e480349..8b778cb 100644
>> --- a/utils/gssd/gssd.c
>> +++ b/utils/gssd/gssd.c
>> @@ -44,11 +44,13 @@
>>  #define _GNU_SOURCE
>>  #endif
>>  
>> +#include <sys/types.h>
>>  #include <sys/param.h>
>>  #include <sys/socket.h>
>>  #include <sys/time.h>
>>  #include <sys/resource.h>
>>  #include <sys/inotify.h>
>> +#include <sys/wait.h>
>>  #include <rpc/rpc.h>
>>  #include <netinet/in.h>
>>  #include <arpa/inet.h>
>> @@ -736,6 +738,21 @@ sig_die(int signal)
>>  	printerr(1, "exiting on signal %d\n", signal);
>>  	exit(0);
>>  }
>> +static void
>> +sig_child(int signal)
>> +{
>> +	int err;
>> +	pid_t pid;
>> +
>> +	/* Parent: just wait on child to exit and return */
>> +	do {
>> +		pid = wait(&err);
>> +	} while(pid == -1 && errno != -ECHILD);
>> +
>> +	if (WIFSIGNALED(err))
>> +		printerr(0, "WARNING: forked child was killed"
>> +			 "with signal %d\n", WTERMSIG(err));
>> +}
>>  
>>  static void
>>  usage(char *progname)
>> @@ -902,6 +919,7 @@ main(int argc, char *argv[])
>>  
>>  	signal(SIGINT, sig_die);
>>  	signal(SIGTERM, sig_die);
>> +	signal(SIGCHLD, sig_child);
>>  	signal_set(&sighup_ev, SIGHUP, gssd_scan_cb, NULL);
>>  	signal_add(&sighup_ev, NULL);
>>  	event_set(&inotify_ev, inotify_fd, EV_READ | EV_PERSIST, gssd_inotify_cb, NULL);
>> diff --git a/utils/gssd/gssd_proc.c b/utils/gssd/gssd_proc.c
>> index 11168b2..8f5ca03 100644
>> --- a/utils/gssd/gssd_proc.c
>> +++ b/utils/gssd/gssd_proc.c
>> @@ -656,16 +656,9 @@ process_krb5_upcall(struct clnt_info *clp, uid_t uid, int fd, char *tgtname,
>>  			/* fork() failed! */
>>  			printerr(0, "WARNING: unable to fork() to handle"
>>  				"upcall: %s\n", strerror(errno));
>> -			return;
>> +			/* FALLTHROUGH */
>>  		default:
>> -			/* Parent: just wait on child to exit and return */
>> -			do {
>> -				pid = wait(&err);
>> -			} while(pid == -1 && errno != -ECHILD);
>> -
>> -			if (WIFSIGNALED(err))
>> -				printerr(0, "WARNING: forked child was killed"
>> -					 "with signal %d\n", WTERMSIG(err));
>> +			/* Parent: Return and wait for the SIGCHLD */
>>  			return;
>>  		}
>>  no_fork:
> 
> I was thinking that there was some reason that we couldn't do this --
> that there were data structures that would get wiped if you got another
> upcall while the first was being processed. The forking should prevent
> that though, so I think this looks reasonable.
> 
> Acked-by: Jeff Layton <jlayton@poochiereds.net>
> 
Self Nak... During my testing there was a large number zombie rpc.gssd
process... I'm not sure why but they are there... 

steved.