About session disconnect and an slow response some time later

["egoitz via Info" <info@xxxxxxxxxxxxxxxxxx>]

Good morning,

I wrote to the list some time ago due to this topic. I have noted that when a user asks to disconnect his/her sessions (imap sessions normally the disconnected ones) after some time like an hour or a half an hour passes I start having some slow responses to new connections. I have monitored them and it's slightly intermittent, KO and OK of server responding alert in less than 3 seconds. This intermittency lasts like 10 minutes and later the response delays increase. Obviously for avoid bigger service outages I stop and start Cyrus and all works fine later.

I have been trying to figure why it could be happening because I have seen it only happens when I launch TERM for a disconnection to several imap proccesses (the user that requested proccesses). It doesn't happen at the moment. As said for instance yesterday happened when an hour passed of the disconnection.

I don't have prefork param set in cyrus.conf, so... Cyrus spawns on demand. After doing some examinations my theory is that perhaps Cyrus is not calling spawn_service() every time gets needed. I would say it should happen in master.c :

2610               if (!in_shutdown && Services[i].exec &&
2611                    Services[i].nactive < Services[i].max_workers &&
2612                    Services[i].ready_workers == 0 &&
2613                    y >= 0 && FD_ISSET(y, &rfds))
2614                {
2615                    /* huh, someone wants to talk to us */
2616                    spawn_service(i);
2617                }

I think that perhaps ready_workers is not properly decremented when I launch this TERM and perhaps this causes Cyrus not to spawn any new more services of the kind of the terminated service. As yesterday happened in a more or less peak accesses hour it needed to have more services spawned but... as it seen ready_workers not to be 0 for that service... it didn't spawn new more services and as consequence, connections started becoming queued in being accepeted awaiting that happened when a proccess gets idle because it's client has disconnected or idled timeout or imap timeout happened....

So, I was wondering if in reap_child() for the states SERVICE_STATE_UNKNOWN and SERVICE_STATE_BUSY shouldn't be decremented the ready_workers in the service struct... Concretely in master.c in :

1121                case SERVICE_STATE_BUSY:
1122                    s->nactive--;
1123                   if (!in_shutdown && failed) {
1124                       syslog(LOG_DEBUG,
1125                               "service %s/%s pid %d in BUSY state: "
1126                               "terminated abnormally",
1127                               SERVICEPARAM(s->name),
1128                               SERVICEPARAM(s->familyname), pid);
1129                    }
1130                    break;
1131
1132                case SERVICE_STATE_UNKNOWN:
1133                    s->nactive--;
1134                   syslog(LOG_WARNING,
1135                           "service %s/%s pid %d in UNKNOWN state: exited",
1136                           SERVICEPARAM(s->name),
1137                           SERVICEPARAM(s->familyname), pid);

Obviously, another possible solution is to specify prefork in cyrus.conf but I'd rather avoid wasting memory when it's not needed...

What's your opinion mates?. Or what do you think Ellie, Bron :) . Have you ever seen something like it?.

Cheers!

Cyrus / Info / see discussions + participants + delivery options Permalink