[Dovecot] pop3-login process leakage

[Clive Jones]

Hi,

I've recently deployed dovecot on our servers, to replace
courier-imapd. I'm delighted with its features and performance,
but there's a major problem - it's leaking pop3-login processes.

We have about 60 client machines, each collecting mail once every
two minutes. In this configuration, the number of pop3-login
processes increases by about ten an hour - apparently without
bound.

My workaround is to increase the open files rlimit, and the kernel
overall limit on open files. This postpones disaster for long
enough that killing and restarting dovecot nightly avoids
problems.

A possibly related problem is that about 1% of attempted POP3
logins fail. At the server, all I see is a syslog entry like this:
  Nov  3 22:17:21 greenwich pop3-login:
    Disconnected: Inactivity [10.76.30.246]

At the client, there is a long (about 60-second) timeout. Then
the client automatically retries, and generally succeeds.

There are *not* exactly as many such disconnects as there are
stray pop3-login processes - I counted roughly 230 inactivity
disconnects in the time it took for 190 stray processes to
accumulate.

We are running dovecot 0.99.10 on OpenBSD 3.2 . The configuration is
fairly normal, except that we are using a custom userdb and
passdb (compiled as shared objects):
  auth_userdb = passwd8
  auth_passdb = smb

passwd8 is just userdb-passwd.c with a trivial tweak to
truncate presented user names to eight characters.

smb is an shim that delegates authentication to smbclient, to
check passwords against NT domain controllers.

I've placed the source code for these two modules in:
  http://www.nsict.org/~clive/misc/dovecot-2003-11-04/
...along with my dovecot.conf, and a file that illustrates the
growth in the number of pop3-login processes over time.

So far as I'm aware, those modules are loaded by the dovecot-auth
process, not the pop3-login processes, and are therefore unlikely
to be the problem?

The client machines are LAN-connected, and running a mixture
of Outlook (2000/XP) and Outlook Express (5.5 and 6). There
are also a couple of people using Mozilla. Most are collecting
e-mail via POP3, a few via IMAP.

This issue didn't show up under beta-test loading before I
deployed dovecot. Since it seems only to be exhibited (or, at
least, noticeably exhibited) under live load, it's very hard for
me to diagnose things further without disrupting service.

I'd be very grateful if anyone could suggest what might be wrong.

Regards,

--Clive.