UNIX Socket Resource Temporarily Unavailable
============================================

Commonly visible as:

---%<-------------------------------------------------------------------------
imap-login: Error: net_connect_unix(imap) failed: Resource temporarily
unavailable
---%<-------------------------------------------------------------------------

This means that there are more imap-login processes trying to connect to the
"imap" UNIX socket than there are imap processes accepting the connections. The
kernel's connection listener queue got full and it started rejecting further
connections. So what can be done about it?

Wrong service settings
----------------------

This can happen if 'service imap { client_limit } ' is set to anything else
than 1. IMAP (and POP3 and other mail) processes do disk IO, lock waiting and
such, so if all the available imap processes are stuck waiting on something,
they can't accept new connections and they queue up in the kernel. For mail
processes only 'client_limit=1' is recommended, except maybe for very tiny
systems with a few users.

It can also happen if 'service imap { process_limit } ' is reached. Dovecot
logs a warning if process_limit or client_limit is reached.

Out of file descriptors
-----------------------

If the "ulimit -n" is too low, kernel stops notifying the process about
incoming connections. Make sure that the limit is at least as high as the
client_limit. Dovecot also internally checks this, and if it's too low it
writes a warning to stderr at startup (and to log in v2.2.33+).

Note that Dovecot is not using "dovecot" user's or PAM's limits in general.
Make sure the limits are correct with:'cat /proc/`pidof dovecot`/limits'

Master process busy
-------------------

Dovecot master process forks all of the new processes. If it's using 100% CPU,
it doesn't have time to fork enough new processes. Even if it's not constantly
using 100% CPU there may be fork bursts where it temporarily gets too busy. The
solution is to make it do less work by forking less processes:

 * Most importantly switch to <high-performance login process mode>
   [LoginProcess.txt]. This alone might be enough.
 * You can also switch (most of the) other commonly forked processes to be
   reused. For example 'service imap { service_count = 100 } ' reuses the imap
   process for 100 different IMAP connections before it dies. This is useful
   mainly for imap, pop3 and managesieve services. It's better to avoid using
   'service_count=0' (unlimited) in case there are memory leaks.
 * You can pre-fork some idling processes to handle bursts with 'service {
   process_min_avail }'.

See <Services.txt> before changing any service settings. Some services require
specific values to work correctly.

Listener queue size
-------------------

Dovecot uses the ' service { client_limit } '*' service { process_limit } ' as
the listener queue size. Dovecot v2.2.25 and older also had a hardcoded maximum
of 511, while later versions have no upper limit. Most OSes use an even lower
limit, typically 128. In Linux you can increase this from
'/proc/sys/net/core/somaxconn'.

Dovecot v2.2.[0-6] bug
----------------------

Dovecot v2.2.0 - v2.2.6 were attempting to optimize host.domain lookups by
doing them only once in the master process. Unfortunately they were actually
doing the lookup every time when creating a new process. In some configuration
this lookup could have done a somewhat slow DNS lookup, causing the process
creation to become very slow and triggering this message.

The fix is in v2.2.7 and you can also workaround this:

 * Add to dovecot.conf: 'import_environment = TZ DEBUG_OUTOFMEM
   DOVECOT_HOSTDOMAIN=mailserver.example.com' (of course changing the value)

(This file was created from the wiki on 2019-06-19 12:42)