[Dovecot] Dovecot 2 and TCP-Keepalive

[Sven Hartge]

Timo Sirainen <tss at iki.fi> wrote:
> On 20.10.2012, at 19.39, Sven Hartge wrote:

>> My question is: does Dovecot2 use TCP-Keepalive on its sockets per
>> default or do I need to enable it some way I have not yet discovered?

> It's the default yes. Of course Linux's default keepalive interval is
> something like 90 minutes, so have you changed that already?..

Yes, I did. For those systems it is set to 15 minutes right now.

>> The manual and wiki only talk about "keepalive" in connection with
>> the IMAP protocol and IDLE and my C-fu is too weak to understand the
>> source code.

> imap_idle_notify_interval (default 2 min) causes Dovecot to send data
> to IDLEing connections, which pretty much makes the TCP keepalive
> irrelevant. For non-IDLE connections Dovecot has a disconnect timeout
> of 30 minutes.

This is fine. As long as the client notices the termination of the
connection, everything should be OK.

Before I switched keepalive on for Perdition, the firewall/NAT would
internally throw away a connection but neither the client or the server
would notice this.

Then if the client tried to do something with this connection, like
select or save a message, the firewall/NAT would send a RST and the
client would then bug the user with a meaningless message like "folder
does not exist" which caused a lot of confusion for the end-user and
created quite the bit of trouble tickets.

This problem mostly happend with an IMAP connection to the "Sent
Messages" folder which normally does not see much changes until the
users writes and sends a mail. Then after the mail was sent via SMTP the
client tries to save the message, gets sent an RST from the firewall/NAT
and presents the user with a wrong and confusing error message. The user
then thinks his mail was not sent and sends it again. This time the
client opens a new connection to select the "Sent Messages" folder and
everything works. But the recipient gets the mail twice.

Again resulting in confusion and trouble tickets to be dealt with.

By switching to TCP keepalive (and reducing the keepalive time to 15
minutes) all those problems were solved and my users (and support staff)
were happy again ;)

Grüße,
Sven.

-- 
Sigmentation fault. Core dumped.