From: Duane Hill (no email)
Date: Tue Feb 27 2007 - 07:41:53 EST
On Tue, 27 Feb 2007, Leandro Santi wrote:
> On 2/26/07, Andrew McNamara <> wrote:
>> >> >>> Does accept() report errors when Postfix is built without MySQL?
>> >> >>
>> >> >> Same errors are showing up in the logs after a rebuild without MySQL.
>> I did some googling, and only found old hits for freebsd - things like the
>> following seem rather similar, although this is from 2005, and surely would
>> have been addressed by now (or there would be a lot more reports):
>> Are you using an SMP machine and kernel? Maybe try without SMP? Might also
>> be worth trying with a 32 bit kernel.
>> I have a vague memory of some old versions of Linux doing something like
>> this when the socket was closed by the remote end between userspace being
>> notified that listening socket was ready, and the accept being performed.
>> If you're feeling keen, you could create a stress test that rapidly opened
>> and closed sockets (or use a port scanning tool to do the same - nmap?)
> FWIW, I've seen Linux 2.2 accept() return EPIPE. This
> made smtpd die, because of the unexpected, and
> undocumented kernel behavior. Sadly, those machines
> do not exist anymore, and the problem was rare, so the
> issue never got enough priority for a full dissection. I recall
> it was a fairly recent kernel, 2.2.19 plus patches. Those
> machines were doing border-mx, with high connection
> rates and traffic levels.
The server that is getting the most of the errors is a Dell PowerEdge
6850. It has 4x3ghz Intel Dual Core Zeon processors and 12 gig ram. As
I've stated earlier in this thread, it is running FreeBSD 5.5-RELEASE-p9
amd64. The version is older because of a previous MTA that was running was
not built for the later versions. This server I can't really "play" with
testing as it is the main border MX server. It gets hit with an average
close to 4.5 million connections and/or messages per day. This server has
the number of inbound connections limited to 750. I see all 750
connections taken during peak times.
I do have another server that doesn't get hit as much and still gets a
handful of the errors. It is running FreeBSD 6.1-RELEASE-p10 i386. I'm not
sure on the brand of server. It is a dual Zeon 2.5ghz with four gig ram.
This server has the number of inbound connections limited to 300. There
are, on occasion, times when all connections are taken. This only happens
if connections to the other server fail twice because of the way I have
the MX records set up.