Tracking down em problem

[Eric Anderson]

Sven Willenberger wrote:
> FreeBSD6.0-RC1 (Wed Oct 26 13:31:21 EDT 2005)
> 
> I seem to have an issue with losing connections to an em interface
> during process of heavy IO load. There are several variables here so I
> am hoping for some guidelines to help troubleshoot this.
> 
> I have a postgresql server (8.0.4) set up on an i386 system. The data
> directory is on its own partition (which is actually a gstripe/gmirror
> setup -- see the footnote after my problem description).
> 
> I have enabled a replication system from another server. When I started
> relication there was a large amount of data that had to be fed to this
> server via the em0 interface. During this process, while ssh'ed to the
> box, my connection would just hang for a few moments, then it would
> recover. However, if I cd to the data directory (stripe/mirror) and
> start ls -alrt several times, the connection actually gets broken; not
> only my ssh connection but the replication connection from the master
> server is broken.
> 
> I have tried to set debug.mpsafenet=0 in /boot/loader.conf to no avail
> -- the same issue happens. Preemption is enabled in the kernel, as is
> sched_4bsd. I don't really know how to proceed at this point to try and
> troubleshoot this issue: as it stands now, it is most definitely a show
> stopper for the purposes of this server.


I've seen something similar on recent 5.4-STABLE, also using emX 
devices.  I have 3 Dell 1850's showing the same exact issue, and a few 
1850's that are not.  The ones that are not, are 5.4-RELEASE, and the 
ones that do, are running 5.4-STABLE.  In dmesg, I see a warning like this:

Nov  1 19:56:06 hal kernel: em1: Link is up 1000 Mbps Full Duplex

I don't see a 'link is down', just 'Link is up'.  One machine I've seen 
this on repeatedly is from about August 16th.

I'm using SCHED_4BSD, SMP, and most of the other GENERIC settings.

If anyone wants more details, let me know.  I have a spare Dell 1850 I 
can play with.


Eric


-- 
------------------------------------------------------------------------
Eric Anderson        Sr. Systems Administrator        Centaur Technology
Anything that works is better than anything that doesn't.
------------------------------------------------------------------------