Instant crash with ZFS + iozone?

[Lassi Tuura]

Hi,

On May 4, 2010, at 02:49 , Jeremy Chadwick wrote:
On Tue, May 04, 2010 at 01:22:35AM +0200, Lassi Tuura wrote:
>> I installed 8.0-RELEASE, then 8.0-STABLE 201004 (amd64) on a system
>> with 4 * 1TB hard drives for ZFS (2 * Hitachi Deskstar 7K2000, 2 *
>> Samsung SpinPoint F3), plus 2GB IDE flash for OS itself (Transcend),
>> 4GB ECC RAM, AMD Athlon II X2 235e CPU, Asus M4A78L-M LE motherboard.
>> 
>> I can use the basic system fine. However when I create a ZFS volume
>> out of the 4 disks and run iozone on it, the system will reliably die
>> within 5 seconds or so, sometimes it takes a little longer, up to a
>> minute or so.
>> 
>> By "die" I mean the screen goes completely blank, and it will no
>> longer respond to anything - no network, not even ping and any
>> existing network connections will die, no keyboard, screen totally
>> black without as much as a cursor... nothing. The soft power button
>> won't work either. AFAIK the only thing that works is the reset
>> button.
> 
> This sounds like a hardware problem to me -- particularly excessive draw
> on the PSU, voltage issues, or other whatnots.  Especially if the screen
> goes black (no cursor, monitor not in power save mode) and requires a
> hard reset.  Possibly the problem is only witnessed under extreme disk
> I/O combined with high CPU usage + memory I/O.

It looks like the freezes are caused by PCIe wifi card, TP-LINK TL-WN350GD
which is identified as "ath0: <Atheros 2417> mem 0xfebf0000-0xfebfffff irq
21 at device 7.0 on pci3". If I remove the card the system no longer hangs.
Swapping PSU, memory or disks does not appear to change this, at least as
far as reusing bits from another system did not add or remove freezes.

I thought these Atheros chips are quite well supported in FreeBSD, but I
am quite new to FreeBSD. Which forum that would be able to help me debug
this, why for example it flakes under disk load? The card was by no means
expensive, but I'd like to return it while I still can if it's just bad;
and obviously, get it to work if this is just some silly conflict issue.

Regards,
Lassi