[CentOS-virt] NIC Stability Problems Under Xen 4.4 / CentOS 6 / Linux 3.18

Tue Jan 31 00:41:53 UTC 2017
Kevin Stange <kevin at steadfast.net>

On 01/30/2017 06:12 PM, Adi Pircalabu wrote:
> On 31/01/17 10:49, Kevin Stange wrote:
>> You said 3.x kernels specifically. The kernel on Xen Made Easy now is a
>> 4.4 kernel.  Any chance you have tested with that one?
> Not yet, however the future Xen nodes we'll deploy will run CentOS 7 and
> Xen with kernel 4.4.

I'll keep you (and others here) posted on my own experiences with that
4.4 build over the next few weeks to report on any issues.  I'm hoping
something happened between 3.18 and 4.4 that fixed underlying problems.

>> Did you ever try without MTU=9000 (default 1500 instead)?
> Yes, also with all sorts of configuration combinations like LACP rate
> slow/fast, "options ixgbe LRO=0,0" and so on. No improvement.

Alright, I'll assume that probably won't help then.  I tried it on one
box which hasn't had the issue again yet, but that doesn't guarantee

>> I am having certain issues on certain hardware where there's no shutting
>> down the affected NICs.  Trying to do so or unload the igb module hangs
>> the entire box.  But in that case they're throwing AER errors instead of
>> just unit hangs:
>> pcieport 0000:00:03.0: AER: Uncorrected (Non-Fatal) error received:
>> id=0000
>> igb 0000:04:00.1: PCIe Bus Error: severity=Uncorrected (Non-Fatal),
>> type=Transaction Layer, id=0401(Requester ID)
>> igb 0000:04:00.1:   device [8086:10a7] error
>> status/mask=00004000/00000000
>> igb 0000:04:00.1:    [14] Completion Timeout     (First)
>> igb 0000:04:00.1: broadcast error_detected message
>> igb 0000:04:00.1: broadcast slot_reset message
>> igb 0000:04:00.1: broadcast resume message
>> igb 0000:04:00.1: AER: Device recovery successful
> This is interesting. We've never had any problems with the 1Gb NICs, but
> we're only using 10Gb for the storage network. Could it be a common
> problem with either the adapters, or the drivers which only replicate
> running the Xen enabled kernel?

Since I've never run the 3.18 kernel on a box of this type without
running in a dom0 and since I can't reproduce this kind of issue without
a fair amount of NIC load over a tremendous period of time, it's
impossible to test if it's tied to Xen.

However, I know this hardware works well under 2.6.32-*.el6 and
3.10.0-*.el7 kernels without stability problems, as it did with
2.6.18-*.el5xen (Xen 3.4.4).

I suspect the above errors are actually due to something PCIe related,
and I have a subset of boxes which are actually being impacted by two
distinct problems with equivalent impact, which increases the likelihood
that the boxes will die.  Another set of boxes only ever sees the unit
hangs which seem unrecoverable even unloading/reloading the driver.  A
third set has random recoverable unit hangs only.  With so much
diversity, it's even harder to pin any specific causes to the problems.

The fact we're both pushing NFS and iSCSI traffic over these links makes
me wonder if there's something about that kind of traffic that increases
the chances of causing these issues.  When I put VM network traffic over
the same NICs, they seem a lot less prone to failures, but also end up
pushing less traffic in general.

>> Switching to Broadcom would be a possibility, though it's tricky because
>> two of the NICs are onboard, so we'd need to replace the dual-port 1G
>> card with a quad-port 1G card.  Since you're saying you're all 10G,
>> maybe you don't know, but if you have any specific Broadcom 1G cards
>> you've had good fortune with, I'd be interested in knowing which models.
>>   Broadcom cards are rarely labeled as such which makes finding them a
>> bit more difficult than Intel ones.
> We've purchased a number of servers with Broadcom BCM957810A1008G, sold
> by Dell as QLogic 57810 dual 10Gb Base-T adapters, none of them going up
> & down like a yo-yo so far.
>> So far the one hypervisor with pci=nomsi has been quiet but that doesn't
>> mean it's fixed.  I need to give it 6 weeks or so. :)
> It'd be more like 6-9 months for us, making it terrible to debug it :-/

I had a bunch of these on relatively light VM load for 3 months for
"burn in" with no issues but they've been pretty aggressively failing
since I started to try to put real loads on them.  Still, it's odd
because some of the boxes with identical hardware and similar VM loads
have not yet blown up after 3 or more weeks, and maybe they won't for
several months.

Kevin Stange
Chief Technology Officer
Steadfast | Managed Infrastructure, Datacenter and Cloud Services
800 S Wells, Suite 190 | Chicago, IL 60607
312.602.2689 X203 | Fax: 312.602.2688
kevin at steadfast.net | www.steadfast.net