Re: [Xen-devel] HVM hangs system on AMD A10-6800K + Hudson-D4

[Darren Shepherd <darren.s.shepherd@xxxxxxxxx>]

On Aug 19, 2013, at 3:31 AM, George Dunlap <George.Dunlap@xxxxxxxxxxxxx> wrote:

> On Sun, Aug 18, 2013 at 3:14 AM, Darren Shepherd
> <darren.s.shepherd@xxxxxxxxx> wrote:
>> 
>> On Wed, Aug 14, 2013 at 11:22 AM, Hurwitz, Sherry <sherry.hurwitz@xxxxxxx>
>> wrote:
>>> 
>>> I would like to help you because my team at AMD does not see this problem
>>> but maybe there is something unique in your environment that we have not
>>> tested with.  So you are using the Xen hypervisor from XenServer 6.2 and you
>>> have tested with Centos and Unbuntu guests?  Could you tell me your Guest
>>> configurations (vcpus and memory per guest).  Any information you can
>>> provide would be appreciated.
>> 
>> 
>> I found another user on the xen-users list who experienced this same problem
>> [1] with the same APU but with a different motherboard that has a Hudson-D3.
>> He suggested compiling and running Xen from the latest source and that
>> seemed to work.  I haven't extensively tested this, but before the system
>> would freeze under a minute and now I've ran it for a couple hours with no
>> issues.  So I have ran all of the following dom0 setups and they have
>> failed:
>> 
>> XenServer 6.2
>> CentOS 6.4
>> Ubuntu 13.04
>> Ubuntu 13.04 + Xen 4.3 compiled from source
>> 
>> So there is something that was checked in since Xen 4.3 that seems to fix
>> the issue or did something to mask the issue.  Specifically what I was doing
>> that made it fail was that I was running a Windows XP installation from ISO.
>> During the installation, around the time of entering your product key or
>> configuring networking the system would hang.  On XenServer the domain was
>> configured just however XenServer does it when selecting the OS type as
>> Windows XP.  On CentOS and Ubuntu I was using the smallest xl cfg I could
>> which was basically name, builder, disks, vif, memory.
>> 
>> I started debugging the issue further on Ubuntu 13.04 and got it so that I
>> can easily reproduce it.  I'm running basically a fresh install of 13.04
>> 64-bit with "apt-get install xen-hypervisor-4.2."  I then disabled all
>> features I could in the BIOS to just start ruling things out.  So I turned
>> off SMP, IOMMU, CPB, NX, C6, and PowerNow!.  All other settings in the BIOS
>> are default.  If I ran the following xl conf
>> 
>> name = 'blank'
>> builder = 'hvm'
>> memory = 1024
>> 
>> I then run "xl create -f blank.cfg" so that domain creates, and then about a
>> second later I run "xl create -f blank.cfg 'name="blank2"'"  Its about 50/50
>> if the system freezes on the first "xl create" or the second.  So basically,
>> I can reproduce the issue with no domU OS.
>> 
>> Unfortunately I don't have a COM port plugged into the motherboard and I
>> haven't had time to go run and buy a com port to plug into the motherboard
>> so I have no clue if Xen is logging any errors, but I definitly don't get
>> anything on the linux console or in logs.  Below are the specs of the system
>> I have
>> 
>> AMD A10-6800K
>> G.SKILL Ripjaws X Series 32GB (4 x 8GB) DDR3 2400 <-- Running at 1333mhz
>> (which is BIOS default)
>> ASUS F2A85-M PRO FM2 AMD A85X
>> 
>> I'm particularly interested in what has made this work in latest git because
>> I'd like the patch to be backed ported into Xen 4.2.x so that CentOS can
>> pick it up.
> 
> Andrew, Jan, any idea which patch might be the culprit for fixing this 
> problem?
> 
> Barring that, Darren, you can try to do a git bisect to figure out
> when things started working again:
> 
> http://webchick.net/node/99
> 
> -George

Okay, so I compiled, ran and tested everything between RELEASE-4.3.0 (bad) with 
master (good).   Unfortunately I came to the conclusion it's not a patch but 
instead it's the debug flag.  So if I build RELEASE-4.3.0 it fails my test case 
because because Config.mk has "debug ?= n". If I then just rebuild xen.gz with 
"make debug=y" it works.

I really couldn't figure out a way to do bisect with debugging.  I tried 
compiling xen.gz with debug on and then just compiling a few obj with debug 
off, but that produced weird results and I have no clue if that is really even 
valid to have a linked file with half debug .o's and half not.  

I'll keep fooling around and see if I can narrow this down any further.  So my 
simple test fails with debug off, but I don't know if debug on just makes it 
less likely to happen.  I'm going to run more extensive tests later.  I'll also 
try to get a serial port setup on the box.

Darren 
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel