Re: x86/AMD: Nested hvm crashes in 4.3

["Egger, Christoph" <chegger@amazon.de>]

Running a PV guest as L2 guest makes no difference how the p2m
code is used in L0 Xen (L0 == OS that runs on bare metal hardware).

I assume you use the default settings which means you use NPT-on-NPT.
Try shadow-on-npt. You can do this with

  cpuid="host,svm_npt=0"

in the guest config file. Then in the L1 guest you should see
the NPT svm feature bit not available.
Then launch a l2 guest and check if it still crashes.

I agree with Jan: Please provide the crash logs he requested.

Christoph

On 27.06.13 11:20, Suravee Suthikulpanit wrote:
> On 6/27/2013 3:22 AM, Jan Beulich wrote:
>>>>> On 27.06.13 at 02:24, Suravee Suthikulanit
>>>>> <suravee.suthikulpanit@amd.com> wrote:
>>> I have found an issue in where the system crash right when I start
>>> another HVM guest inside an HVM guest.  I have traced back to the patch
>>> which the issue started.
>>>
>>> commit f1bde87fc08ce8c818a1640a8fe4765d48923091
>>> Author: Jan Beulich <jbeulich@suse.com>
>>> Date:   Fri Feb 8 11:06:04 2013 +0100
>>>
>>>       x86: debugging code for testing 16Tb support on smaller memory
>>> systems
>>>
>>>       Signed-off-by: Jan Beulich <jbeulich@suse.com>
>>>       Acked-by: Keir Fraser <keir@xen.org>
>> We had issues exposed by this patch before, but any such issue
>> would just have been masked before that patch (and would
>> surface on a system with more than 5Tb of memory anyway).
> 
> The system I am having the issue has 48GB of memory.
> 
>> So it is very unlikely for the patch itself to be at fault.
> 
> I have traced the issue and found that the system crashing starts from
> this commit id and onward.
> (i.e. The system does not crash with commit id
> ed759d20249197cf87b338ff0ed328052ca3b8e7)
> So, I am still believe that this patch has somehow triggered the issue.
> 
>> Furthermore, the crash them supposedly is because of the code
>> added conditional upon NDEBUG, and hence would (on a smaller
>> memory system) otherwise not surface at all for a production
>> (debug=n) build.
>>
>>> The issue doesn't reproduce when starting a PV (L2) guest inside an HVM
>>> (L1) guest.
>> "Does not" or "does"? In the former case - what is this supposed to
>> tell me?
> 
> What I am trying to say here is that the system_does not_ crash when
> starting the PV guest as level 2 guest.
> This is meant to be another data point to help analyzing the issue.
> 
>> In any case - without you sharing technical details (register/stack
>> dump of the crash at the very least) I don't think I have anything
>> at hand to look for possible problems.
> 
> At this point, I am just reporting the issue.  I have not been able to
> get the crash dump because the system immediately reboot.
> I'll try to boot Xen with "noreboot" and inspect the log for more clues.
> Any suggestions are welcome.
> 
> Thank you,
> 
> Suravee