Re: [PATCH] x86/hvm: implement save/restore for posted interrupts

[Malcolm Crossley <malcolm.crossley@citrix.com>]

On 22/09/14 23:43, Tian, Kevin wrote:
>> From: Andrew Cooper
>> Sent: Friday, September 19, 2014 10:52 AM
>>
>> On 19/09/2014 14:39, Jan Beulich wrote:
>>>>>> On 19.09.14 at 15:32, <olaf@aepfle.de> wrote:
>>>> On Thu, Sep 18, Zhang, Yang Z wrote:
>>>>
>>>>> Zhang, Yang Z wrote on 2014-09-01:
>>>>>> Zhang, Yang Z wrote on 2014-08-08:
>>>>>>> Can you tell how you reproduce it? I am following the steps you told
>>>>>>> me previously, but I cannot reproduce it with this fixing.
>>>>>> Any update on this?
>>>>> Ping again.
>>>> I will have a look at this issue next week.
>>> Btw., Andrew Cooper hinted at the possibility of this series
>>>
>>> http://lists.xenproject.org/archives/html/xen-devel/2014-09/msg02753.html
>>>
>>> (particularly patch 2) having an effect on migration problems
>>> between APICV-capable and -incapable systems, i.e. perhaps
>>> worth a try in this context too.
>>>
>>> Jan
>>
>> The issue XenServer observes is that problems occur even when migrating
>> between two identical machines, one with APICV enabled, and one disabled.
>>
>> Given the now-discovered issues with Xen's x2apic handling, it is
>> entirely possible that the differences between the emulated and real
>> APICs was enough to cause issues.   And yet, it could be a mixture of
>> all of the issues identified so far.
>>
> 
> Have to say I didn't see any explicit message that this issue is related to 
> two machines with different capabilities (one with APICv enabled, and
> one disabled).
> 
> by reading original commit message:
> ----
> Saving and restoring a PVonHVM guest on a host which has the VMX
> "Posted Interrupt Processing" feature enabled will fail because the
> xen-platform-pci device does not receive interrupts anymore after
> restore. The reason is that the IRQs are not maintained in APIC_IRR,
> but in a separate PIR array. This info is lost during the save
> operation.
> ----
> 
> I really thought it's related to migration between two machines which
> both have APICv enabled. Not sure whether Yang realized this point.
> 

We have done some internal XenServer testing on this issue:

We install Citrix Windows PV drivers into a Windows Server 2008 32bit
VM, this is effectively a PVHVM guest.

After migrating the VM between two Haswell-EP servers, the Windows PV
drivers stop working and we are unable to access storage or network.

When we apply Olaf original patch then we don't lose storage or network
accesss via the PV drivers.

We are continuing to test but I thought we'd share that Olaf's original
patch resolves the PVHVM APIC-v issue for Xenserver.

Malcolm


> Thanks
> Kevin
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel
>