Re: [RFC Patch v3 18/18] x86/hvm: Always set pending event injection when loading VMC[BS] state.

[Wen Congyang <wency@cn.fujitsu.com>]

On 09/11/2014 06:35 PM, Tim Deegan wrote:
> At 10:06 -0500 on 10 Sep (1410340016), Aravind Gopalakrishnan wrote:
>> Hi,
>> You mention that this 'fix' is just copied for svm. But you have not 
>> seen the problem of "VM_ENTRY_INTR_INFO may be valid" (whose svm 
>> equivalent is "vmcb->eventinj.bytes is valid").
>> My concern is that we should test colo mode for svm first, since, if the 
>> problem is never really seen on svm, then fix may not be _necessary_
> 
> I think it's obvious by inspection of the code (and indeed the patch)
> that SVM has the equivalent bug.
> 
>> At this point, my problems are with test setups. I can help testing 
>> scenarios, but as Wen had mentioned, 'colo testing' might be the way to 
>> test.
>> So, if I can get some pointers to how I can reproduce the issue at hand, 
>> then it would be very helpful.
>>
>> (Tim had mentioned we could try to simulate it by running a guest that 
>> takes lot of faults and save-restore another guest over it. However,
>> I am not having much luck following this route. I got a hvm guest to 
>> continuously take SW exceptions on all vcpus and tried to save-restore.
>> I can't see vmcb->eventinj.bytes containing any valid info.)
> 
> Although I think that getting COLO running on SVM is a gret idea, it
> shouldn't block acceptance of this fix.  I've attached a program
> that tests for the bug.  Run it against any HVM guest (and then
> destroy the guest because the test will have corrupted its state).

Thanks for providing test program.
I run it against a HVM guest, and unpause the guest by hand. The guest
doesn't response anymore. Apply this patch and test it again, the problem
is gone.

Wen Congyang

> 
> Cheers,
> 
> Tim.
>