Re: [PATCH v2 0/5] Export offsets of VMCS fields as note information for kdump

[Yanfei Zhang <zhangyanfei@cn.fujitsu.com>]

于 2012年05月21日 17:36, Avi Kivity 写道:
> On 05/21/2012 12:08 PM, Yanfei Zhang wrote:
>> 于 2012年05月21日 16:34, Avi Kivity 写道:
>>> On 05/21/2012 05:32 AM, Yanfei Zhang wrote:
>>>> 于 2012年05月21日 01:43, Avi Kivity 写道:
>>>>> On 05/16/2012 10:50 AM, zhangyanfei wrote:
>>>>>> This patch set exports offsets of VMCS fields as note information for
>>>>>> kdump. We call it VMCSINFO. The purpose of VMCSINFO is to retrieve
>>>>>> runtime state of guest machine image, such as registers, in host
>>>>>> machine's crash dump as VMCS format. The problem is that VMCS internal
>>>>>> is hidden by Intel in its specification. So, we slove this problem
>>>>>> by reverse engineering implemented in this patch set. The VMCSINFO
>>>>>> is exported via sysfs to kexec-tools just like VMCOREINFO.
>>>>>>
>>>>>> Here are two usercases for two features that we want.
>>>>>>
>>>>>> 1) Create guest machine's crash dumpfile from host machine's crash dumpfile
>>>>>>
>>>>>> In general, we want to use this feature on failure analysis for the system
>>>>>> where the processing depends on the communication between host and guest
>>>>>> machines to look into the system from both machines's viewpoints.
>>>>>>
>>>>>> As a concrete situation, consider where there's heartbeat monitoring
>>>>>> feature on the guest machine's side, where we need to determine in
>>>>>> which machine side the cause of heartbeat stop lies. In our actual
>>>>>> experiments, we encountered such situation and we found the cause of
>>>>>> the bug was in host's process schedular so guest machine's vcpu stopped
>>>>>> for a long time and then led to heartbeat stop.
>>>>>>
>>>>>> The module that judges heartbeat stop is on guest machine, so we need
>>>>>> to debug guest machine's data. But if the cause lies in host machine
>>>>>> side, we need to look into host machine's crash dump.
>>>>>
>>>>> Do you mean, that a heartbeat failure in the guest lead to host panic?
>>>>>
>>>>> My expectation is that a problem in the guest will cause the guest to
>>>>> panic and perhaps produce a dump; the host will remain up.
>>>>>
>>>>
>>>> The point is that before our investigation, we didn't know which side 
>>>> leads to this buggy situation. Maybe a bug in host machine or the guest
>>>> machine itself causes a heartbeat failure.
>>>
>>> How can a guest bug cause a host panic?
>>>
>>>> So we want to get both host machine's crash dump and guest machine's
>>>> crash dump *at the same time*. Then we could use userspace tools to
>>>> get guest machine crash dump from host machine's and analyse them
>>>> separately to find which side causes the problem.
>>>>
>>>
>>> If the guest caused the problem, there would be no panic; therefore
>>> there was a host bug.
>>>
>>
>> Yes, a guest bug cannot cause a host panic. When heartbeat stops in guest
>> machine, we could trigger the host dump mechanism to work. This is because
>> we want to get the status of both host and guest machine at the same time
>> when heartbeat stops in guest machine. Then we can look for bug reasons
>> from both host machine's and guest machine's views.
> 
> That sounds like a bad idea. Can you explain in what situation it makes
> sense for a guest to stop the host (and all other guests running on it)
> rather than just restarting the failed services (on the host or other
> guests)?
> 

We never do this on customer's environment which maybe a host with many guests
running on it. We do this on another environment to reproduce the buggy
situation; or we do this in testing phase on development environment towards
production one on the customer's site.

>>>>>> Without this feature, we first create guest machine's dump and then
>>>>>> create host mahine's, but there's only a short time between two
>>>>>> processings, during which it's unlikely that buggy situation remains.
>>>>>>
>>>>>> So, we think the feature is useful to debug both guest machine's and
>>>>>> host machine's sides at the same time, and expect we can make failure
>>>>>> analysis efficiently.
>>>>>>
>>>>>> Of course, we believe this feature is commonly useful on the situation
>>>>>> where guest machine doesn't work well due to something of host machine's.
>>>>>>
>>>>>> 2) Get offsets of VMCS information on the CPU running on the host machine
>>>>>>
>>>>>> If kdump doesn't work well, then it means we cannot use kvm API to get
>>>>>> register values of guest machine and they are still left on its vmcs
>>>>>> region. In the case, we use crash dump mechanism running outside of
>>>>>> linux kernel, such as sadump, a firmware-based crash dump. Then VMCS
>>>>>> information is then necessary.
>>>>>
>>>>> Shouldn't sadump then expose the VMCS offsets? Perhaps bundling them
>>>>> into its dump file?
>>>>>
>>>>
>>>> Firmware-based crash dump doesn't concern the os running on the machine.
>>>> So it will not do any os handling when machine crashes.
>>>
>>> Seems to me the VMCS offsets are OS independent.
>>>
>> Hmm, you mean we could get VMCS offsets in sadump itself?
>> But I think if we just export VMCS offsets in kernel, we could use the current
>> existing dump tools with no or just very tiny change. I think this could be
>> a more general mechanism than making changes in all kinds of dump tools.
> 
> The sadump tool generates a core file with the OS image, right? Can it
> not attach the offsets to a note, just like you propose for kdump?
> 

Both are right.