Re: [BUG] kvm: guest can not startup due to triple fault

[Jan Kiszka]

Am 18.10.2010 06:43, Dave Young wrote:
> On Sun, Oct 17, 2010 at 8:15 PM, Avi Kivity <avi@xxxxxxxxxx> wrote:
>>  On 10/17/2010 12:14 PM, Dave Young wrote:
>>>
>>> On Sun, Oct 17, 2010 at 5:34 PM, Avi Kivity<avi@xxxxxxxxxx>  wrote:
>>>>    On 10/17/2010 10:54 AM, Dave Young wrote:
>>>>>
>>>>>  linux guest can not startup normally, qemu keep rebooting.
>>>>>
>>>>>  host&    guest both use linux-2.6.36-rc8+ kernel
>>>>>
>>>>>  Add a dump_stack in x86.c around EXIT_SHUTDOWN code, I got:
>>>>>  [ 3924.503041] Pid: 1685, comm: qemu-system-x86 Not tainted
>>>>> 2.6.36-rc8+
>>>>>  #62
>>>>>  [ 3924.503045] Call Trace:
>>>>>  [ 3924.503055]  [<ffffffffa0176047>] handle_triple_fault+0x11/0x26
>>>>>  [kvm_intel]
>>>>>  [ 3924.503060]  [<ffffffffa01772d2>] vmx_handle_exit+0x3df/0x474
>>>>>  [kvm_intel]
>>>>>  [ 3924.503080]  [<ffffffffa021f203>]
>>>>> kvm_arch_vcpu_ioctl_run+0x95f/0xc43
>>>>>  [kvm]
>>>>>  [ 3924.503087]  [<ffffffff81503e54>] ?
>>>>> __mutex_unlock_slowpath+0x11c/0x131
>>>>>  [ 3924.503101]  [<ffffffffa021e1f9>] ? kvm_arch_vcpu_load+0x90/0xe6
>>>>> [kvm]
>>>>>  [ 3924.503112]  [<ffffffffa0210faf>] kvm_vcpu_ioctl+0x109/0x54a [kvm]
>>>>>  [ 3924.503117]  [<ffffffff810bee76>] ?
>>>>>  fire_user_return_notifiers+0x2c/0x36
>>>>>  [ 3924.503122]  [<ffffffff8106be93>] ? trace_hardirqs_on+0xd/0xf
>>>>>  [ 3924.503127]  [<ffffffff81002789>] ? do_notify_resume+0x717/0x73d
>>>>>  [ 3924.503131]  [<ffffffff8103994e>] ? finish_task_switch+0x69/0xda
>>>>>  [ 3924.503135]  [<ffffffff8110628c>] do_vfs_ioctl+0x4c3/0x50d
>>>>>  [ 3924.503139]  [<ffffffff8110632b>] sys_ioctl+0x55/0x77
>>>>>  [ 3924.503142]  [<ffffffff81002c02>] system_call_fastpath+0x16/0x1b
>>>>>
>>>>>  Anyone can help on this?
>>>>
>>>>  Please provide a trace (trace-cmd record -e kvm) leading to the
>>>> failure.
>>>>
>>>
>>> Hi, please see:
>>> http://userweb.kernel.org/~davey/tmp/trace.dat
>>>
>>
>> Was this done with an old version of trace-cmd?  Please use the latest git.
>>  Make sure you have udis86 and udis86-devel installed when you build it.
> 
> Hi, the trace-cmd was build from git version, but without udis86.
> 
> I updated trace.dat on the same url, is it a correct one?
> 

IIUC, plugin_kvm is only involved in reporting, not recording, so udis86
shouldn't matter.

I had a look out of curiosity (just managed to make my trace tools work
smoothly), and the data is properly parsed by kernelshark here. There is
a strange NMI storm going on right before the triple fault (on some MSR
write) hits the fan.

Jan


>>
>> --
>> error compiling committee.c: too many arguments to function
>>
>>
> 
> 
>

Attachment: signature.asc
Description: OpenPGP digital signature