Re: [Qemu-devel] Re: [PATCH 0/7] KVM SMP support, early version

[Jan Kiszka]

Glauber Costa wrote:
> On Tue, Dec 1, 2009 at 10:17 AM, Jan Kiszka <address@hidden> wrote:
>> Avi Kivity wrote:
>>> On 11/30/2009 07:30 PM, Jan Kiszka wrote:
>>>>>> No.  But what's wrong with on_vcpu?
>>>>>>
>>>>> intrinsically racy. signal passing slow down things, etc.
>>>>>
>>>>> That said, as I've stated many times: I don't believe there's anything
>>>>> fundamentally wrong with on_vcpu. But we might get benefits from a 
>>>>> re-design
>>>>> of things to avoid it whenever possible. (just like the vcpu_reset()
>>>>> I've just posted)
>>>>>
>>>>>
>>>> If you don't want immediate execution of update_guest_debug, save the
>>>> state that shall be transferred, set some flag, and run the transfer
>>>> before guest entry inside the vcpu threads (after putting the registers
>>>> as older kernels may otherwise overwrite the flags register). Should
>>>> work, may even avoid redundant calls during a gdb session.
>>>>
>>> There's no guarantee the vcpu will ever exit to qemu, so you have to
>>> signal the vcpu thread anyway.  When you do that, you might as well load
>>> the new state.
>> Debugging is special here as it involves vmstop before you start playing
>> with the debug settings. But I may also oversee some corner case right now.
>>
> 
> I imagined so. In this case, it might be better to set a flag before
> vmstop, and then
> honor it on vcpu entry. I had this feeling that in most cases where we
> current signal the
> vcpu, the VM will be already stopped anyway, so we don't need to resignal.
> 

Do not set the flag before the machine stop, otherwise you risk to race
with vcpus that exit and re-enter briefly before that. This manipulation
should really be done under vmstop protection.

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux