Re: [PATCH 3/6] Add KVM_CAP_DIRTY_QUOTA_MIGRATION and handle vCPU page faults.

[Shivam Kumar <shivam.kumar1@nutanix.com>]

On 20/11/21 1:51 am, Shivam Kumar wrote:
>
> On 20/11/21 1:36 am, Sean Christopherson wrote:
>> On Sat, Nov 20, 2021, Shivam Kumar wrote:
>>> On 18/11/21 11:27 pm, Sean Christopherson wrote:
>>>>> +        return -EINVAL;
>>>> Probably more idiomatic to return 0 if the desired value is the 
>>>> current value.
>>> Keeping the case in mind when the userspace is trying to enable it 
>>> while the
>>> migration is already going on(which shouldn't happen), we are returning
>>> -EINVAL. Please let me know if 0 still makes more sense.
>> If the semantics are not "enable/disable", but rather "(re)set the 
>> quota",
>> then it makes sense to allow changing the quota arbitrarily.
> I agree that the semantics are not apt. Will modify it. Thanks.
>>
>>>>> +    mutex_lock(&kvm->lock);
>>>>> +    kvm->dirty_quota_migration_enabled = enabled;
>>>> Needs to check vCPU creation.
>>> In our current implementation, we are using the
>>> KVM_CAP_DIRTY_QUOTA_MIGRATION ioctl to start dirty logging (through 
>>> dirty
>>> counter) on the kernel side. This ioctl is called each time a new 
>>> migration
>>> starts and ends.
>> Ah, and from the cover letter discussion, you want the count and 
>> quota to be
>> reset when a new migration occurs.  That makes sense.
>>
>> Actually, if we go the route of using kvm_run to report and update 
>> the count/quota,
>> we don't even need a capability.  Userspace can signal each vCPU to 
>> induce an exit
>> to userspace, e.g. at the start of migration, then set the desired 
>> quota/count in
>> vcpu->kvm_run and stuff exit_reason so that KVM updates the 
>> quota/count on the
>> subsequent KVM_RUN.  No locking or requests needed, and userspace can 
>> reset the
>> count at will, it just requires a signal.
>>
>> It's a little weird to overload exit_reason like that, but if that's 
>> a sticking
>> point we could add a flag in kvm_run somewhere.  Requiring an exit to 
>> userspace
>> at the start of migration doesn't seem too onerous.
> Yes, this path looks flaw-free. We will explore the complexity and how 
> we can simplify its implementation.
Is it okay to define the per-vcpu dirty quota and dirty count in the 
kvm_run structure itself? It can save space and reduce the complexity of 
the implemenation by large margin.