Re: [PATCH 6/6] KVM: Dirty memory tracking for performant checkpointing and improved live migration

["Radim Krčmář" <rkrcmar@redhat.com>]

2016-04-29 18:47+0000, Cao, Lei:
> On 4/28/2016 2:08 PM, Radim Krčmář wrote:
>> 2016-04-26 19:26+0000, Cao, Lei:
>> * Is there a reason to call KVM_ENABLE_MT often?
> 
> KVM_ENABLE_MT can be called multiple times during a protected
> VM's lifecycle in a checkpointing system. A protected VM has two
> instances, primary and secondary. Memory tracking is only enabled on
> the primary. When we do a polite failover, memory tracking is
> disabled on the old primary and enabled on the new primary. Memory
> tracking is also disabled when the secondary goes away, in which case
> checkpoint cycle stops and there is no need for memory tracking. When
> the secondary comes back, memory tracking is re-enabled and the two
> instances sync up and checkpoint cycle starts.

Makes sense.

>> * How significant is the benefit of MT_FETCH_WAIT?
> 
> This allows the user thread that harvest dirty pages to park instead
> of doing busy wait when there is no or very few dirty pages.

True, mandatory polling could be ugly.

>> * When would you disable MT_FETCH_REARM?
> 
> In a checkpointing system, dirty pages are harvested after the VM is
> paused. Userspace can choose to rearm the write traps all at once after
> all the dirty pages have been fetched using KVM_REARM_DIRTY_PAGES, in
> which case the traps don't need to be armed during each fetch.

Ah, it makes a difference when you don't plan to run the VM again.

I guess all three of them are worth it.
(Might change my mind when I gain better understanding.)

>> * What drawbacks had an interface without explicit checkpointing cycles?
> 
> Checkpointing cycle has to be implemented in userspace to use this
> interface. 

But isn't the explicit cycle necessary only in userspace?
The dirty list could be implemented as a circullar buffer, so KVM
wouldn't need an explicit notification about the new cycle -- the
userspace would just drain all dirty pages and unpause vcpus.
(Quiesced can be stateless one-time kick of waiters instead.)

Thanks.