Re: [PATCH v3 0/6] KVM: MMU: fast invalidate all mmio sptes

[Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>]

Sorry for the delay reply since i was on vacation.

On 06/15/2013 10:22 AM, Takuya Yoshikawa wrote:
> On Thu, 13 Jun 2013 21:08:21 -0300
> Marcelo Tosatti <mtosatti@redhat.com> wrote:
> 
>> On Fri, Jun 07, 2013 at 04:51:22PM +0800, Xiao Guangrong wrote:
> 
>> - Where is the generation number increased?
> 
> Looks like when a new slot is installed in update_memslots() because
> it's based on slots->generation.  This is not restricted to "create"
> and "move".

Yes. It reuses slots->generation to avoid unnecessary synchronizations
(RCU, memory barrier).

Increasing mmio generation number in the case of "create" and "move"
is ok - it is no addition work unless mmio generation number is overflow
which is hardly triggered (since the valid mmio generation number is
large enough and zap_all is scale well now.) and the mmio spte is updated
only when it is used in the future.

> 
>> - Should use spinlock breakable code in kvm_mmu_zap_mmio_sptes()
>> (picture guest with 512GB of RAM, even walking all those pages is
>> expensive) (ah, patch to remove kvm_mmu_zap_mmio_sptes does that).
>> - Is -13 enough to test wraparound? Its highly likely the guest has 
>> not began executing by the time 13 kvm_set_memory_calls are made
>> (so no sptes around). Perhaps -2000 is more sensible (should confirm
>> though).
> 
> In the future, after we've tested enough, we should change the testing
> code to be executed only for some debugging configs.  Especially, if we
> change zap_mmio_sptes() to zap_all_shadows(), very common guests, even
> without huge memory like 512GB, can see the effect induced by sudden page
> faults unnecessarily.
> 
> If necessary, developers can test the wraparound code by lowering the
> max_gen itself anyway.

I agree.

> 
>> - Why remove "if (change == KVM_MR_CREATE) || (change
>> ==  KVM_MR_MOVE)" from kvm_arch_commit_memory_region?
>> Its instructive.
> 
> There may be a chance that we miss generation wraparounds if we don't
> check other cases: seems unlikely, but theoretically possible.
> 
> In short, all memory slot changes make mmio sptes stored in shadow pages
> obsolete, or zapped for wraparounds, in the new way -- am I right?

Yes. You are definitely right. :)

Takuya-san, thank you very much for you answering the questions for me and thanks
all of you for patiently reviewing my patches.

Marcelo, your points?