Re: [PATCH v7 10/11] KVM: MMU: collapse TLB flushes when zap all pages

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

On 05/23/2013 03:24 PM, Gleb Natapov wrote:
> On Thu, May 23, 2013 at 02:26:57PM +0800, Xiao Guangrong wrote:
>> On 05/23/2013 02:12 PM, Gleb Natapov wrote:
>>> On Thu, May 23, 2013 at 03:55:59AM +0800, Xiao Guangrong wrote:
>>>> kvm_zap_obsolete_pages uses lock-break technique to zap pages,
>>>> it will flush tlb every time when it does lock-break
>>>>
>>>> We can reload mmu on all vcpus after updating the generation
>>>> number so that the obsolete pages are not used on any vcpus,
>>>> after that we do not need to flush tlb when obsolete pages
>>>> are zapped
>>>>
>>>> Note: kvm_mmu_commit_zap_page is still needed before free
>>>> the pages since other vcpus may be doing locklessly shadow
>>>> page walking
>>>>
>>> Since obsolete pages are not accessible for lockless page walking after
>>> reload of all roots I do not understand why additional tlb flush is
>>
>> kvm_reload_remote_mmus() forces vcpus to leave guest mode, but if the
>> vcpu is not running on guest mode, it does nothing except set the request
>> bit. So, the vcpu can go lockless page walking after kvm_reload_remote_mmus()
>> return on other vcpu.
>>
>> Like this scenario:
>>
>> VCPU 0                              VCPU 1
>>                                  exit when it encounters #PF
>>
>> kvm_reload_remote_mmus(){
>>     set vcpu1->request bit;
>>
>>     do not send IPI due to
>>     vcpu 1 not running on guest mode
>>
>>                                  call page-fault handler then go lockless walking !!!
>>     return
>> }
>>
>>
>>> needed. Also why tlb flush should prevent lockless-walking from using
>>> the page? Making page unreachable from root_hpa does that, no?
>>
>> lockless-walking disables the interrupt and makes the vcpu state as
>> READING_SHADOW_PAGE_TABLES, this state is treated as GUEST_MODE,
>> kvm_flush_remote_tlbs() should send IPI to this vcpu in this case.
> 
> kvm_flush_remote_tlbs() uses the same make_all_cpus_request() as
> kvm_reload_remote_mmus() does, so why the same scenario you describe
> above cannot happen with kvm_flush_remote_tlbs()?


After call kvm_flush_remote_tlbs(), the page still exists on vcpu->root,
so we can not protect the page is being used by other vcpu.

But before call kvm_mmu_commit_zap_page(), the page has been deleted from
vcpu's page table, after call kvm_flush_remote_tlbs(), we can ensure that
other vcpus can not find these pages.