Re: [PATCH v2 00/16] KVM: MMU: fast page fault

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

On 04/14/2012 11:37 AM, Takuya Yoshikawa wrote:

> On Fri, 13 Apr 2012 18:05:29 +0800
> Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> wrote:
> 
>> Thanks for Avi and Marcelo's review, i have simplified the whole things
>> in this version:
>> - it only fix the page fault with PFEC.P = 1 && PFEC.W = 0 that means
>>   unlock set_spte path can be dropped.
>>
>> - it only fixes the page fault caused by dirty-log
>>
>> In this version, all the information we need is from spte, the
>> SPTE_ALLOW_WRITE bit and SPTE_WRITE_PROTECT bit:
>>    - SPTE_ALLOW_WRITE is set if the gpte is writable and the pfn pointed
>>      by the spte is writable on host.
>>    - SPTE_WRITE_PROTECT is set if the spte is write-protected by shadow
>>      page table protection.
>>
>> All these bits can be protected by cmpxchg, now, all the things is fairly
>> simple than before. :)
> 
> Well, could you remove cleanup patches not needed for "lock-less" from
> this patch series?
> 
> I want to see them separately.
> 
> Or everything was needed for "lock-less" ?
> 


The cleanup patches do the prepare work for fast page fault, the later path will
be easily implemented, for example, the for_each_spte_rmap patches make "store
more bits in rmap" patch doing little change since spte_list_walk is removed.

>> Performance test:
>>
>> autotest migration:
>> (Host: Intel(R) Xeon(R) CPU           X5690  @ 3.47GHz * 12 + 32G)
> 
> Please explain what this test result means, not just numbers.
> 
> There are many aspects:
> 	- how fast migration can converge/complete
> 	- how fast programs inside the guest can run during migration:
> 	  -- throughput
> 	  -- latency
> 	- ...
> 


The result is rather straightforward, i think explanation is not needed.

> I think lock-less will reduce latency a lot, but not sure about convergence:
> why it became fast?
> 


It is hard to understand? It is faster since it can be parallel.

>> - For ept:
>>
>> Before:
>>                     smp2.Fedora.16.64.migrate
>> Times   .unix      .with_autotest.dbench.unix     total
>>  1       104           214                         323
>>  2       68            238                         310
>>  3       68            242                         314
>>
>> After:
>>                     smp2.Fedora.16.64.migrate
>> Times   .unix      .with_autotest.dbench.unix     total
>>  1       101           190                         295
>>  2       67            188                         259
>>  3       66            217                         289
>>
> 
> As discussed on v1-threads, the main goal of this "lock-less" should be
> the elimination of mmu_lock contentions
> 
> So what we should measure is latency.
> 


I think the test of migration time is fairly enough to see the effect.