Re: [PATCH 5/6] KVM: Separate rmap_pde from kvm_lpage_info->write_count

[Avi Kivity <avi@redhat.com>]

On 06/28/2012 06:45 AM, Takuya Yoshikawa wrote:
> On Thu, 28 Jun 2012 11:12:51 +0800
> Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> wrote:
> 
>> >  struct kvm_arch_memory_slot {
>> > +	unsigned long *rmap_pde[KVM_NR_PAGE_SIZES - 1];
>> >  	struct kvm_lpage_info *lpage_info[KVM_NR_PAGE_SIZES - 1];
>> >  };
>> > 
>> 
>> It looks little complex than before - need manage more alloc-ed/freed buffers.
> 
> Actually I want to integrate rmap and rmap_pde in the future:
> 
> 	rmap[KVM_NR_PAGE_SIZES]

That's a good direction.

> 
> For this we need to modify some unrelated ppc code, so I just
> avoided the integration in this series.
> 
> Note: write_count: 4 bytes, rmap_pde: 8 bytes.  So we are wasting
> extra paddings by packing them into lpage_info.

The wastage is quite low since it's just 4 bytes per 2MB.

> 
>> Why not just introduce a function to get the next rmap? Something like this:
> 
> I want to eliminate this kind of overheads.

I don't think the overhead is significant.  rmap walk speed is largely a
function of cache misses IMO, and we may even be adding cache misses by
splitting lpage_info.

But I still think it's the right thing since it simplifies the code.
Maybe we should add a prefetch() on write_count do mitigate the
overhead, if it starts showing up in profiles.

-- 
error compiling committee.c: too many arguments to function