Re: [PATCH 2/2] KVM: Scalable memslots implementation

[Paolo Bonzini]

On 03/02/21 14:52, David Hildenbrand wrote:
> > However, note that the TDP MMU does not need an rmap at all.  Since that
> > one is getting ready to become the default, the benefits of working on
> > the rmap would be quite small and only affect nested virtualization.
>
> Right, but we currently always have to allocate it.
>
> 8 bytes per 4k page, 8 bytes per 2M page, 8 bytes per 1G page.
>
> The 4k part alone is 0.2% of the memblock size.
>
> For a 1 TB memslot we might "waste" > 2 GB on rmap arrays.
>
> (that's why I am asking :) )

Yes, we're in violent agreement. :)  I mean that the simplest solution 
would be to allocate it lazily when building a non-TDP root.

Paolo