Re: [Qemu-devel] [RFC] Host vs Guest memory allocation

[Avi Kivity]

On 04/12/2010 06:56 PM, Alexander Graf wrote:
> For fully system emulation on the other hand I can imagine quite some nice 
> tricks one could pull.
>
> On PPC hosts you get a huge number of VSIDs that are basically like tags on the 
> TLB. So if you'd give every x86 page table one VSID you'd potentially have 
> really great and fast shadow PTEs.
>    

You mean, if you have lots of ppc machines but no x86?

smp would be a problem because of the relaxed memory model (of course 
tcg needs a lot of work before it can do smp anyway).

> On x86 hosts you can just keep several page tables around. You can then map for 
> example every combination of guest VSIDs to one page table each.
>    

Yeah.

> I'm sure there are similar fun things you can do with the other supported 
> archs. The hard part is to come up with something generic enough so it works on 
> all hosts and guests with little effort. Oh well :)
>    

Well, x86 page tables are pretty flexible, the memory model is strict, 
the atomics are rich, and you have both unaligned and trapping 
accesses.  So if you restrict yourself to x86 hosts I think you can do 
anything with page size >= 4k.

Run both user and kernel mode in guest user mode, do tcg and mmu in 
kernel mode.  Should be fun.

--
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.