Re: [kvm-devel] [PATCH] export notifier #1

[Avi Kivity <avi@qumranet.com>]

Andrea Arcangeli wrote:
> On Thu, Jan 24, 2008 at 07:56:43AM +0200, Avi Kivity wrote:
>   
>> What I need is a list of (mm, va) that map the page.  kvm doesn't have 
>> access to that, export notifiers do.  It seems reasonable that export 
>> notifier do that rmap walk since they are part of core mm, not kvm.
>>     
>
> Yes. Like said in earlier email we could ignore the slowdown and
> duplicate the mm/rmap.c code inside kvm, but that looks a bad layering
> violation and it's unnecessary, dirty and suboptimal IMHO.
>
>   

Historical note: old kvm versions (like the what will eventually ship in 
2.6.24) have a page-based rmap (hooking the rmap list off 
page->private).  We changed that to an mm based rmap since page->private 
is not available when kvm maps general userspace memory.


>> Alternatively, kvm can change its internal rmap structure to be page based 
>> instead of (mm, hva) based.  The problem here is to size this thing, as we 
>> don't know in advance (when the kvm module is loaded) whether 0% or 100% 
>> (or some value in between) of system memory will be used for kvm.
>>     
>
> Another issue is that for things like the page sharing driver, it's
> more convenient to be able to know exactly which "sptes" belongs to a
> certain userland mapping, and only that userland mapping (not all
> others mappings of the physical page). So if the rmap becomes page
> based, it'd be nice to still be able to find the "mm" associated with
> that certain spte pointer to skip all sptes in the other "mm" during
> the invalidate.
>   

You also need the mm (or rather, the kvm structure, but they have a 1:1 
relationship) to be able to lock and maintain the shadow structures 
properly.

-- 
Any sufficiently difficult bug is indistinguishable from a feature.