Re: [PATCHv4 2/2] x86/ept: defer the invalidation until the p2m lock is released

[David Vrabel <david.vrabel@citrix.com>]

On 14/12/15 14:39, David Vrabel wrote:
> Holding the p2m lock while calling ept_sync_domain() is very expensive
> since it does a on_selected_cpus() call.  IPIs on many socket machines
> can be very slows and on_selected_cpus() is serialized.
> 
> Defer the invalidate until the p2m lock is released.  Since the processor
> may cache partial translations, we also need to make sure any page table
> pages to be freed are not freed until the invalidate is complete.  Such
> pages are temporarily stored in a list.

This breaks PoD because we check for zeroed pages without doing a sync
after the type change.  This allows other VCPUs to write to a page that
is now in the pod pool.  The write will both be lost and will corrupt
another zero page when the page is used to populate another GFN.

David