On Wed, 17 Mar 2021 10:55:00 +0000, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: > > On 17/03/21 11:53, Marc Zyngier wrote: > > On Wed, 17 Mar 2021 10:40:23 +0000, > > Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: > >> > >> On 17/03/21 10:10, Marc Zyngier wrote: > >>>> @@ -366,7 +366,7 @@ static int hyp_map_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep, > >>>> if (WARN_ON(level == KVM_PGTABLE_MAX_LEVELS - 1)) > >>>> return -EINVAL; > >>>> - childp = (kvm_pte_t *)get_zeroed_page(GFP_KERNEL); > >>>> + childp = (kvm_pte_t *)get_zeroed_page(GFP_KERNEL_ACCOUNT); > >>> No, this is wrong. > >>> > >>> You cannot account the hypervisor page tables to the guest because we > >>> don't ever unmap them, and that we can't distinguish two data > >>> structures from two different VMs occupying the same page. > >> > >> If you never unmap them, there should at least be a shrinker to get > >> rid of unused pages in the event of memory pressure. > > > > We don't track where these pages are coming from or whether they can > > safely be unmapped. Until we can track such ownership and deal with > > page sharing, these mappings have to stay, > > > > At most, this represent the amount of memory required to map the whole > > of the linear mapping. > > Ah, these are the EL2 pages, not the stage2 page tables, right? If > so, sorry for the noise. Yes, EL2 page tables when running non-VHE. VHE doesn't have that problem for obvious reasons. Stage-2 page tables can be completely discarded at any point, and the MMU notifiers already deal with that. M. -- Without deviation from the norm, progress is not possible.
