[PATCH v3 00/23] KVM: Extend Eager Page Splitting to the shadow MMU

[David Matlack]

On Wed, Apr 13, 2022 at 11:28 AM Sean Christopherson <seanjc at google.com> wrote:
>
> On Wed, Apr 13, 2022, David Matlack wrote:
> > On Wed, Apr 13, 2022 at 01:02:51AM +0000, Sean Christopherson wrote:
> > > There will be one wart due to unsync pages needing @vcpu, but we can pass in NULL
> > > for the split case and assert that @vcpu is non-null since all of the children
> > > should be direct.
> >
> > The NULL vcpu check will be a little gross,
>
> Yeah, I would even call it a lot gross :-)
>
> > but it should never trigger in practice since eager page splitting always
> > requests direct SPs. My preference has been to enforce that in code by
> > splitting out
>
> It still is enforced in code, just at different points.  The split version WARNs
> and continues after finding a page, the below WARNs and rejects _while_ finding
> the page.
>
> Speaking of WARNs, that reminds me... it might be worth adding a WARN in
> kvm_mmu_get_child_sp() to document (and detect, but more to document) that @direct
> should never encounter an page with unsync or unsync_children, e.g.
>
>         union kvm_mmu_page_role role;
>         struct kvm_mmu_page *sp;
>
>         role = kvm_mmu_child_role(sptep, direct, access);
>         sp = kvm_mmu_get_page(vcpu, gfn, role);
>
>         /* Comment goes here about direct pages in shadow MMUs? */
>         WARN_ON(direct && (sp->unsync || sp->unsync_children));
>         return sp;
>
> The indirect walk of FNAME(fetch)() handles unsync_children, but none of the other
> callers do.  Obviously shouldn't happen, but especially in the huge page split
> case it took me a second to understand exactly why it can't happen.

Will do.

>
> > but I can see the advantage of your proposal is that eager page splitting and
> > faults will go through the exact same code path to get a kvm_mmu_page.
> > __kvm_mmu_find_shadow_page(), but I can see the advantage of your
> > proposal is that eager page splitting and faults will go through the
> > exact same code path to get a kvm_mmu_page.
> >
> > >
> > >             if (sp->unsync) {
> > >                     if (WARN_ON_ONCE(!vcpu)) {
> > >                             kvm_mmu_prepare_zap_page(kvm, sp,
> > >                                                      &invalid_list);
> > >                             continue;
> > >                     }
> > >
> > >                     /*
> > >                      * The page is good, but is stale.  kvm_sync_page does
> > >                      * get the latest guest state, but (unlike mmu_unsync_children)
> > >                      * it doesn't write-protect the page or mark it synchronized!
> > >                      * This way the validity of the mapping is ensured, but the
> > >                      * overhead of write protection is not incurred until the
> > >                      * guest invalidates the TLB mapping.  This allows multiple
> > >                      * SPs for a single gfn to be unsync.
> > >                      *
> > >                      * If the sync fails, the page is zapped.  If so, break
> > >                      * in order to rebuild it.
> > >                      */
> > >                     if (!kvm_sync_page(vcpu, sp, &invalid_list))
> > >                             break;
> > >
> > >                     WARN_ON(!list_empty(&invalid_list));
> > >                     kvm_flush_remote_tlbs(kvm);
> > >             }