On Fri, Jul 7, 2023 at 12:58 AM Isaku Yamahata <isaku.yamahata@xxxxxxxxx> wrote:
>
> On Thu, Jul 06, 2023 at 01:52:08PM +0900,
> David Stevens <stevensd@xxxxxxxxxxxx> wrote:
>
> > On Wed, Jul 5, 2023 at 7:17 PM Yu Zhang <yu.c.zhang@xxxxxxxxxxxxxxx> wrote:
> > >
> > > On Tue, Jul 04, 2023 at 04:50:50PM +0900, David Stevens wrote:
> > > > From: David Stevens <stevensd@xxxxxxxxxxxx>
> > > >
> > > > Stop passing FOLL_GET to __kvm_follow_pfn. This allows the host to map
> > > > memory into the guest that is backed by un-refcounted struct pages - for
> > > > example, higher order non-compound pages allocated by the amdgpu driver
> > > > via ttm_pool_alloc_page.
> > >
> > > I guess you mean the tail pages of the higher order non-compound pages?
> > > And as to the head page, it is said to be set to one coincidentally[*],
> > > and shall not be considered as refcounted. IIUC, refcount of this head
> > > page will be increased and decreased soon in hva_to_pfn_remapped(), so
> > > this may not be a problem(?). But treating this head page differently,
> > > as a refcounted one(e.g., to set the A/D flags), is weired.
> > >
> > > Or maybe I missed some context, e.g., can the head page be allocted to
> > > guest at all?
> >
> > Yes, this is to allow mapping the tail pages of higher order
> > non-compound pages - I should have been more precise in my wording.
> > The head pages can already be mapped into the guest.
> >
> > Treating the head and tail pages would require changing how KVM
> > behaves in a situation it supports today (rather than just adding
> > support for an unsupported situation). Currently, without this series,
> > KVM can map VM_PFNMAP|VM_IO memory backed by refcounted pages into the
> > guest. When that happens, KVM sets the A/D flags. I'm not sure whether
> > that's actually valid behavior, nor do I know whether anyone actually
> > cares about it. But it's what KVM does today, and I would shy away
> > from modifying that behavior without good reason.
> >
> > > >
> > > > The bulk of this change is tracking the is_refcounted_page flag so that
> > > > non-refcounted pages don't trigger page_count() == 0 warnings. This is
> > > > done by storing the flag in an unused bit in the sptes.
> > >
> > > Also, maybe we should mention this only works on x86-64.
> > >
> > > >
> > > > Signed-off-by: David Stevens <stevensd@xxxxxxxxxxxx>
> > > > ---
> > > > arch/x86/kvm/mmu/mmu.c | 44 +++++++++++++++++++++------------
> > > > arch/x86/kvm/mmu/mmu_internal.h | 1 +
> > > > arch/x86/kvm/mmu/paging_tmpl.h | 9 ++++---
> > > > arch/x86/kvm/mmu/spte.c | 4 ++-
> > > > arch/x86/kvm/mmu/spte.h | 12 ++++++++-
> > > > arch/x86/kvm/mmu/tdp_mmu.c | 22 ++++++++++-------
> > > > 6 files changed, 62 insertions(+), 30 deletions(-)
> > > >
> > > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > > > index e44ab512c3a1..b1607e314497 100644
> > > > --- a/arch/x86/kvm/mmu/mmu.c
> > > > +++ b/arch/x86/kvm/mmu/mmu.c
> > >
> > > ...
> > >
> > > > @@ -2937,6 +2943,7 @@ static int mmu_set_spte(struct kvm_vcpu *vcpu, struct kvm_memory_slot *slot,
> > > > bool host_writable = !fault || fault->map_writable;
> > > > bool prefetch = !fault || fault->prefetch;
> > > > bool write_fault = fault && fault->write;
> > > > + bool is_refcounted = !fault || fault->is_refcounted_page;
> > >
> > > Just wonder, what if a non-refcounted page is prefetched? Or is it possible in
> > > practice?
> >
> > Prefetching is still done via gfn_to_page_many_atomic, which sets
> > FOLL_GET. That's fixable, but it's not something this series currently
> > does.
>
> So if we prefetch a page, REFCOUNTED bit is cleared unconditionally with this
> hunk. kvm_set_page_{dirty, accessed} won't be called as expected for prefetched
> spte. If I read the patch correctly, REFCOUNTED bit in SPTE should represent
> whether the corresponding page is ref-countable or not, right?
>
> Because direct_pte_prefetch_many() is for legacy KVM MMU and FNAME(prefetch_pte)
> is shadow paging, we need to test it with legacy KVM MMU or shadow paging to hit
> the issue, though.
>
direct_pte_prefetch_many and prefetch_gpte both pass NULL for the
fault parameter, so is_refcounted will evaluate to true. So the spte's
refcounted bit will get set in that case.
-David
