On Tue, Nov 05, 2019 at 11:02:46AM +0100, David Hildenbrand wrote:
> On 05.11.19 10:49, David Hildenbrand wrote:
> >On 05.11.19 10:17, David Hildenbrand wrote:
> >>On 05.11.19 05:38, Dan Williams wrote:
> >>>On Thu, Oct 24, 2019 at 5:11 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >>>>
> >>>>Right now, ZONE_DEVICE memory is always set PG_reserved. We want to
> >>>>change that.
> >>>>
> >>>>KVM has this weird use case that you can map anything from /dev/mem
> >>>>into the guest. pfn_valid() is not a reliable check whether the memmap
> >>>>was initialized and can be touched. pfn_to_online_page() makes sure
> >>>>that we have an initialized memmap (and don't have ZONE_DEVICE memory).
> >>>>
> >>>>Rewrite kvm_is_reserved_pfn() to make sure the function produces the
> >>>>same result once we stop setting ZONE_DEVICE pages PG_reserved.
> >>>>
> >>>>Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> >>>>Cc: "Radim Krčmář" <rkrcmar@xxxxxxxxxx>
> >>>>Cc: Michal Hocko <mhocko@xxxxxxxxxx>
> >>>>Cc: Dan Williams <dan.j.williams@xxxxxxxxx>
> >>>>Cc: KarimAllah Ahmed <karahmed@xxxxxxxxx>
> >>>>Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> >>>>---
> >>>> virt/kvm/kvm_main.c | 10 ++++++++--
> >>>> 1 file changed, 8 insertions(+), 2 deletions(-)
> >>>>
> >>>>diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> >>>>index e9eb666eb6e8..9d18cc67d124 100644
> >>>>--- a/virt/kvm/kvm_main.c
> >>>>+++ b/virt/kvm/kvm_main.c
> >>>>@@ -151,9 +151,15 @@ __weak int kvm_arch_mmu_notifier_invalidate_range(struct kvm *kvm,
> >>>>
> >>>> bool kvm_is_reserved_pfn(kvm_pfn_t pfn)
> >>>> {
> >>>>- if (pfn_valid(pfn))
> >>>>- return PageReserved(pfn_to_page(pfn));
> >>>>+ struct page *page = pfn_to_online_page(pfn);
> >>>>
> >>>>+ /*
> >>>>+ * We treat any pages that are not online (not managed by the buddy)
> >>>>+ * as reserved - this includes ZONE_DEVICE pages and pages without
> >>>>+ * a memmap (e.g., mapped via /dev/mem).
> >>>>+ */
> >>>>+ if (page)
> >>>>+ return PageReserved(page);
> >>>> return true;
> >>>> }
> >>>
> >>>So after this all the pfn_valid() usage in kvm_main.c is replaced with
> >>>pfn_to_online_page()? Looks correct to me.
> >>>
> >>>However, I'm worried that kvm is taking reference on ZONE_DEVICE pages
> >>>through some other path resulting in this:
> >>>
> >>> https://lore.kernel.org/linux-nvdimm/20190919154708.GA24650@xxxxxxxxxx/
> >>>
> >>>I'll see if this patch set modulates or maintains that failure mode.
> >>>
> >>
> >>I'd assume that the behavior is unchanged. Ithink we get a reference to
> >>these ZONE_DEVICE pages via __get_user_pages_fast() and friends in
> >>hva_to_pfn_fast() and friends in virt/kvm/kvm_main.c
> >>
> >
> >I think I know what's going wrong:
> >
> >Pages that are pinned via gfn_to_pfn() and friends take a references,
> >however are often released via
> >kvm_release_pfn_clean()/kvm_release_pfn_dirty()/kvm_release_page_clean()...
> >
> >
> >E.g., in arch/x86/kvm/x86.c:reexecute_instruction()
> >
> >...
> >pfn = gfn_to_pfn(vcpu->kvm, gpa_to_gfn(gpa));
> >...
> >kvm_release_pfn_clean(pfn);
> >
> >
> >
> >void kvm_release_pfn_clean(kvm_pfn_t pfn)
> >{
> > if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn))
> > put_page(pfn_to_page(pfn));
> >}
> >
> >This function makes perfect sense as the counterpart for kvm_get_pfn():
> >
> >void kvm_get_pfn(kvm_pfn_t pfn)
> >{
> > if (!kvm_is_reserved_pfn(pfn))
> > get_page(pfn_to_page(pfn));
> >}
> >
> >
> >As all ZONE_DEVICE pages are currently reserved, pages pinned via
> >gfn_to_pfn() and friends will often not see a put_page() AFAIKS.
Assuming gup() takes a reference for ZONE_DEVICE pages, yes, this is a
KVM bug.
> >Now, my patch does not change that, the result of
> >kvm_is_reserved_pfn(pfn) will be unchanged. A proper fix for that would
> >probably be
> >
> >a) To drop the reference to ZONE_DEVICE pages in gfn_to_pfn() and
> >friends, after you successfully pinned the pages. (not sure if that's
> >the right thing to do but you're the expert)
> >
> >b) To not use kvm_release_pfn_clean() and friends on pages that were
> >definitely pinned.
This is already KVM's intent, i.e. the purpose of the PageReserved() check
is simply to avoid putting a non-existent reference. The problem is that
KVM assumes pages with PG_reserved set are never pinned, which AFAICT was
true when the code was first added.
> (talking to myself, sorry)
>
> Thinking again, dropping this patch from this series could effectively also
> fix that issue. E.g., kvm_release_pfn_clean() and friends would always do a
> put_page() if "pfn_valid() and !PageReserved()", so after patch 9 also on
> ZONDE_DEVICE pages.
Yeah, this appears to be the correct fix.
> But it would have side effects that might not be desired. E.g.,:
>
> 1. kvm_pfn_to_page() would also return ZONE_DEVICE pages (might even be the
> right thing to do).
This should be ok, at least on x86. There are only three users of
kvm_pfn_to_page(). Two of those are on allocations that are controlled by
KVM and are guaranteed to be vanilla MAP_ANONYMOUS. The third is on guest
memory when running a nested guest, and in that case supporting ZONE_DEVICE
memory is desirable, i.e. KVM should play nice with a guest that is backed
by ZONE_DEVICE memory.
> 2. kvm_set_pfn_dirty() would also set ZONE_DEVICE pages dirty (might be
> okay)
This is ok from a KVM perspective.
The scarier code (for me) is transparent_hugepage_adjust() and
kvm_mmu_zap_collapsible_spte(), as I don't at all understand the
interaction between THP and _PAGE_DEVMAP.
_______________________________________________
devel mailing list
devel@xxxxxxxxxxxxxxxxxxxxxx
http://driverdev.linuxdriverproject.org/mailman/listinfo/driverdev-devel
