On Wed, Nov 06, 2019 at 07:56:34AM +0100, David Hildenbrand wrote:
> On 06.11.19 01:08, Dan Williams wrote:
> >On Tue, Nov 5, 2019 at 4:03 PM Sean Christopherson
> >>But David's proposed fix for the above refcount bug is to omit the patch
> >>so that KVM no longer treats ZONE_DEVICE pages as reserved. That seems
> >>like the right thing to do, including for thp_adjust(), e.g. it would
> >>naturally let KVM use 2mb pages for the guest when a ZONE_DEVICE page is
> >>mapped with a huge page (2mb or above) in the host. The only hiccup is
> >>figuring out how to correctly transfer the reference.
> >
> >That might not be the only hiccup. There's currently no such thing as
> >huge pages for ZONE_DEVICE, there are huge *mappings* (pmd and pud),
> >but the result of pfn_to_page() on such a mapping does not yield a
> >huge 'struct page'. It seems there are other paths in KVM that assume
> >that more typical page machinery is active like SetPageDirty() based
> >on kvm_is_reserved_pfn(). While I told David that I did not want to
> >see more usage of is_zone_device_page(), this patch below (untested)
> >seems a cleaner path with less surprises:
> >
> >diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> >index 4df0aa6b8e5c..fbea17c1810c 100644
> >--- a/virt/kvm/kvm_main.c
> >+++ b/virt/kvm/kvm_main.c
> >@@ -1831,7 +1831,8 @@ EXPORT_SYMBOL_GPL(kvm_release_page_clean);
> >
> > void kvm_release_pfn_clean(kvm_pfn_t pfn)
> > {
> >- if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn))
> >+ if ((!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn)) ||
The is_error_noslot_pfn() check shouldn't be overriden by zone_device.
> >+ (pfn_valid(pfn) && is_zone_device_page(pfn_to_page(pfn))))
But rather than special case kvm_release_pfn_clean(), I'd rather KVM
explicitly handle ZONE_DEVICE pages, there are other flows where KVM
really should be aware of ZONE_DEVICE pages, e.g. for sanity checks and
whatnot. There are surprisingly few callers of kvm_is_reserved_pfn(), so
it's actually not too big of a change.
> > put_page(pfn_to_page(pfn));
> > }
> > EXPORT_SYMBOL_GPL(kvm_release_pfn_clean);
>
> I had the same thought, but I do wonder about the kvm_get_pfn() users,
> e.g.,:
>
> hva_to_pfn_remapped():
> r = follow_pfn(vma, addr, &pfn);
> ...
> kvm_get_pfn(pfn);
> ...
>
> We would not take a reference for ZONE_DEVICE, but later drop one reference
> via kvm_release_pfn_clean(). IOW, kvm_get_pfn() gets *really* dangerous to
> use. I can't tell if this can happen right now.
>
> We do have 3 users of kvm_get_pfn() that we have to audit before this
> change. Also, we should add a comment to kvm_get_pfn() that it should never
> be used with possible ZONE_DEVICE pages.
>
> Also, we should add a comment to kvm_release_pfn_clean(), describing why we
> treat ZONE_DEVICE in a special way here.
>
>
> We can then progress like this
>
> 1. Get this fix upstream, it's somewhat unrelated to this series.
> 2. This patch here remains as is in this series (+/- documentation update)
> 3. Long term, rework KVM to not have to not treat ZONE_DEVICE like reserved
> pages. E.g., get rid of kvm_get_pfn(). Then, this special zone check can go.
Dropping kvm_get_pfn() is less than ideal, and at this point unnecessary.
I'm 99% sure the existing call sites for kvm_get_pfn() can never be
reached with ZONE_DEVICE pages. I think we can do:
1. Get a fix upstream to have KVM stop treating ZONE_DEVICE pages as
reserved PFNs, i.e. exempt them in kvm_is_reserved_pfn() and change
the callers of kvm_is_reserved_pfn() to handle ZONE_DEVICE pages.
2. Drop this patch from the series, and instead remove the special
treatment of ZONE_DEVICE pages from kvm_is_reserved_pfn().
Give me a few minutes to prep a patch.
>
> --
>
> Thanks,
>
> David / dhildenb
>
