On 2018-10-31 at 09:49 Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> > On 2018-10-29 at 20:10 Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
> >> The property of DAX pages that requires special coordination is the
> >> fact that the device hosting the pages can be disabled at will. The
> >> get_dev_pagemap() api is the interface to pin a device-pfn so that you
> >> can safely perform a pfn_to_page() operation.
> >>
> >> Have the pages that kvm uses in this path already been pinned by vfio?
>
> No, VFIO is not involved here.
>
> The pages that KVM uses are never pinned. Soon after we grab them and
> we build KVM's page table, we do put_page in mmu_set_spte (via
> kvm_release_pfn_clean). From that point on the MMU notifier will take
> care of invalidating SPT before the page disappears from the mm's page
> table.
I looked into this a little, and I think it's safe in terms of DAX's
get_dev_pagemap() refcounts to have kvm_is_reserved_pfn() treat
DAX pages as not reserved. kvm_is_reserved_pfn() is used before a
pfn_to_page() call, so the concern was that the pages didn't have a DAX
refcount.
Many of the times that KVM looks at the PFN, it's holding the KVM
mmu_lock, which keeps the pages in the host-side VMA. (IIUC).
Functions like kvm_set_pfn_dirty() fall into this category.
The other times, such as the vmexit path I mentioned before, go through
a gfn_to_pfn call. Under the hood, those call one of the
get_user_pages() functions during hva_to_pfn, and those do the
right thing w.r.t. get_dev_pagemap().
One of the other things I noticed was some places in KVM make a
distinction between kvm_is_reserved_pfn and PageReserved:
void kvm_set_pfn_dirty(kvm_pfn_t pfn)
{
if (!kvm_is_reserved_pfn(pfn)) {
struct page *page = pfn_to_page(pfn);
if (!PageReserved(page))
SetPageDirty(page);
}
}
I think we want to SetPageDirty for DAX, so making PageReserved be true
for DAX seems like the way to go, or we'll need more KVM-specific
changes. Apologies is this was discussed in the previous thread on this
topic and is redundant.
Thanks,
Barret
