On Mon, Nov 11, 2019 at 2:12 PM Sean Christopherson
<sean.j.christopherson@xxxxxxxxx> wrote:
>
> Explicitly exempt ZONE_DEVICE pages from kvm_is_reserved_pfn() and
> instead manually handle ZONE_DEVICE on a case-by-case basis. For things
> like page refcounts, KVM needs to treat ZONE_DEVICE pages like normal
> pages, e.g. put pages grabbed via gup(). But for flows such as setting
> A/D bits or shifting refcounts for transparent huge pages, KVM needs to
> to avoid processing ZONE_DEVICE pages as the flows in question lack the
> underlying machinery for proper handling of ZONE_DEVICE pages.
>
> This fixes a hang reported by Adam Borowski[*] in dev_pagemap_cleanup()
> when running a KVM guest backed with /dev/dax memory, as KVM straight up
> doesn't put any references to ZONE_DEVICE pages acquired by gup().
>
> Note, Dan Williams proposed an alternative solution of doing put_page()
> on ZONE_DEVICE pages immediately after gup() in order to simplify the
> auditing needed to ensure is_zone_device_page() is called if and only if
> the backing device is pinned (via gup()). But that approach would break
> kvm_vcpu_{un}map() as KVM requires the page to be pinned from map() 'til
> unmap() when accessing guest memory, unlike KVM's secondary MMU, which
> coordinates with mmu_notifier invalidations to avoid creating stale
> page references, i.e. doesn't rely on pages being pinned.
>
> [*] http://lkml.kernel.org/r/20190919115547.GA17963@xxxxxxxxxx
>
> Reported-by: Adam Borowski <kilobyte@xxxxxxxxxx>
> Debugged-by: David Hildenbrand <david@xxxxxxxxxx>
> Cc: Dan Williams <dan.j.williams@xxxxxxxxx>
Acked-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> Cc: stable@xxxxxxxxxxxxxxx
Perhaps add:
Fixes: 3565fce3a659 ("mm, x86: get_user_pages() for dax mappings")
...since that was the first kernel that broke KVM's assumption about
which pfn types needed to have the reference count managed.
