Re: [PATCH v1 03/10] KVM: Prepare kvm_is_reserved_pfn() for PG_reserved changes

[David Hildenbrand <david@redhat.com>]

On 06.11.19 01:08, Dan Williams wrote:
> On Tue, Nov 5, 2019 at 4:03 PM Sean Christopherson
> <sean.j.christopherson@intel.com> wrote:
>>
>> On Tue, Nov 05, 2019 at 03:43:29PM -0800, Dan Williams wrote:
>>> On Tue, Nov 5, 2019 at 3:30 PM Dan Williams <dan.j.williams@intel.com> wrote:
>>>>
>>>> On Tue, Nov 5, 2019 at 3:13 PM Sean Christopherson
>>>> <sean.j.christopherson@intel.com> wrote:
>>>>>
>>>>> On Tue, Nov 05, 2019 at 03:02:40PM -0800, Dan Williams wrote:
>>>>>> On Tue, Nov 5, 2019 at 12:31 PM David Hildenbrand <david@redhat.com> wrote:
>>>>>>>> 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.
>>>>>>>
>>>>>>> The x86 KVM MMU code is one of the ugliest code I know (sorry, but it
>>>>>>> had to be said :/ ). Luckily, this should be independent of the
>>>>>>> PG_reserved thingy AFAIKs.
>>>>>>
>>>>>> Both transparent_hugepage_adjust() and kvm_mmu_zap_collapsible_spte()
>>>>>> are honoring kvm_is_reserved_pfn(), so again I'm missing where the
>>>>>> page count gets mismanaged and leads to the reported hang.
>>>>>
>>>>> When mapping pages into the guest, KVM gets the page via gup(), which
>>>>> increments the page count for ZONE_DEVICE pages.  But KVM puts the page
>>>>> using kvm_release_pfn_clean(), which skips put_page() if PageReserved()
>>>>> and so never puts its reference to ZONE_DEVICE pages.
>>>>
>>>> Oh, yeah, that's busted.
>>>
>>> Ugh, it's extra busted because every other gup user in the kernel
>>> tracks the pages resulting from gup and puts them (put_page()) when
>>> they are done. KVM wants to forget about whether it did a gup to get
>>> the page and optionally trigger put_page() based purely on the pfn.
>>> Outside of VFIO device assignment that needs pages pinned for DMA, why
>>> does KVM itself need to pin pages? If pages are pinned over a return
>>> to userspace that needs to be a FOLL_LONGTERM gup.
>>
>> Short answer, KVM pins the page to ensure correctness with respect to the
>> primary MMU invalidating the associated host virtual address, e.g. when
>> the page is being migrated or unmapped from host userspace.
>>
>> The main use of gup() is to handle guest page faults and map pages into
>> the guest, i.e. into KVM's secondary MMU.  KVM uses gup() to both get the
>> PFN and to temporarily pin the page.  The pin is held just long enough to
>> guaranteed that any invalidation via the mmu_notifier will be stalled
>> until after KVM finishes installing the page into the secondary MMU, i.e.
>> the pin is short-term and not held across a return to userspace or entry
>> into the guest.  When a subsequent mmu_notifier invalidation occurs, KVM
>> pulls the PFN from the secondary MMU and uses that to update accessed
>> and dirty bits in the host.
>>
>> There are a few other KVM flows that eventually call into gup(), but those
>> are "traditional" short-term pins and use put_page() directly.
> 
> Ok, I was misinterpreting the effect of the bug with what KVM is using
> the reference to do.
> 
> To your other point:
> 
>> 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)) ||
> +           (pfn_valid(pfn) && is_zone_device_page(pfn_to_page(pfn))))
>                  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.

-- 

Thanks,

David / dhildenb