Re: [PATCH v1 04/10] vfio/type1: Prepare is_invalid_reserved_pfn() for PG_reserved changes

[David Hildenbrand <david@redhat.com>]

On 08.11.19 08:14, David Hildenbrand wrote:
> On 08.11.19 06:09, Dan Williams wrote:
>> On Thu, Nov 7, 2019 at 2:07 PM David Hildenbrand <david@redhat.com> wrote:
>>>
>>> On 07.11.19 19:22, David Hildenbrand wrote:
>>>>
>>>>
>>>>> Am 07.11.2019 um 16:40 schrieb Dan Williams <dan.j.williams@intel.com>:
>>>>>
>>>>> On Thu, Oct 24, 2019 at 5:12 AM David Hildenbrand <david@redhat.com> 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 is_invalid_reserved_pfn() similar to kvm_is_reserved_pfn() to make
>>>>>> sure the function produces the same result once we stop setting ZONE_DEVICE
>>>>>> pages PG_reserved.
>>>>>>
>>>>>> Cc: Alex Williamson <alex.williamson@redhat.com>
>>>>>> Cc: Cornelia Huck <cohuck@redhat.com>
>>>>>> Signed-off-by: David Hildenbrand <david@redhat.com>
>>>>>> ---
>>>>>> drivers/vfio/vfio_iommu_type1.c | 10 ++++++++--
>>>>>> 1 file changed, 8 insertions(+), 2 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
>>>>>> index 2ada8e6cdb88..f8ce8c408ba8 100644
>>>>>> --- a/drivers/vfio/vfio_iommu_type1.c
>>>>>> +++ b/drivers/vfio/vfio_iommu_type1.c
>>>>>> @@ -299,9 +299,15 @@ static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
>>>>>>     */
>>>>>> static bool is_invalid_reserved_pfn(unsigned long pfn)
>>>>>> {
>>>>>> -       if (pfn_valid(pfn))
>>>>>> -               return PageReserved(pfn_to_page(pfn));
>>>>>> +       struct page *page = pfn_to_online_page(pfn);
>>>>>
>>>>> Ugh, I just realized this is not a safe conversion until
>>>>> pfn_to_online_page() is moved over to subsection granularity. As it
>>>>> stands it will return true for any ZONE_DEVICE pages that share a
>>>>> section with boot memory.
>>>>
>>>> That should not happen right now and I commented back when you introduced subsection support that I don’t want to have ZONE_DEVICE mixed with online pages in a section. Having memory block devices that partially span ZONE_DEVICE would be ... really weird. With something like pfn_active() - as discussed - we could at least make this check work - but I am not sure if we really want to go down that path. In the worst case, some MB of RAM are lost ... I guess this needs more thought.
>>>>
>>>
>>> I just realized the "boot memory" part. Is that a real thing? IOW, can
>>> we have ZONE_DEVICE falling into a memory block (with holes)? I somewhat
>>> have doubts that this would work ...
>>
>> One of the real world failure cases that started the subsection effect
>> is that Persistent Memory collides with System RAM on a 64MB boundary
>> on shipping platforms. System RAM ends on a 64MB boundary and due to a
>> lack of memory controller resources PMEM is mapped contiguously at the
>> end of that boundary. Some more details in the subsection cover letter
>> / changelogs [1] [2]. It's not sufficient to just lose some memory,
>> that's the broken implementation that lead to the subsection work
>> because the lost memory may change from one boot to the next and
>> software can't reliably inject a padding that conforms to the x86
>> 128MB section constraint.
> 
> Thanks, I thought it was mostly for weird alignment where other parts of
> the section are basically "holes" and not memory.
> 
> Yes, it is a real bug that ZONE_DEVICE pages fall into sections that are
> marked SECTION_IS_ONLINE.
> 
>>
>> Suffice to say I think we need your pfn_active() to get subsection
>> granularity pfn_to_online_page() before PageReserved() can be removed.
> 
> I agree that we have to fix this. I don't like ZONE_DEVICE pages falling
> into memory device blocks (e.g., cannot get offlined), but I guess that
> train is gone :) As long as it's not for memory hotplug, I can most
> probably live with this.
> 
> Also, I'd like to get Michals opinion on this and the pfn_active()
> approach, but I can understand he's busy.
> 
> This patch set can wait, I won't be working next week besides
> reading/writing mails either way.
> 
> Is anybody looking into the pfn_active() thingy?
> 

I wonder if we should do something like this right now to fix this 
(exclude the false positive ZONE_DEVICE pages we could have within an 
online section, which was not possible before subsection hotplug):

diff --git a/include/linux/memory_hotplug.h b/include/linux/memory_hotplug.h
index 384ffb3d69ab..490a9e9358b3 100644
--- a/include/linux/memory_hotplug.h
+++ b/include/linux/memory_hotplug.h
@@ -30,6 +30,8 @@ struct vmem_altmap;
         if (___nr < NR_MEM_SECTIONS && online_section_nr(___nr) && \
             pfn_valid_within(___pfn))                              \
                 ___page = pfn_to_page(___pfn);                     \
+       if (unlikely(___page && is_zone_device_page(___page)))     \
+               ___page = NULL;                                    \
         ___page;                                                   \
  })


Yeah, it's another is_zone_device_page(), but it should not be racy 
here, as we want to exclude, not include ZONE_DEVICE.

I don't have time to look into this right now, unfortunately.

-- 

Thanks,

David / dhildenb