Re: [RFC PATCH v2 04/51] KVM: guest_memfd: Introduce KVM_GMEM_CONVERT_SHARED/PRIVATE ioctls

[Alexey Kardashevskiy <aik@amd.com>]

On 28/6/25 01:17, Vishal Annapurve wrote:
> On Thu, Jun 26, 2025 at 9:50 PM Alexey Kardashevskiy <aik@amd.com> wrote:
>>
>>
>>
>> On 25/6/25 00:10, Vishal Annapurve wrote:
>>> On Tue, Jun 24, 2025 at 6:08 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>>>>
>>>> On Tue, Jun 24, 2025 at 06:23:54PM +1000, Alexey Kardashevskiy wrote:
>>>>
>>>>> Now, I am rebasing my RFC on top of this patchset and it fails in
>>>>> kvm_gmem_has_safe_refcount() as IOMMU holds references to all these
>>>>> folios in my RFC.
>>>>>
>>>>> So what is the expected sequence here? The userspace unmaps a DMA
>>>>> page and maps it back right away, all from the userspace? The end
>>>>> result will be the exactly same which seems useless. And IOMMU TLB
>>>
>>>    As Jason described, ideally IOMMU just like KVM, should just:
>>> 1) Directly rely on guest_memfd for pinning -> no page refcounts taken
>>> by IOMMU stack
>>> 2) Directly query pfns from guest_memfd for both shared/private ranges
>>> 3) Implement an invalidation callback that guest_memfd can invoke on
>>> conversions.
> 
> Conversions and truncations both.
> 
>>>
>>> Current flow:
>>> Private to Shared conversion via kvm_gmem_convert_range() -
>>>       1) guest_memfd invokes kvm_gmem_invalidate_begin() for the ranges
>>> on each bound memslot overlapping with the range
>>>            -> KVM has the concept of invalidation_begin() and end(),
>>> which effectively ensures that between these function calls, no new
>>> EPT/NPT entries can be added for the range.
>>>        2) guest_memfd invokes kvm_gmem_convert_should_proceed() which
>>> actually unmaps the KVM SEPT/NPT entries.
>>>        3) guest_memfd invokes kvm_gmem_execute_work() which updates the
>>> shareability and then splits the folios if needed
>>>
>>> Shared to private conversion via kvm_gmem_convert_range() -
>>>       1) guest_memfd invokes kvm_gmem_invalidate_begin() for the ranges
>>> on each bound memslot overlapping with the range
>>>        2) guest_memfd invokes kvm_gmem_convert_should_proceed() which
>>> actually unmaps the host mappings which will unmap the KVM non-seucure
>>> EPT/NPT entries.
>>>        3) guest_memfd invokes kvm_gmem_execute_work() which updates the
>>> shareability and then merges the folios if needed.
>>>
>>> ============================
>>>
>>> For IOMMU, could something like below work?
>>>
>>> * A new UAPI to bind IOMMU FDs with guest_memfd ranges
>>
>> Done that.
>>
>>> * VFIO_DMA_MAP/UNMAP operations modified to directly fetch pfns from
>>> guest_memfd ranges using kvm_gmem_get_pfn()
>>
>> This API imho should drop the confusing kvm_ prefix.
>>
>>>       -> kvm invokes kvm_gmem_is_private() to check for the range
>>> shareability, IOMMU could use the same or we could add an API in gmem
>>> that takes in access type and checks the shareability before returning
>>> the pfn.
>>
>> Right now I cutnpasted kvm_gmem_get_folio() (which essentially is filemap_lock_folio()/filemap_alloc_folio()/__filemap_add_folio()) to avoid new links between iommufd.ko and kvm.ko. It is probably unavoidable though.
> 
> I don't think that's the way to avoid links between iommufd.ko and
> kvm.ko. Cleaner way probably is to have gmem logic built-in and allow
> runtime registration of invalidation callbacks from KVM/IOMMU
> backends. Need to think about this more.

Yeah, otherwise iommufd.ko will have to install a hook in guest_memfd (==kvm.ko) in run time so more beloved symbol_get() :)

> 
>>
>>
>>> * IOMMU stack exposes an invalidation callback that can be invoked by
>>> guest_memfd.
>>>
>>> Private to Shared conversion via kvm_gmem_convert_range() -
>>>       1) guest_memfd invokes kvm_gmem_invalidate_begin() for the ranges
>>> on each bound memslot overlapping with the range
>>>        2) guest_memfd invokes kvm_gmem_convert_should_proceed() which
>>> actually unmaps the KVM SEPT/NPT entries.
>>>              -> guest_memfd invokes IOMMU invalidation callback to zap
>>> the secure IOMMU entries.
>>>        3) guest_memfd invokes kvm_gmem_execute_work() which updates the
>>> shareability and then splits the folios if needed
>>>        4) Userspace invokes IOMMU map operation to map the ranges in
>>> non-secure IOMMU.
>>>
>>> Shared to private conversion via kvm_gmem_convert_range() -
>>>       1) guest_memfd invokes kvm_gmem_invalidate_begin() for the ranges
>>> on each bound memslot overlapping with the range
>>>        2) guest_memfd invokes kvm_gmem_convert_should_proceed() which
>>> actually unmaps the host mappings which will unmap the KVM non-seucure
>>> EPT/NPT entries.
>>>            -> guest_memfd invokes IOMMU invalidation callback to zap the
>>> non-secure IOMMU entries.
>>>        3) guest_memfd invokes kvm_gmem_execute_work() which updates the
>>> shareability and then merges the folios if needed.
>>>        4) Userspace invokes IOMMU map operation to map the ranges in secure IOMMU.
>>
>>
>> Alright (although this zap+map is not necessary on the AMD hw).
> 
> IMO guest_memfd ideally should not directly interact or cater to arch
> specific needs, it should implement a mechanism that works for all
> archs. KVM/IOMMU implement invalidation callbacks and have all the
> architecture specific knowledge to take the right decisions.


Every page conversion will go through:

kvm-amd.ko -1-> guest_memfd (kvm.ko) -2-> iommufd.ko -3-> amd-iommu (build-in).

Which one decides on IOMMU not needing (un)mapping? Got to be (1) but then it need to propagate the decision to amd-iommu (and we do not have (3) at the moment in that path).

Or we just always do unmap+map (and trigger unwanted page huge page smashing)? All is doable and neither particularly horrible, I'm trying to see where the consensus is now. Thanks,


>>
>>> There should be a way to block external IOMMU pagetable updates while
>>> guest_memfd is performing conversion e.g. something like
>>> kvm_invalidate_begin()/end().
>>>
>>>>> is going to be flushed on a page conversion anyway (the RMPUPDATE
>>>>> instruction does that). All this is about AMD's x86 though.
>>>>
>>>> The iommu should not be using the VMA to manage the mapping. It should
>>>
>>> +1.
>>
>> Yeah, not doing this already, because I physically cannot map gmemfd's memory in IOMMU via VMA (which allocates memory via gup() so wrong memory is mapped in IOMMU). Thanks,
>>
>>
>>>> be directly linked to the guestmemfd in some way that does not disturb
>>>> its operations. I imagine there would be some kind of invalidation
>>>> callback directly to the iommu.
>>>>
>>>> Presumably that invalidation call back can include a reason for the
>>>> invalidation (addr change, shared/private conversion, etc)
>>>>
>>>> I'm not sure how we will figure out which case is which but guestmemfd
>>>> should allow the iommu to plug in either invalidation scheme..
>>>>
>>>> Probably invalidation should be a global to the FD thing, I imagine
>>>> that once invalidation is established the iommu will not be
>>>> incrementing page refcounts.
>>>
>>> +1.
>>
>> Alright. Thanks for the comments.
>>
>>>
>>>>
>>>> Jason
>>
>> --
>> Alexey
>>

-- 
Alexey