Re: [RFC PATCH v1 2/9] KVM: guest_memfd: Add guest_memfd support to kvm_(read|/write)_guest_page()

[Patrick Roy <roypat@amazon.co.uk>]

On Thu, 2025-01-23 at 12:28 +0000, Fuad Tabba wrote:
> Hi Patrick,
> 
> On Thu, 23 Jan 2025 at 11:57, Patrick Roy <roypat@amazon.co.uk> wrote:
>>
>>
>>
>> On Thu, 2025-01-23 at 11:39 +0000, David Hildenbrand wrote:
>>> On 23.01.25 10:48, Fuad Tabba wrote:
>>>> On Wed, 22 Jan 2025 at 22:10, David Hildenbrand <david@redhat.com> wrote:
>>>>>
>>>>> On 22.01.25 16:27, Fuad Tabba wrote:
>>>>>> Make kvm_(read|/write)_guest_page() capable of accessing guest
>>>>>> memory for slots that don't have a userspace address, but only if
>>>>>> the memory is mappable, which also indicates that it is
>>>>>> accessible by the host.
>>>>>
>>>>> Interesting. So far my assumption was that, for shared memory, user
>>>>> space would simply mmap() guest_memdd and pass it as userspace address
>>>>> to the same memslot that has this guest_memfd for private memory.
>>>>>
>>>>> Wouldn't that be easier in the first shot? (IOW, not require this patch
>>>>> with the cost of faulting the shared page into the page table on access)
>>>>
>>>
>>> In light of:
>>>
>>> https://lkml.kernel.org/r/20250117190938.93793-4-imbrenda@linux.ibm.com
>>>
>>> there can, in theory, be memslots that start at address 0 and have a
>>> "valid" mapping. This case is done from the kernel (and on special s390x
>>> hardware), though, so it does not apply here at all so far.
>>>
>>> In practice, getting address 0 as a valid address is unlikely, because
>>> the default:
>>>
>>> $ sysctl  vm.mmap_min_addr
>>> vm.mmap_min_addr = 65536
>>>
>>> usually prohibits it for good reason.
>>>
>>>> This has to do more with the ABI I had for pkvm and shared memory
>>>> implementations, in which you don't need to specify the userspace
>>>> address for memory in a guestmem memslot. The issue is there is no
>>>> obvious address to map it to. This would be the case in kvm:arm64 for
>>>> tracking paravirtualized time, which the userspace doesn't necessarily
>>>> need to interact with, but kvm does.
>>>
>>> So I understand correctly: userspace wouldn't have to mmap it because it
>>> is not interested in accessing it, but there is nothing speaking against
>>> mmaping it, at least in the first shot.
>>>
>>> I assume it would not be a private memslot (so far, my understanding is
>>> that internal memslots never have a guest_memfd attached).
>>> kvm_gmem_create() is only called via KVM_CREATE_GUEST_MEMFD, to be set
>>> on user-created memslots.
>>>
>>>>
>>>> That said, we could always have a userspace address dedicated to
>>>> mapping shared locations, and use that address when the necessity
>>>> arises. Or we could always require that memslots have a userspace
>>>> address, even if not used. I don't really have a strong preference.
>>>
>>> So, the simpler version where user space would simply mmap guest_memfd
>>> to provide the address via userspace_addr would at least work for the
>>> use case of paravirtualized time?
>>
>> fwiw, I'm currently prototyping something like this for x86 (although
>> not by putting the gmem address into userspace_addr, but by adding a new
>> field to memslots, so that memory attributes continue working), based on
>> what we talked about at the last guest_memfd sync meeting (the whole
>> "how to get MMIO emulation working for non-CoCo VMs in guest_memfd"
>> story). So I guess if we're going down this route for x86, maybe it
>> makes sense to do the same on ARM, for consistency?
>>
>>> It would get rid of the immediate need for this patch and patch #4 to
>>> get it flying.
>>>
>>>
>>> One interesting question is: when would you want shared memory in
>>> guest_memfd and *not* provide it as part of the same memslot.
>>
>> In my testing of non-CoCo gmem VMs on ARM, I've been able to get quite
>> far without giving KVM a way to internally access shared parts of gmem -
>> it's why I was probing Fuad for this simplified series, because
>> KVM_SW_PROTECTED_VM + mmap (for loading guest kernel) is enough to get a
>> working non-CoCo VM on ARM (although I admittedly never looked at clocks
>> inside the guest - maybe that's one thing that breaks if KVM can't
>> access gmem. How to guest and host agree on the guest memory range
>> used to exchange paravirtual timekeeping information? Could that exchange
>> be intercepted in userspace, and set to shared via memory attributes (e.g.
>> placed outside gmem)? That's the route I'm going down the paravirtual
>> time on x86).
> 
> For an idea of what it looks like on arm64, here's how kvmtool handles it:
> https://github.com/kvmtool/kvmtool/blob/master/arm/aarch64/pvtime.c
> 
> Cheers,
> /fuad
 
Thanks! In that example, kvmtool actually allocates a separate memslot for
the pvclock stuff, so I guess it's always possible to simply put it into
a non-gmem memslot, which indeed sidesteps this issue as you mention in
your reply to David :D
  
>>> One nice thing about the mmap might be that access go via user-space
>>> page tables: E.g., __kvm_read_guest_page can just access the memory
>>> without requiring the folio lock and an additional temporary folio
>>> reference on every access -- it's handled implicitly via the mapcount.
>>>
>>> (of course, to map the page we still need that once on the fault path)
>>
>> Doing a direct map access in kvm_{read,write}_guest() and friends will
>> also get tricky if guest_memfd folios ever don't have direct map
>> entries. On-demand restoration is painful, both complexity and
>> performance wise [1], while going through a userspace mapping of
>> guest_memfd would "just work".
>>
>>> --
>>> Cheers,
>>>
>>> David / dhildenb
>>>