On 7/9/24 3:20 PM, Patrick Roy wrote:
> Hey all,
>
> This RFC series is a rough draft adding support for running
> non-confidential compute VMs in guest_memfd, based on prior discussions
> with Sean [1]. Our specific usecase for this is the ability to unmap
> guest memory from the host kernel's direct map, as a mitigation against
> a large class of speculative execution issues.
>
> === Implementation ===
>
> This patch series introduces a new flag to the `KVM_CREATE_GUEST_MEMFD`
> to remove its pages from the direct map when they are allocated. When
> trying to run a guest from such a VM, we now face the problem that
> without either userspace or kernelspace mappings of guest_memfd, KVM
> cannot access guest memory to, for example, do MMIO emulation of access
> memory used to guest/host communication. We have multiple options for
> solving this when running non-CoCo VMs: (1) implement a TDX-light
> solution, where the guest shares memory that KVM needs to access, and
> relies on paravirtual solutions where this is not possible (e.g. MMIO),
> (2) have KVM use userspace mappings of guest_memfd (e.g. a
> memfd_secret-style solution), or (3) dynamically reinsert pages into the
> direct map whenever KVM wants to access them.
>
> This RFC goes for option (3). Option (1) is a lot of overhead for very
> little gain, since we are not actually constrained by a physical
> inability to access guest memory (e.g. we are not in a TDX context where
> accesses to guest memory cause a #MC). Option (2) has previously been
> rejected [1].
Do the pages have to have the same address when they are temporarily mapped?
Wouldn't it be easier to do something similar to kmap_local_page() used for
HIMEM? I.e. you get a temporary kernel mapping to do what's needed, but it
doesn't have to alter the shared directmap.
Maybe that was already discussed somewhere as unsuitable but didn't spot it
here.
> In this patch series, we make sufficient parts of KVM gmem-aware to be
> able to boot a Linux initrd from private memory on x86. These include
> KVM's MMIO emulation (including guest page table walking) and kvm-clock.
> For VM types which do not allow accessing gmem, we return -EFAULT and
> attempt to prepare a KVM_EXIT_MEMORY_FAULT.
>
> Additionally, this patch series adds support for "restricted" userspace
> mappings of guest_memfd, which work similar to memfd_secret (e.g.
> disallow get_user_pages), which allows handling I/O and loading the
> guest kernel in a simple way. Support for this is completely independent
> of the rest of the functionality introduced in this patch series.
> However, it is required to build a minimal hypervisor PoC that actually
> allows booting a VM from a disk.
>
> === Performance ===
>
> We have run some preliminary performance benchmarks to assess the impact
> of on-the-fly direct map manipulations. We were mainly interested in the
> impact of manipulating the direct map for MMIO emulation on virtio-mmio.
> Particularly, we were worried about the impact of the TLB and L1/2/3
> Cache flushes that set_memory_[n]p entails.
>
> In our setup, we have taken a modified Firecracker VMM, spawned a Linux
> guest with 1 vCPU, and used fio to stress a virtio_blk device. We found
> that the cache flushes caused throughput to drop from around 600MB/s to
> ~50MB/s (~90%) for both reads and writes (on a Intel(R) Xeon(R) Platinum
> 8375C CPU with 64 cores). We then converted our prototype to use
> set_direct_map_{invalid,default}_noflush instead of set_memory_[n]p and
> found that without cache flushes the pure impact of the direct map
> manipulation is indistinguishable from noise. This is why we use
> set_direct_map_{invalid,default}_noflush instead of set_memory_[n]p in
> this RFC.
>
> Note that in this comparison, both the baseline, as well as the
> guest_memfd-supporting version of Firecracker were made to bounce I/O
> buffers in VMM userspace. As GUP is disabled for the guest_memfd VMAs,
> the virtio stack cannot directly pass guest buffers to read/write
> syscalls.
>
> === Security ===
>
> We want to use unmapping guest memory from the host kernel as a security
> mitigation against transient execution attacks. Temporarily restoring
> direct map entries whenever KVM requires access to guest memory leaves a
> gap in this mitigation. We believe this to be acceptable for the above
> cases, since pages used for paravirtual guest/host communication (e.g.
> kvm-clock) and guest page tables do not contain sensitive data. MMIO
> emulation will only end up reading pages containing privileged
> instructions (e.g. guest kernel code).
>
> === Summary ===
>
> Patches 1-4 are about hot-patching various points inside of KVM that
> access guest memory to correctly handle the case where memory happens to
> be guest-private. This means either handling the access as a memory
> error, or simply accessing the memslot's guest_memfd instead of looking
> at the userspace provided VMA if the VM type allows these kind of
> accesses. Patches 5-6 add a flag to KVM_CREATE_GUEST_MEMFD that will
> make it remove its pages from the kernel's direct map. Whenever KVM
> wants to access guest-private memory, it will temporarily re-insert the
> relevant pages. Patches 7-8 allow for restricted userspace mappings
> (e.g. get_user_pages paths are disabled like for memfd_secret) of
> guest_memfd, so that userspace has an easy path for loading the guest
> kernel and handling I/O-buffers.
>
> === ToDos / Limitations ===
>
> There are still a few rough edges that need to be addressed before
> dropping the "RFC" tag, e.g.
>
> * Handle errors of set_direct_map_default_not_flush in
> kvm_gmem_invalidate_folio instead of calling BUG_ON
> * Lift the limitation of "at most one gfn_to_pfn_cache for each
> gfn/pfn" in e1c61f0a7963 ("kvm: gmem: Temporarily restore direct map
> entries when needed"). It currently means that guests with more than 1
> vcpu fail to boot, because multiple vcpus can put their kvm-clock PV
> structures into the same page (gfn)
> * Write selftests, particularly around hole punching, direct map removal,
> and mmap.
>
> Lastly, there's the question of nested virtualization which Sean brought
> up in previous discussions, which runs into similar problems as MMIO. I
> have looked at it very briefly. On Intel, KVM uses various gfn->uhva
> caches, which run in similar problems as the gfn_to_hva_caches dealt
> with in 200834b15dda ("kvm: use slowpath in gfn_to_hva_cache if memory
> is private"). However, previous attempts at just converting this to
> gfn_to_pfn_cache (which would make them work with guest_memfd) proved
> complicated [2]. I suppose initially, we should probably disallow nested
> virtualization in VMs that have their memory removed from the direct
> map.
>
> Best,
> Patrick
>
> [1]: https://lore.kernel.org/linux-mm/cc1bb8e9bc3e1ab637700a4d3defeec95b55060a.camel@xxxxxxxxxx/
> [2]: https://lore.kernel.org/kvm/ZBEEQtmtNPaEqU1i@xxxxxxxxxx/
>
> Patrick Roy (8):
> kvm: Allow reading/writing gmem using kvm_{read,write}_guest
> kvm: use slowpath in gfn_to_hva_cache if memory is private
> kvm: pfncache: enlighten about gmem
> kvm: x86: support walking guest page tables in gmem
> kvm: gmem: add option to remove guest private memory from direct map
> kvm: gmem: Temporarily restore direct map entries when needed
> mm: secretmem: use AS_INACCESSIBLE to prohibit GUP
> kvm: gmem: Allow restricted userspace mappings
>
> arch/x86/kvm/mmu/paging_tmpl.h | 94 +++++++++++++++++++-----
> include/linux/kvm_host.h | 5 ++
> include/linux/kvm_types.h | 1 +
> include/linux/secretmem.h | 13 +++-
> include/uapi/linux/kvm.h | 2 +
> mm/secretmem.c | 6 +-
> virt/kvm/guest_memfd.c | 83 +++++++++++++++++++--
> virt/kvm/kvm_main.c | 112 +++++++++++++++++++++++++++-
> virt/kvm/pfncache.c | 130 +++++++++++++++++++++++++++++----
> 9 files changed, 399 insertions(+), 47 deletions(-)
>
>
> base-commit: 890a64810d59b1a58ed26efc28cfd821fc068e84
