Thread (128 messages) 128 messages, 13 authors, 2024-09-25

Re: [RFC PATCH 13/21] KVM: X86: Handle private MMIO as shared

From: Xu Yilun <hidden>
Date: 2024-08-30 16:59:42
Also in: kvm, linux-iommu, linux-pci

On Fri, Aug 23, 2024 at 11:21:27PM +1000, Alexey Kardashevskiy wrote:
Currently private MMIO nested page faults are not expected so when such
fault occurs, KVM tries moving the faulted page from private to shared
which is not going to work as private MMIO is not backed by memfd.

Handle private MMIO as shared: skip page state change and memfd
This means host keeps the mapping for private MMIO, which is different
from private memory. Not sure if it is expected, and I want to get
some directions here.

From HW perspective, private MMIO is not intended to be accessed by
host, but the consequence may varies. According to TDISP spec 11.2,
my understanding is private device (known as TDI) should reject the
TLP and transition to TDISP ERROR state. But no further error
reporting or logging is mandated. So the impact to the host system
is specific to each device. In my test environment, an AER
NonFatalErr is reported and nothing more, much better than host
accessing private memory.

On SW side, my concern is how to deal with mmu_notifier. In theory, if
we get pfn from hva we should follow the userspace mapping change. But
that makes no sense. Especially for TDX TEE-IO, private MMIO mapping
in SEPT cannot be changed or invalidated as long as TDI is running.

Another concern may be specific for TDX TEE-IO. Allowing both userspace
mapping and SEPT mapping may be safe for private MMIO, but on
KVM_SET_USER_MEMORY_REGION2,  KVM cannot actually tell if a userspace
addr is really for private MMIO. I.e. user could provide shared memory
addr to KVM but declare it is for private MMIO. The shared memory then
could be mapped in SEPT and cause problem.

So personally I prefer no host mapping for private MMIO.

Thanks,
Yilun
quoted hunk ↗ jump to hunk
page state tracking.

The MMIO KVM memory slot is still marked as shared as the guest can
access it as private or shared so marking the MMIO slot as private
is not going to help.

Signed-off-by: Alexey Kardashevskiy <redacted>
---
 arch/x86/kvm/mmu/mmu.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index 928cf84778b0..e74f5c3d0821 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -4366,7 +4366,11 @@ static int __kvm_faultin_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault
 {
 	bool async;
 
-	if (fault->is_private)
+	if (fault->slot && fault->is_private && !kvm_slot_can_be_private(fault->slot) &&
+	    (vcpu->kvm->arch.vm_type == KVM_X86_SNP_VM))
+		pr_warn("%s: private SEV TIO MMIO fault for fault->gfn=%llx\n",
+			__func__, fault->gfn);
+	else if (fault->is_private)
 		return kvm_faultin_pfn_private(vcpu, fault);
 
 	async = false;
-- 
2.45.2
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help