Add hypercall-based PCI CAM for x86-64 pKVM.

src/transport/mod.rs

@@ -5,6 +5,8 @@ pub mod fake;
 pub mod mmio;
 pub mod pci;
 mod some;
+#[cfg(target_arch = "x86_64")]
+pub mod x86_64;
 
 use crate::{PhysAddr, Result, PAGE_SIZE};
 use bitflags::{bitflags, Flags};

src/transport/x86_64.rs

@@ -0,0 +1,124 @@
+//! x86-64 specific transports.
+
+use super::pci::bus::{Cam, ConfigurationAccess, DeviceFunction};
+use core::arch::asm;
+
+/// This CPUID returns the signature and should be used to determine if VM is running under pKVM,
+/// KVM or not. See the Linux header `arch/x86/include/uapi/asm/kvm_para.h`.
+const KVM_CPUID_SIGNATURE: u32 = 0x40000000;
+
+// See `include/uapi/linux/kvm_para.h`. (These hypercalls numbers can change depending on the
+// upstream progress.)
+const KVM_HC_PKVM_OP: u32 = 20;
+const PKVM_GHC_IOREAD: u32 = KVM_HC_PKVM_OP + 3;
+const PKVM_GHC_IOWRITE: u32 = KVM_HC_PKVM_OP + 4;
+
+const PKVM_SIGNATURE: &[u8] = b"PKVM";
+
+/// A PCI configuration access mechanism using hypercalls implemented by the x86-64 pKVM hypervisor.
+pub struct HypCam {
+    /// The physical base address of the PCI root complex.
+    phys_base: usize,
+    cam: Cam,
+}
+
+impl HypCam {
+    /// Creates a new `HypCam` for the PCI root complex at the given physical base address.
+    pub fn new(phys_base: usize, cam: Cam) -> Self {
+        Self { phys_base, cam }
+    }
+
+    /// Returns whether we are running under pKVM by checking the CPU ID signature.
+    pub fn is_pkvm() -> bool {
+        cpuid_signature() == PKVM_SIGNATURE
+    }
+}
+
+impl ConfigurationAccess for HypCam {
+    fn read_word(&self, device_function: DeviceFunction, register_offset: u8) -> u32 {
+        let address = self.cam.cam_offset(device_function, register_offset);
+        hyp_io_read(self.phys_base + (address as usize), 4)
+    }
+
+    fn write_word(&mut self, device_function: DeviceFunction, register_offset: u8, data: u32) {
+        let address = self.cam.cam_offset(device_function, register_offset);
+        hyp_io_write(self.phys_base + (address as usize), 4, data);
+    }
+
+    unsafe fn unsafe_clone(&self) -> Self {
+        Self {
+            phys_base: self.phys_base,
+            cam: self.cam,
+        }
+    }
+}
+
+/// Gets the signature CPU ID.
+fn cpuid_signature() -> [u8; 4] {
+    let signature: u32;
+    unsafe {
+        // The argument for cpuid is passed via rax and in case of KVM_CPUID_SIGNATURE returned via
+        // rbx, rcx and rdx. Ideally using a named argument in inline asm for rbx would be more
+        // straightforward, but when "rbx" is directly used LLVM complains that it is used
+        // internally.
+        //
+        // Therefore use r8 instead and push rbx to the stack before making cpuid call, store
+        // rbx content to r8 as use it as inline asm output and pop the rbx.
+        asm!(
+            "push rbx",
+            "cpuid",
+            "mov r8, rbx",
+            "pop rbx",
+            in("eax") KVM_CPUID_SIGNATURE,
+            out("r8") signature,
+            out("rcx") _,
+            out("rdx") _,
+        );
+    };
+    signature.to_le_bytes()
+}
+
+/// Asks the hypervisor to perform an IO read at the given physical address.
+fn hyp_io_read(address: usize, size: usize) -> u32 {
+    // Arguments for vmcall are passed via rax, rbx, rcx and rdx. Ideally using a named argument in
+    // the inline asm for rbx would be more straightforward, but when "rbx" is used directly LLVM
+    // complains that it is used internally.
+    //
+    // Therefore use r8 temporary, push rbx to the stack, perform proper call and pop rbx
+    // again
+    let data;
+    unsafe {
+        asm!(
+            "push rbx",
+            "mov rbx, r8",
+            "vmcall",
+            "pop rbx",
+            inout("rax") PKVM_GHC_IOREAD => data,
+            in("r8") address,
+            in("rcx") size,
+        );
+    }
+    data
+}
+
+/// Asks the hypervisor to perform an IO write at the given physical address.
+fn hyp_io_write(address: usize, size: usize, data: u32) {
+    unsafe {
+        // Arguments for vmcall are passed via rax, rbx, rcx and rdx. Ideally using a named argument
+        // in the inline asm for rbx would be more straightforward but when "rbx" is used directly
+        // used LLVM complains that it is used internally.
+        //
+        // Therefore use r8 temporary, push rbx to the stack, perform proper call and pop rbx
+        // again
+        asm!(
+            "push rbx",
+            "mov rbx, r8",
+            "vmcall",
+            "pop rbx",
+            in("rax") PKVM_GHC_IOWRITE,
+            in("r8") address,
+            in("rcx") size,
+            in("rdx") data,
+        );
+    }
+}