diff mbox series

[v12,44/46] KVM: selftests: evmcs_test: Introduce L2 TLB flush test

Message ID 20221021153521.1216911-45-vkuznets@redhat.com
State New
Headers
Series KVM: x86: hyper-v: Fine-grained TLB flush + L2 TLB flush features |

Commit Message

Vitaly Kuznetsov Oct. 21, 2022, 3:35 p.m. UTC 
  Enable Hyper-V L2 TLB flush and check that Hyper-V TLB flush hypercalls
from L2 don't exit to L1 unless 'TlbLockCount' is set in the
Partition assist page.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
---
 .../selftests/kvm/include/x86_64/evmcs.h      |  2 +
 .../testing/selftests/kvm/x86_64/evmcs_test.c | 52 ++++++++++++++++++-
 2 files changed, 52 insertions(+), 2 deletions(-)

Comments

Sean Christopherson Oct. 26, 2022, 10:21 p.m. UTC | #1 
On Fri, Oct 21, 2022, Vitaly Kuznetsov wrote:
> @@ -48,6 +49,8 @@ static inline void rdmsr_gs_base(void)
>  
>  void l2_guest_code(void)
>  {
> +	u64 unused;
> +
>  	GUEST_SYNC(7);
>  
>  	GUEST_SYNC(8);
> @@ -64,15 +67,33 @@ void l2_guest_code(void)
>  	vmcall();
>  	rdmsr_gs_base(); /* intercepted */
>  
> +	/* L2 TLB flush tests */
> +	hyperv_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE | HV_HYPERCALL_FAST_BIT, 0x0,
> +			 HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES | HV_FLUSH_ALL_PROCESSORS);
> +	rdmsr_fs_base();
> +	/*
> +	 * Note: hypercall status (RAX) is not preserved correctly by L1 after
> +	 * synthetic vmexit, use unchecked version.
> +	 */
> +	__hyperv_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE | HV_HYPERCALL_FAST_BIT, 0x0,
> +			   HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES | HV_FLUSH_ALL_PROCESSORS,
> +			   &unused);
> +	/* Make sure we're not issuing Hyper-V TLB flush call again */
> +	__asm__ __volatile__ ("mov $0xdeadbeef, %rcx");

This needs a clobber.  It won't cause problems in the current code, but it will
make someone really sad if they add more code after this.

Using %ecx instead of %rcx will also suffice, 32-bit accesses clear bits 63:32.

Even better, if a "nop" isn't required to get the compiler to emit preamble, would
be to load ECX through an input constraint, that way the compiler "knows" the value
of ECX and can optimize for it (though it's extremely unlikely 0xdeadbeef will be
a useful value).

Actually, not setting RCX in vmcall() is a nasty bug waiting to happen, e.g. if
RCX just happens to contain a value that gets routed to L0.

Rather than handle this as a one-off, can you insert a prep patch to have the
common vmcall() stuff RCX with a "safe" value?

Related side topic, rdmsr_{f,g}s_base() should also use input constraints, and
should use a proper #define for the MSRs.  Also, why earth do those clobber all GPRs?
Oooh, because they get routed to L1 and L1 doesn't preserve GPRs.

Related side topic #2, KVM's kvm_xen_hypercall() is broken, it checks the wrong
input register for 64-bit (checks RAX instead of RCX).  Not sure that's even a
fixable bug though.
diff mbox series

Patch

diff --git a/tools/testing/selftests/kvm/include/x86_64/evmcs.h b/tools/testing/selftests/kvm/include/x86_64/evmcs.h
index 94d6059e9a12..901caf0e0939 100644
--- a/tools/testing/selftests/kvm/include/x86_64/evmcs.h
+++ b/tools/testing/selftests/kvm/include/x86_64/evmcs.h
@@ -237,6 +237,8 @@  struct hv_enlightened_vmcs {
 #define HV_VMX_ENLIGHTENED_CLEAN_FIELD_ENLIGHTENMENTSCONTROL    BIT(15)
 #define HV_VMX_ENLIGHTENED_CLEAN_FIELD_ALL                      0xFFFF
 
+#define HV_VMX_SYNTHETIC_EXIT_REASON_TRAP_AFTER_FLUSH 0x10000031
+
 extern struct hv_enlightened_vmcs *current_evmcs;
 
 int vcpu_enable_evmcs(struct kvm_vcpu *vcpu);
diff --git a/tools/testing/selftests/kvm/x86_64/evmcs_test.c b/tools/testing/selftests/kvm/x86_64/evmcs_test.c
index 74f076ba574b..16d98f8fc50c 100644
--- a/tools/testing/selftests/kvm/x86_64/evmcs_test.c
+++ b/tools/testing/selftests/kvm/x86_64/evmcs_test.c
@@ -16,6 +16,7 @@ 
 
 #include "kvm_util.h"
 
+#include "hyperv.h"
 #include "vmx.h"
 
 static int ud_count;
@@ -48,6 +49,8 @@  static inline void rdmsr_gs_base(void)
 
 void l2_guest_code(void)
 {
+	u64 unused;
+
 	GUEST_SYNC(7);
 
 	GUEST_SYNC(8);
@@ -64,15 +67,33 @@  void l2_guest_code(void)
 	vmcall();
 	rdmsr_gs_base(); /* intercepted */
 
+	/* L2 TLB flush tests */
+	hyperv_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE | HV_HYPERCALL_FAST_BIT, 0x0,
+			 HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES | HV_FLUSH_ALL_PROCESSORS);
+	rdmsr_fs_base();
+	/*
+	 * Note: hypercall status (RAX) is not preserved correctly by L1 after
+	 * synthetic vmexit, use unchecked version.
+	 */
+	__hyperv_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE | HV_HYPERCALL_FAST_BIT, 0x0,
+			   HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES | HV_FLUSH_ALL_PROCESSORS,
+			   &unused);
+	/* Make sure we're not issuing Hyper-V TLB flush call again */
+	__asm__ __volatile__ ("mov $0xdeadbeef, %rcx");
+
 	/* Done, exit to L1 and never come back.  */
 	vmcall();
 }
 
-void guest_code(struct vmx_pages *vmx_pages, struct hyperv_test_pages *hv_pages)
+void guest_code(struct vmx_pages *vmx_pages, struct hyperv_test_pages *hv_pages,
+		vm_vaddr_t hv_hcall_page_gpa)
 {
 #define L2_GUEST_STACK_SIZE 64
 	unsigned long l2_guest_stack[L2_GUEST_STACK_SIZE];
 
+	wrmsr(HV_X64_MSR_GUEST_OS_ID, HYPERV_LINUX_OS_ID);
+	wrmsr(HV_X64_MSR_HYPERCALL, hv_hcall_page_gpa);
+
 	x2apic_enable();
 
 	GUEST_SYNC(1);
@@ -102,7 +123,17 @@  void guest_code(struct vmx_pages *vmx_pages, struct hyperv_test_pages *hv_pages)
 	vmwrite(PIN_BASED_VM_EXEC_CONTROL, vmreadz(PIN_BASED_VM_EXEC_CONTROL) |
 		PIN_BASED_NMI_EXITING);
 
+	/* L2 TLB flush setup */
+	current_evmcs->partition_assist_page = hv_pages->partition_assist_gpa;
+	current_evmcs->hv_enlightenments_control.nested_flush_hypercall = 1;
+	current_evmcs->hv_vm_id = 1;
+	current_evmcs->hv_vp_id = 1;
+	current_vp_assist->nested_control.features.directhypercall = 1;
+	*(u32 *)(hv_pages->partition_assist) = 0;
+
 	GUEST_ASSERT(!vmlaunch());
+	GUEST_ASSERT_EQ(vmreadz(VM_EXIT_REASON), EXIT_REASON_EXCEPTION_NMI);
+	GUEST_ASSERT_EQ((vmreadz(VM_EXIT_INTR_INFO) & 0xff), NMI_VECTOR);
 	GUEST_ASSERT(vmptrstz() == hv_pages->enlightened_vmcs_gpa);
 
 	/*
@@ -146,6 +177,18 @@  void guest_code(struct vmx_pages *vmx_pages, struct hyperv_test_pages *hv_pages)
 	GUEST_ASSERT(vmreadz(VM_EXIT_REASON) == EXIT_REASON_MSR_READ);
 	current_evmcs->guest_rip += 2; /* rdmsr */
 
+	/*
+	 * L2 TLB flush test. First VMCALL should be handled directly by L0,
+	 * no VMCALL exit expected.
+	 */
+	GUEST_ASSERT(!vmresume());
+	GUEST_ASSERT(vmreadz(VM_EXIT_REASON) == EXIT_REASON_MSR_READ);
+	current_evmcs->guest_rip += 2; /* rdmsr */
+	/* Enable synthetic vmexit */
+	*(u32 *)(hv_pages->partition_assist) = 1;
+	GUEST_ASSERT(!vmresume());
+	GUEST_ASSERT(vmreadz(VM_EXIT_REASON) == HV_VMX_SYNTHETIC_EXIT_REASON_TRAP_AFTER_FLUSH);
+
 	GUEST_ASSERT(!vmresume());
 	GUEST_ASSERT(vmreadz(VM_EXIT_REASON) == EXIT_REASON_VMCALL);
 	GUEST_SYNC(11);
@@ -199,6 +242,7 @@  static struct kvm_vcpu *save_restore_vm(struct kvm_vm *vm,
 int main(int argc, char *argv[])
 {
 	vm_vaddr_t vmx_pages_gva = 0, hv_pages_gva = 0;
+	vm_vaddr_t hcall_page;
 
 	struct kvm_vcpu *vcpu;
 	struct kvm_vm *vm;
@@ -212,12 +256,16 @@  int main(int argc, char *argv[])
 	TEST_REQUIRE(kvm_has_cap(KVM_CAP_NESTED_STATE));
 	TEST_REQUIRE(kvm_has_cap(KVM_CAP_HYPERV_ENLIGHTENED_VMCS));
 
+	hcall_page = vm_vaddr_alloc_pages(vm, 1);
+	memset(addr_gva2hva(vm, hcall_page), 0x0,  getpagesize());
+
 	vcpu_set_hv_cpuid(vcpu);
 	vcpu_enable_evmcs(vcpu);
 
 	vcpu_alloc_vmx(vm, &vmx_pages_gva);
 	vcpu_alloc_hyperv_test_pages(vm, &hv_pages_gva);
-	vcpu_args_set(vcpu, 2, vmx_pages_gva, hv_pages_gva);
+	vcpu_args_set(vcpu, 3, vmx_pages_gva, hv_pages_gva, addr_gva2gpa(vm, hcall_page));
+	vcpu_set_msr(vcpu, HV_X64_MSR_VP_INDEX, vcpu->id);
 
 	vm_init_descriptor_tables(vm);
 	vcpu_init_descriptor_tables(vcpu);