[PATCH v4 03/34] KVM: selftests: Add a shameful hack to preserve/clobber GPRs across ucall

[Sean Christopherson <seanjc@google.com>]

Preserve or clobber all GPRs (except RIP and RSP, as they're saved and
restored via the VMCS) when performing a ucall on x86 to fudge around a
horrific long-standing bug in selftests' nested VMX support where L2's
GPRs are not preserved across a nested VM-Exit.  I.e. if a test triggers a
nested VM-Exit to L1 in response to a ucall, e.g. GUEST_SYNC(), then L2's
GPR state can be corrupted.

The issues manifests as an unexpected #GP in clear_bit() when running the
hyperv_evmcs test due to RBX being used to track the ucall object, and RBX
being clobbered by the nested VM-Exit.  The problematic hyperv_evmcs
testcase is where L0 (test's host userspace) injects an NMI in response to
GUEST_SYNC(8) from L2, but the bug could "randomly" manifest in any test
that induces a nested VM-Exit from L0.  The bug hasn't caused failures in
the past due to sheer dumb luck.

The obvious fix is to rework the nVMX helpers to save/restore L2 GPRs
across VM-Exit and VM-Enter, but that is a much bigger task and carries
its own risks, e.g. nSVM does save/restore GPRs, but not in a thread-safe
manner, and there is a _lot_ of cleanup that can be done to unify code
for doing VM-Enter on nVMX, nSVM, and eVMCS.

Signed-off-by: Sean Christopherson <seanjc@google.com>
---
 .../testing/selftests/kvm/lib/x86_64/ucall.c  | 32 +++++++++++++++++--
 1 file changed, 30 insertions(+), 2 deletions(-)

diff --git a/tools/testing/selftests/kvm/lib/x86_64/ucall.c b/tools/testing/selftests/kvm/lib/x86_64/ucall.c
index 4d41dc63cc9e..a53df3ece2f8 100644
--- a/tools/testing/selftests/kvm/lib/x86_64/ucall.c
+++ b/tools/testing/selftests/kvm/lib/x86_64/ucall.c
@@ -14,8 +14,36 @@ void ucall_arch_init(struct kvm_vm *vm, vm_paddr_t mmio_gpa)
 
 void ucall_arch_do_ucall(vm_vaddr_t uc)
 {
-	asm volatile("in %[port], %%al"
-		: : [port] "d" (UCALL_PIO_PORT), "D" (uc) : "rax", "memory");
+	/*
+	 * FIXME: Revert this hack (the entire commit that added it) once nVMX
+	 * preserves L2 GPRs across a nested VM-Exit.  If a ucall from L2, e.g.
+	 * to do a GUEST_SYNC(), lands the vCPU in L1, any and all GPRs can be
+	 * clobbered by L1.  Save and restore non-volatile GPRs (clobbering RBP
+	 * in particular is problematic) along with RDX and RDI (which are
+	 * inputs), and clobber volatile GPRs. *sigh*
+	 */
+#define HORRIFIC_L2_UCALL_CLOBBER_HACK	\
+	"rcx", "rsi", "r8", "r9", "r10", "r11"
+
+	asm volatile("push %%rbp\n\t"
+		     "push %%r15\n\t"
+		     "push %%r14\n\t"
+		     "push %%r13\n\t"
+		     "push %%r12\n\t"
+		     "push %%rbx\n\t"
+		     "push %%rdx\n\t"
+		     "push %%rdi\n\t"
+		     "in %[port], %%al\n\t"
+		     "pop %%rdi\n\t"
+		     "pop %%rdx\n\t"
+		     "pop %%rbx\n\t"
+		     "pop %%r12\n\t"
+		     "pop %%r13\n\t"
+		     "pop %%r14\n\t"
+		     "pop %%r15\n\t"
+		     "pop %%rbp\n\t"
+		: : [port] "d" (UCALL_PIO_PORT), "D" (uc) : "rax", "memory",
+		     HORRIFIC_L2_UCALL_CLOBBER_HACK);
 }
 
 void *ucall_arch_get_ucall(struct kvm_vcpu *vcpu)
-- 
2.41.0.487.g6d72f3e995-goog