On 01/09/16 13:46, Christoffer Dall wrote:

On Fri, Aug 19, 2016 at 01:38:13PM +0100, Marc Zyngier wrote:

quoted

In order to efficiently perform the GICV access on behalf of the
guest, we need to be able to do avoid going back all the way to

s/do//

quoted

the host kernel.

For this, we introduce a new hook in the world switch code,
conveniently placed just after populating the fault info.
At that point, we only have saved/restored the GP registers,
and we can quickly perform all the required checks (data abort,
translation fault, valid faulting syndrome, not an external
abort, not a PTW).

Coming back from the emulation code, we need to skip the emulated
instruction. This involves an additional bit of save/restore in
order to be able to access the guest's PC (and possibly CPSR if
this is a 32bit guest).

At this stage, no emulation code is provided.

Signed-off-by: Marc Zyngier <redacted>
---
 arch/arm64/include/asm/kvm_hyp.h |  1 +
 arch/arm64/kvm/hyp/switch.c      | 32 ++++++++++++++++++++++++++++++++
 include/kvm/arm_vgic.h           |  3 +++
 virt/kvm/arm/hyp/vgic-v2-sr.c    |  7 +++++++
 virt/kvm/arm/vgic/vgic-v2.c      |  2 ++
 5 files changed, 45 insertions(+)

diff --git a/arch/arm64/include/asm/kvm_hyp.h b/arch/arm64/include/asm/kvm_hyp.h
index cff5105..88ec3ac 100644
--- a/arch/arm64/include/asm/kvm_hyp.h
+++ b/arch/arm64/include/asm/kvm_hyp.h

@@ -123,6 +123,7 @@ typeof(orig) * __hyp_text fname(void)					\
 
 void __vgic_v2_save_state(struct kvm_vcpu *vcpu);
 void __vgic_v2_restore_state(struct kvm_vcpu *vcpu);
+bool __vgic_v2_perform_cpuif_access(struct kvm_vcpu *vcpu);
 
 void __vgic_v3_save_state(struct kvm_vcpu *vcpu);
 void __vgic_v3_restore_state(struct kvm_vcpu *vcpu);

diff --git a/arch/arm64/kvm/hyp/switch.c b/arch/arm64/kvm/hyp/switch.c
index ae7855f..0be1594 100644
--- a/arch/arm64/kvm/hyp/switch.c
+++ b/arch/arm64/kvm/hyp/switch.c

@@ -17,6 +17,7 @@
 
 #include <linux/types.h>
 #include <asm/kvm_asm.h>
+#include <asm/kvm_emulate.h>
 #include <asm/kvm_hyp.h>
 
 static bool __hyp_text __fpsimd_enabled_nvhe(void)

@@ -232,6 +233,21 @@ static bool __hyp_text __populate_fault_info(struct kvm_vcpu *vcpu)
 	return true;
 }
 
+static void __hyp_text __skip_instr(struct kvm_vcpu *vcpu)
+{
+	vcpu->arch.ctxt.gp_regs.regs.pc	= read_sysreg_el2(elr);
+
+	if (vcpu_mode_is_32bit(vcpu)) {
+		vcpu->arch.ctxt.gp_regs.regs.pstate = read_sysreg_el2(spsr);
+		kvm_skip_aarch32_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
+		write_sysreg_el2(vcpu->arch.ctxt.gp_regs.regs.pstate, spsr);
+	} else {
+		*vcpu_pc(vcpu) += 4;
+	}
+
+	write_sysreg_el2(vcpu->arch.ctxt.gp_regs.regs.pc, elr);
+}
+
 static int __hyp_text __guest_run(struct kvm_vcpu *vcpu)
 {
 	struct kvm_cpu_context *host_ctxt;

@@ -270,6 +286,22 @@ again:
 	if (exit_code == ARM_EXCEPTION_TRAP && !__populate_fault_info(vcpu))
 		goto again;
 
+	if (static_branch_unlikely(&vgic_v2_cpuif_trap) &&
+	    exit_code == ARM_EXCEPTION_TRAP) {

do you get the static branch benefit when the test contains an &&
clause?  (I haven't looked at the generated assembly, no)

You do, otherwise the C semantics would be broken. This is strictly
equivalent to:

	if (static_branch_unlikely()) {
		if (exit_code == ...) {
			[...]
		}
	}

Also, if you flip this static branch for code both mapped in EL1 and
EL2, would you potentially need some form of additional icache
maintenance here?

Or are you relying on the static branch being set at boot time and hold
forever true/false?

I asked myself this exact question when I did this, and became convinced
that this was OK for two reasons:
- we do it only once
- when we do it, we haven't executed that code yet, so it cannot be in
the cache yet

quoted

+		bool valid;
+
+		valid = kvm_vcpu_trap_get_class(vcpu) == ESR_ELx_EC_DABT_LOW &&
+			kvm_vcpu_trap_get_fault_type(vcpu) == FSC_FAULT &&
+			kvm_vcpu_dabt_isvalid(vcpu) &&
+			!kvm_vcpu_dabt_isextabt(vcpu) &&
+			!kvm_vcpu_dabt_iss1tw(vcpu);
+
+		if (valid &&  __vgic_v2_perform_cpuif_access(vcpu)) {

extra whitespace

quoted

+			__skip_instr(vcpu);

does this interact in any amusing way with single-step guest debugging?

Ouch. Good point. Actually, nothing that uses kvm_skip_instr() works for
singlestep/watchpoint either (trapped sysreg, WFx, spurious traps,
MMIO). I guess that's something we need to fix overall.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...