Re: [PATCH] x86: kvm: Avoid guest page table walk when gpa_available is set

From: Brijesh Singh
Date: Mon Apr 24 2017 - 18:14:37 EST


Hi Radim,


/* GPA available (AMD only) */
bool gpa_available;
+ gpa_t gpa_val;

Can't we pass this information through function parameters?

(I'd rather avoid intractable variables.)


I also wanted to avoid adding yet another variable but we can't depend on
cr2 parameters passed into x86_emulate_instruction().

The x86_emulate_instruction() function is called from two places:

1) handling the page-fault.
pf_interception [svm.c]
kvm_mmu_page_fault [mmu.c]
x86_emulate_instruction [x86.c]

2) completing the IO/MMIO's from previous instruction decode
kvm_arch_vcpu_ioctl_run
complete_emulated_io
emulate_instruction
x86_emulate_instruction(vcpu, 0, emulation_type, NULL, 0)

In #1, we are guaranteed that cr2 variable will contain a valid GPA but
in #2, CR2 is set to zero.

};

struct kvm_lpage_info {
diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
index 5fba706..8827e4b 100644
--- a/arch/x86/kvm/svm.c
+++ b/arch/x86/kvm/svm.c
@@ -4159,6 +4159,10 @@ static int handle_exit(struct kvm_vcpu *vcpu)

vcpu->arch.gpa_available = (exit_code == SVM_EXIT_NPF);

+ /* On #NPF, exit_info_2 contain a valid GPA */
+ if (vcpu->arch.gpa_available)
+ vcpu->arch.gpa_val = svm->vmcb->control.exit_info_2;

How is vcpu->arch.gpa_val used between here and the NPF handler?


handle_exit [svm.c]
pf_interception [svm.c]
/* it invokes the fault handler with CR2 = svm->vmcb->control.exit_info_2 */
kvm_mmu_page_fault [mmu.c]
x86_emulate_instruction [x86.c]
emulator_read_write_onepage [x86.c]
/*
*this is where we walk the guest page table to translate
* a GVA to GPA. If gpa_available is set then we use the
* gpa_val instead of walking the pgtable.
*/

+
if (!is_cr_intercept(svm, INTERCEPT_CR0_WRITE))
vcpu->arch.cr0 = svm->vmcb->save.cr0;
if (npt_enabled)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
@@ -4653,18 +4653,16 @@ static int emulator_read_write_onepage(unsigned long addr, void *val,
* occurred.
*/
if (vcpu->arch.gpa_available &&
- emulator_can_use_gpa(ctxt) &&
- vcpu_is_mmio_gpa(vcpu, addr, exception->address, write) &&
- (addr & ~PAGE_MASK) == (exception->address & ~PAGE_MASK)) {
+ emulator_can_use_gpa(ctxt) &&
+ (addr & ~PAGE_MASK) == (exception->address & ~PAGE_MASK)) {
gpa = exception->address;
- goto mmio;
+ ret = vcpu_is_mmio_gpa(vcpu, addr, gpa, write);
+ } else {
+ ret = vcpu_mmio_gva_to_gpa(vcpu, addr, &gpa, exception, write);
+ if (ret < 0)
+ return X86EMUL_PROPAGATE_FAULT;
}

- ret = vcpu_mmio_gva_to_gpa(vcpu, addr, &gpa, exception, write);
-
- if (ret < 0)
- return X86EMUL_PROPAGATE_FAULT;
-
/* For APIC access vmexit */
if (ret)
goto mmio;
@@ -5675,8 +5673,14 @@ int x86_emulate_instruction(struct kvm_vcpu *vcpu,
}

restart:
- /* Save the faulting GPA (cr2) in the address field */
- ctxt->exception.address = cr2;
+ /*
+ * Save the faulting GPA (cr2) in the address field
+ * NOTE: If gpa_available is set then gpa_val will contain a valid GPA
+ */
+ if (vcpu->arch.gpa_available)
+ ctxt->exception.address = vcpu->arch.gpa_val;
+ else
+ ctxt->exception.address = cr2;

And related, shouldn't vcpu->arch.gpa_val be in cr2?


See my previous comment. In some cases CR2 may be set to zero
(e.g when completing the instruction from previous io/mmio page-fault).

If we are decide to add the gpa_val then we can remove above if
statement from x86_emulate_instruction() and update emulator_read_write_onepage
to use the vcpu->arch.gpa_val instead of exception->address.

if (vcpu->arch.gpa_available &&
emulator_can_use_gpa(ctxt) &&
(addr & ~PAGE_MASK) == (exception->address & ~PAGE_MASK)) {
gpa = vcpu=>arch.gpa_val;
...
...
}

-Brijesh