On 21.11.2008, at 17:53, Avi Kivity wrote:
Alexander Graf wrote:
Until this patch we bounced between the VM and the VMM for a couple
of
instructions after CLGI, only to find out that yet another SVM
instruction
follows.
Since roundtrips are really expensive, it's a lot faster to emulate
these
few instructions. Now we can execute CLGI/VMLOAD/VMRUN on one
intercept and
VMEXIT/VMSAVE on another. This is a major speed improvement!
Neat trick. Have you looked at svm.c to see if we can cut out
extraneous instructions (or move them out of the gif=0 area)? Could
mean big savings.
Well for now I only emulate pre-vmrun and vmsave, because the emulator
breaks on lldt and set dr6 and ltr and ...
+static int nested_svm_emulate(struct vcpu_svm *svm, struct kvm_run
*kvm_run)
+{
+ int er;
+ u32 opcode = 0;
+ unsigned long rip;
+ unsigned long rip_linear;
+
+ svm->vmcb->save.rax = svm->vcpu.arch.regs[VCPU_REGS_RAX];
+ svm->vmcb->save.rsp = svm->vcpu.arch.regs[VCPU_REGS_RSP];
+ svm->vmcb->save.rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
+ rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
+ rip_linear = rip + svm_seg(&svm->vcpu, VCPU_SREG_CS)->base;
+
+ er = emulator_read_std(rip_linear, (void *)&opcode, 3, &svm->vcpu);
+ if (er != X86EMUL_CONTINUE)
+ return er;
+ er = EMULATE_FAIL;
+
+ switch (opcode) {
+ case 0xda010f:
+ vmload_interception(svm, kvm_run);
+ er = EMULATE_DONE;
+ break;
+ case 0xd8010f:
+ vmrun_interception(svm, kvm_run);
+ er = EMULATE_DONE;
+ break;
+ case 0xdb010f:
+ vmsave_interception(svm, kvm_run);
+ er = EMULATE_DONE;
+ break;
+ case 0xdc010f:
+ stgi_interception(svm, kvm_run);
+ er = EMULATE_DONE;
+ break;
+ default:
+ nsvm_printk("NSVM: Opcode %x unknown\n", opcode);
+ }
+
+ nsvm_printk("NSVM: svm emul at 0x%lx -> %d\n", rip, er);
+
+ return er;
+}
+
Move to the regular x86 emulator, so if we extend it with debug flag
support, privilege checking, etc, we get that for svm as well.
Do you seriously want to have a callback to svm specific code in the
x86 emulator? I was thinking of doing that, but this way looked way
cleaner to me. I would actually rather use prefixes as hint on pattern
matching.
So e.g. if we find a clgi instruction with some random prefix (or
somehow marked in other ways) we could just run a pattern that does
what the entry code would do. This way no emulation would be needed.
static int nested_svm_vmexit_real(struct vcpu_svm *svm, void *arg1,
void *arg2, void *opaque)
{
@@ -1551,6 +1600,9 @@ static int nested_svm_vmexit(struct vcpu_svm
*svm)
kvm_mmu_reset_context(&svm->vcpu);
kvm_mmu_load(&svm->vcpu);
+ /* KVM calls vmsave after vmrun, so let's run it now if we can */
+ nested_svm_emulate(svm, NULL);
+
Will also call stgi eventually, so it may make sense to loop here too.
See above - I did that, it broke, I tried to debug it for ~2 days and
just figured I'll leave it as is for now. The current version gives
_huge_ savings already.
Alex
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
