Re: [PATCH 5/5] kvm/svm: copy instruction bytes from VMCB

[Avi Kivity <avi@redhat.com>]

On 12/10/2010 03:51 PM, Andre Przywara wrote:
> In case of a nested page fault or an intercepted #PF newer SVM
> implementations provide a copy of the faulting instruction bytes
> in the VMCB.
> Use these bytes to feed the instruction emulator and avoid the costly
> guest instruction fetch in this case.
>
>
>
> +static int svm_prefetch_instruction(struct kvm_vcpu *vcpu)
> +{
> +	struct vcpu_svm *svm = to_svm(vcpu);
> +	uint8_t len;
> +	struct fetch_cache *fetch;
> +
> +	len = svm->vmcb->control.insn_len&  0x0F;
> +	if (len == 0)
> +		return 1;
> +
> +	fetch =&svm->vcpu.arch.emulate_ctxt.decode.fetch;
> +	fetch->start = kvm_rip_read(&svm->vcpu);
> +	fetch->end = fetch->start + len;
> +	memcpy(fetch->data, svm->vmcb->control.insn_bytes, len);
> +
> +	return 0;
> +}

This reaching in into the emulator internals from svm code is not very 
good.  It also assumes ->prefetch_instruction() is called immediately 
after an exit; this isn't true in vmx and at least was considered for 
svm (emulating multiple instructions during the nsvm vmexit sequence).

Alternatives are:
- add the insn data to emulate_instruction() and friends (my first 
suggestion)
- adding x86_decode_insn_init(), which initializes the decode cache, and 
x86_decode_insn_prefill_cache(), called only if we have the insn data

Another one: teach kvm_fetch_guest_virt() to check if addr/bytes 
intersects with csbase+rip/len; if so, use that instead of doing the 
page table dance.

-- 
error compiling committee.c: too many arguments to function