Re: [PATCH 3/4] x86,module: Detect VMX vs SLD conflicts

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Apr 08, 2020 at 05:09:34PM +0900, Masami Hiramatsu wrote:
> On Tue, 07 Apr 2020 13:02:39 +0200
> Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > +static bool insn_is_vmx(struct insn *insn)
> > +{
> > +	u8 modrm = insn->modrm.bytes[0];
> > +	u8 modrm_mod = X86_MODRM_MOD(modrm);
> > +	u8 modrm_reg = X86_MODRM_REG(modrm);
> > +
> > +	u8 prefix = insn->prefixes.bytes[0];
> 
> This should be the last prefix,
> 
> 	u8 prefix = insn->prefixes.bytes[3];
> 
> (The last prefix always copied on the bytes[3])

And that is 0 on no-prefix, right?

> > +
> > +	if (insn->opcode.bytes[0] != 0x0f)
> > +		return false;
> > +
> > +	switch (insn->opcode.bytes[1]) {
> > +	case 0x01:
> > +		switch (insn->opcode.bytes[2]) {
> 
> Sorry, VMCALL etc. is in Grp7 (0f 01), the 3rd code is embedded
> in modrm instead of opcode. Thus it should be,
> 
> 		switch (insn->modrm.value) {

Indeed, I was hoping (I really should've checked) that that byte was
duplicated in opcodes.

Also, since I already have modrm = insn->modrm.bytes[0], I should
probably use that anyway.

> > +		case 0xc1: /* VMCALL */
> > +		case 0xc2: /* VMLAUNCH */
> > +		case 0xc3: /* VMRESUME */
> > +		case 0xc4: /* VMXOFF */
> 
> 		case 0xd4:	/* VMFUNC */

As per Andrew, VMCALL and VMFUNC are SMV, and I really only need VMX in
this case. Including SMV is probably harmless, but I'm thinking a
smaller function is better.

> > +			return true;
> > +
> > +		default:
> > +			break;
> > +		}
> > +		break;
> > +
> > +	case 0x78: /* VMREAD */
> > +	case 0x79: /* VMWRITE */
> 
> 		return !insn_is_evex(insn);
> 
> With EVEX prefix, these becomes vcvt* instructions.

Thanks!